new

Get trending papers in your email inbox!

Subscribe

Daily Papers

byAK and the research community

Nov 20

Safe Actions Alone Do Not Ensure Safe Agents: Identifying Unfulfilled Obligations with Guard Models

Guard models are increasingly used to safeguard LLM-based agents, primarily by identifying actions that agents are forbidden to perform. However, identifying forbidden actions alone is insufficient to ensure agent safety. In this paper, we argue that agent safety also depends on identifying required yet unperformed safety-critical actions, which we call obligations. Our preliminary study on a popular benchmark for evaluating safety shows that 56.92% of GLM-5.3 trajectories contain unfulfilled obligations, compared with only 30.00% containing forbidden actions. This finding reveals unfulfilled obligations as a major and previously overlooked source of safety risk. However, to our knowledge, no existing benchmark evaluates whether guard models can identify these obligations. To close this gap, we introduce ObligationBench, the first benchmark for evaluating the capability of obligation identification, comprising 240 expert-validated trajectories covering issue resolution, feature development, and terminal operations. Our evaluation of 14 representative models reveals substantial limitations: the highest recall and exact-match rate are only 48.97% and 10.00%, respectively. To address these limitations, we develop ObligationGuard using 40,000 synthetic training examples. ObligationGuard achieves 57.52% recall and an exact-match rate of 21.67%, surpassing all evaluated models on both metrics. We call on the community to incorporate obligation identification into the design and evaluation of future guard models to improve agent safety.

  • 4 authors
·
Oct 7

MMAE: A Massive Multitask Audio Editing Benchmark

We introduce MMAE, a Massive Multitask Audio Editing benchmark, serving as the first comprehensive evaluation testbed designed for general-purpose instruction-based audio editing. Spurred by the shift toward intelligent creation, interactive editing has rapidly expanded from visual domains, pioneered by models like Nano-banana 2 for images and Gemini-Omni for video, into audio. However, the current evaluation infrastructure lags severely, remaining highly fragmented and restricted to specific subdomains or basic operations. Unlike existing benchmarks that are limited in scope, MMAE extends to a broad spectrum of real-world scenarios, encompassing 7 distinct audio modalities, including sound, speech, music, and their mixtures. Furthermore, we establish a comprehensive taxonomy spanning 6 levels of task complexity, from basic modifications to multi-hop reasoning and multi-round editing, 2 levels of granularity, and 8 distinct operation types. Meticulously curated through human-agent collaboration, MMAE comprises 2,000 high-fidelity samples paired with a pioneering rubric-based evaluation framework. By decomposing free-form tasks into 17,741 verifiable criteria, this robust rubric-based paradigm enables a precise, multi-dimensional assessment of both instruction following and context consistency. Our extensive evaluation of leading models reveals that current systems remain far from achieving reliable edits. Strikingly, the Exact Match Rate (EMR) consistently falls below 5% and plummets to an absolute 0% in complex, mixed-modality tasks, exposing critical bottlenecks in precise execution and structural robustness. We hope MMAE will serve as a catalyst for future advances in the intelligent creation community, providing a clear diagnostic roadmap and establishing a standardized, long-lasting evaluation paradigm for next-generation audio editing systems.

  • 38 authors
·
Jun 4 3

Investigating Data Contamination in Modern Benchmarks for Large Language Models

Recent observations have underscored a disparity between the inflated benchmark scores and the actual performance of LLMs, raising concerns about potential contamination of evaluation benchmarks. This issue is especially critical for closed-source models and certain open-source models where training data transparency is lacking. In this paper we study data contamination by proposing two methods tailored for both open-source and proprietary LLMs. We first introduce a retrieval-based system to explore potential overlaps between evaluation benchmarks and pretraining corpora. We further present a novel investigation protocol named Testset Slot Guessing (TS-Guessing), applicable to both open and proprietary models. This approach entails masking a wrong answer in a multiple-choice question and prompting the model to fill in the gap. Additionally, it involves obscuring an unlikely word in an evaluation example and asking the model to produce it. We find that certain commercial LLMs could surprisingly guess the missing option in various test sets. Specifically, in the TruthfulQA benchmark, we find that LLMs exhibit notable performance improvement when provided with additional metadata in the benchmark. Further, in the MMLU benchmark, ChatGPT and GPT-4 demonstrated an exact match rate of 52\% and 57\%, respectively, in guessing the missing options in benchmark test data. We hope these results underscore the need for more robust evaluation methodologies and benchmarks in the field.

  • 5 authors
·
Nov 16, 2023

RANGER -- Repository-Level Agent for Graph-Enhanced Retrieval

General-purpose automated software engineering (ASE) includes tasks such as code completion, retrieval, repair, QA, and summarization. These tasks require a code retrieval system that can handle specific queries about code entities, or code entity queries (for example, locating a specific class or retrieving the dependencies of a function), as well as general queries without explicit code entities, or natural language queries (for example, describing a task and retrieving the corresponding code). We present RANGER, a repository-level code retrieval agent designed to address both query types, filling a gap in recent works that have focused primarily on code-entity queries. We first present a tool that constructs a comprehensive knowledge graph of the entire repository, capturing hierarchical and cross-file dependencies down to the variable level, and augments graph nodes with textual descriptions and embeddings to bridge the gap between code and natural language. RANGER then operates on this graph through a dual-stage retrieval pipeline. Entity-based queries are answered through fast Cypher lookups, while natural language queries are handled by MCTS-guided graph exploration. We evaluate RANGER across four diverse benchmarks that represent core ASE tasks including code search, question answering, cross-file dependency retrieval, and repository-level code completion. On CodeSearchNet and RepoQA it outperforms retrieval baselines that use embeddings from strong models such as Qwen3-8B. On RepoBench, it achieves superior cross-file dependency retrieval over baselines, and on CrossCodeEval, pairing RANGER with BM25 delivers the highest exact match rate in code completion compared to other RAG methods.

  • 4 authors
·
Sep 26, 2025

CoEdPilot: Recommending Code Edits with Learned Prior Edit Relevance, Project-wise Awareness, and Interactive Nature

Recent years have seen the development of LLM-based code generation. Compared to generating code in a software project, incremental code edits are empirically observed to be more frequent. The emerging code editing approaches usually formulate the problem as generating an edit based on known relevant prior edits and context. However, practical code edits can be more complicated. First, an editing session can include multiple (ir)relevant edits to the code under edit. Second, the inference of the subsequent edits is non-trivial as the scope of its ripple effect can be the whole project. In this work, we propose CoEdPilot, an LLM-driven solution to recommend code edits by discriminating the relevant edits, exploring their interactive natures, and estimating its ripple effect in the project. Specifically, CoEdPilot orchestrates multiple neural transformers to identify what and how to edit in the project regarding both edit location and edit content. When a user accomplishes an edit with an optional editing description, a Subsequent Edit Analysis first reports the most relevant files in the project with what types of edits (e.g., keep, insert, and replace) can happen for each line of their code. Next, an Edit-content Generator generates concrete edit options for the lines of code, regarding its relevant prior changes reported by an Edit-dependency Analyzer. Lastly, both the Subsequent Edit Analysis and the Edit-content Generator capture relevant prior edits as feedback to readjust their recommendations. We train our models by collecting over 180K commits from 471 open-source projects in 5 programming languages. Our extensive experiments show that CoEdPilot can well predict the edits (i.e., predicting edit location with an accuracy of 70.8%-85.3%, and the edit content with an exact match rate of 41.8% and BLEU4 score of 60.7)...

  • 9 authors
·
Aug 2, 2024

RTLRepoCoder: Repository-Level RTL Code Completion through the Combination of Fine-Tuning and Retrieval Augmentation

As an essential part of modern hardware design, manually writing Register Transfer Level (RTL) code such as Verilog is often labor-intensive. Following the tremendous success of large language models (LLMs), researchers have begun to explore utilizing LLMs for generating RTL code. However, current studies primarily focus on generating simple single modules, which can not meet the demands in real world. In fact, due to challenges in managing long-context RTL code and complex cross-file dependencies, existing solutions cannot handle large-scale Verilog repositories in practical hardware development. As the first endeavor to exclusively adapt LLMs for large-scale RTL development, we propose RTLRepoCoder, a groundbreaking solution that incorporates specific fine-tuning and Retrieval-Augmented Generation (RAG) for repository-level Verilog code completion. Open-source Verilog repositories from the real world, along with an extended context size, are used for domain-specific fine-tuning. The optimized RAG system improves the information density of the input context by retrieving relevant code snippets. Tailored optimizations for RAG are carried out, including the embedding model, the cross-file context splitting strategy, and the chunk size. Our solution achieves state-of-the-art performance on public benchmark, significantly surpassing GPT-4 and advanced domain-specific LLMs on Edit Similarity and Exact Match rate. Comprehensive experiments demonstrate the remarkable effectiveness of our approach and offer insights for future work.

  • 5 authors
·
Apr 11, 2025

Interplay between thermal and compositional gradients decides the microstructure during thermomigration: a phase-field study

The presence of thermal gradients in alloys often leads to non-uniformity in concentration profiles, which can induce the thermomigration of microstructural features such as precipitates. To investigate such microstructural changes, we present a phase-field model that incorporates coupling between concentration and thermal gradients. First, we simulated the evolution of non-uniform concentration profiles in the single-phase regions of Fe-C and Fe-N alloy systems due to imposed thermal gradients. To validate our model with the classical experiments performed by Darken and Oriani, we studied the evolution of spatially varying concentration profiles where thermal gradients encompass single-phase and two-phase regions. We developed a parameterized thermodynamic description of the two-phase region of a binary alloy to systematically study the effect of interactions between chemically-driven and thermal gradient-driven diffusion of solute on the evolution of precipitates. Our simulations show how thermal gradient, precipitate size, and interparticle distance influence the migration and associated morphological changes of precipitates. The composition profiles and migration rates obtained from single-particle simulations show an exact match with our analytical model. We use twoparticle simulations to show conditions under which thermomigration induces the growth of the smaller particle and shrinkage of the larger one in contrast to the isothermal Ostwald ripening behavior. Our multiparticle simulations show similar behavior during coarsening. Moreover, in the presence of a thermal gradient, there is a shift in the center of mass of the precipitates towards the high-temperature region. Thus, our study offers new insights into the phenomena of microstructure evolution in the presence of thermal gradient.

  • 4 authors
·
Jun 2, 2024

LATTE: Improving Latex Recognition for Tables and Formulae with Iterative Refinement

Portable Document Format (PDF) files are dominantly used for storing and disseminating scientific research, legal documents, and tax information. LaTeX is a popular application for creating PDF documents. Despite its advantages, LaTeX is not WYSWYG -- what you see is what you get, i.e., the LaTeX source and rendered PDF images look drastically different, especially for formulae and tables. This gap makes it hard to modify or export LaTeX sources for formulae and tables from PDF images, and existing work is still limited. First, prior work generates LaTeX sources in a single iteration and struggles with complex LaTeX formulae. Second, existing work mainly recognizes and extracts LaTeX sources for formulae; and is incapable or ineffective for tables. This paper proposes LATTE, the first iterative refinement framework for LaTeX recognition. Specifically, we propose delta-view as feedback, which compares and pinpoints the differences between a pair of rendered images of the extracted LaTeX source and the expected correct image. Such delta-view feedback enables our fault localization model to localize the faulty parts of the incorrect recognition more accurately and enables our LaTeX refinement model to repair the incorrect extraction more accurately. LATTE improves the LaTeX source extraction accuracy of both LaTeX formulae and tables, outperforming existing techniques as well as GPT-4V by at least 7.07% of exact match, with a success refinement rate of 46.08% (formula) and 25.51% (table).

  • 5 authors
·
Sep 21, 2024

Kencorpus: A Kenyan Language Corpus of Swahili, Dholuo and Luhya for Natural Language Processing Tasks

Indigenous African languages are categorized as under-served in Natural Language Processing. They therefore experience poor digital inclusivity and information access. The processing challenge with such languages has been how to use machine learning and deep learning models without the requisite data. The Kencorpus project intends to bridge this gap by collecting and storing text and speech data that is good enough for data-driven solutions in applications such as machine translation, question answering and transcription in multilingual communities. The Kencorpus dataset is a text and speech corpus for three languages predominantly spoken in Kenya: Swahili, Dholuo and Luhya. Data collection was done by researchers from communities, schools, media, and publishers. The Kencorpus' dataset has a collection of 5,594 items - 4,442 texts (5.6M words) and 1,152 speech files (177hrs). Based on this data, Part of Speech tagging sets for Dholuo and Luhya (50,000 and 93,000 words respectively) were developed. We developed 7,537 Question-Answer pairs for Swahili and created a text translation set of 13,400 sentences from Dholuo and Luhya into Swahili. The datasets are useful for downstream machine learning tasks such as model training and translation. We also developed two proof of concept systems: for Kiswahili speech-to-text and machine learning system for Question Answering task, with results of 18.87% word error rate and 80% Exact Match (EM) respectively. These initial results give great promise to the usability of Kencorpus to the machine learning community. Kencorpus is one of few public domain corpora for these three low resource languages and forms a basis of learning and sharing experiences for similar works especially for low resource languages.

  • 6 authors
·
Aug 25, 2022

Character-Level Transformer for Tajik-Persian Transliteration with a Parallel Lexical Corpus

This study addresses automatic transliteration from Tajik (Cyrillic script) to Persian (Perso-Arabic script). We present a curated, lexicographically verified parallel corpus of 52,152 Tajik--Persian words and short phrases, compiled from printed dictionaries, encyclopedic sources, and manually verified online resources. To the best of our knowledge, this is one of the largest publicly available word-level corpora for Tajik--Persian transliteration. Using this corpus, we train a character-level sequence-to-sequence Transformer model and evaluate it using Character Error Rate (CER) and exact-match accuracy. The Transformer achieves a CER of 0.3216 and an exact-match accuracy of 0.3133, outperforming both dictionary-based rule-based and recurrent neural baselines. With beam search (k=3), performance further improves to CER 0.3182 and accuracy 0.3215. We describe the data collection and preprocessing pipeline, model architecture, and experimental protocol, and report a part-of-speech analysis showing performance differences across lexical categories. All preprocessing scripts, deterministic splits into training, validation, and test sets, and training configurations are released to support reproducibility and further research on Tajik and related Persian dialects. The corpus supports research in character-level transliteration, cross-script NLP, and lexicographic applications.

  • 1 authors
·
May 8

VoiceCodeBench: Evaluating Exact Structured-Token Recovery in Automatic Speech Recognition

Automatic speech recognition (ASR) systems are commonly evaluated with word error rate (WER), yet many voice workflows depend on exact written values for identifiers, paths, and measured quantities. A transcript can appear fluent and achieve low WER while corrupting a value that a downstream system must parse, store, or execute. We introduce VoiceCodeBench, a benchmark for evaluating exact structured-token recovery in English ASR. It contains 300 human-recorded workplace segments spanning eight workflow domains and 1,482 audited target entities across 26 entity types, each with a canonical written form recoverable from the audio. Under a raw-audio-only protocol, systems receive audio bytes without additional context or metadata. Alongside WER, we evaluate Canonical Token/Entity Match (CTEM), Task Success Rate (TSR), and per-type exact recovery. Across 12 baseline ASR systems, lower WER generally corresponded to better structured-token recovery but did not fully determine it: Spearman correlations were -0.73 for both WER versus CTEM and WER versus TSR. The strongest baseline by TSR reached only 68.7%, leaving nearly one third of recordings with at least one unrecovered workflow-critical value. These results show that entity-sensitive metrics are needed to assess whether ASR output preserves exact values that production systems must parse, route, store, compare, or execute.

  • 7 authors
·
Aug 27

OTAP: Structure-Aware Optimal Transport for Evaluating Planning and Execution in Agent Trajectories

Large language model agents solve tasks by generating trajectories that interleave planning, tool calls, and intermediate results. Current evaluation metrics reduce such a trajectory to a binary success flag, compare it against a reference by exact matching, or delegate judgment to another language model. A success flag cannot distinguish a sound solution from one that succeeds by luck, and says nothing about why a failed run went wrong. Exact matching penalizes plans that are valid but reordered or decomposed differently from the reference. We reframe trajectory evaluation as a distance between the agent's execution graph and a set of valid solution graphs, and instantiate it via an unbalanced fused Gromov-Wasserstein transport problem over attributed dependency graphs. The resulting score, termed OTAP (Optimal Transport for Agentic Planning), is a pseudo-metric that is provably invariant to dependency-preserving reorderings and has bounded sensitivity to redundant steps. Its unbalanced marginals handle missing or hallucinated steps without forcing a match, and its soft coupling accommodates variation in plan granularity. On controlled perturbations and three public benchmarks, OTAP separates valid from invalid trajectories in a regime where semantics-only metrics score below chance. Its advantage tracks the fidelity of the dependency graph: largest where edges follow from operator semantics, smallest where they are inferred from free text. Where a formal verifier exists, strict surface metrics predict validity better than OTAP does, which places OTAP in open-ended domains where no verifier is available.

  • 3 authors
·
Jul 31

Looking Complete Is Not Being Grounded: Evaluating LLM Sitreps Against Human-Authored Situation Reports

Evaluation for automatic crisis reporting should measure what practitioners need, not what is convenient to compare. Through LLM-assisted, author-approved content analysis, we derive a 15-slot information schema from professional humanitarian situation reports using qualitative content analysis, then evaluate 121 human and 261 LLM-generated reports across five disasters and three hazard types. Under google/gemini-3.7-flash on Cyclone Idai, the best machine arm is nominally above the empirical human ceiling (0.827 vs. 0.821), but the difference is statistically indistinguishable from zero (95\% bootstrap CI [-.049,.059]; permutation p=.852), and the ordering does not replicate under openai/gpt-5.6-luna. On a blind 100-item Idai sample, the author agrees with the Gemini judge at 0.910 accuracy and unweighted Cohen's κ=.799. All four judges clear the plan's raw-agreement criterion of .7; only Gemini reaches .7 on chance-corrected κ. Across the three primary events, machine reports reproduce 5.4-11.0% of same-day official figures. Across those same events, raw exact-number matching links 14.3-19.6% of official figures and 85.4-90.1% of machine figures to the social stream; manual audit confirms 45/50 declared absences, 4/25 official matches, and 24/25 machine matches semantically. The issuing-provenance gap is positive in all 12 observed event-judge cells: report date, organisation, series, and authority were not supplied at professional-report coverage in any cell. On the three primary events, ROUGE-L is uninformative, BERTScore is negatively associated with completeness after multiplicity correction, and a generic LLM judge changes sign.

  • 1 authors
·
Oct 13 2

Same Agent, Different Answers: A Repeat-Aware Audit of Corpus-Induced Answer Churn in Retrieval-Augmented QA

A retrieval-augmented QA system can return different answers after an index expansion even when its requested model identifier, prompt, retrieval policy, evidence depth, rendering, and exposed generation controls are held fixed. Aggregate accuracy may hide these changes when gains and losses cancel, while ordinary generation variability makes one-shot comparisons overstate update effects. We call the hidden phenomenon accuracy-blind answer churn and introduce the Snapshot Compatibility Audit, which estimates excess answer churn by subtracting same-snapshot repeat disagreement from cross-snapshot disagreement. We instantiate it by expanding one frozen FineWeb prefix from one to seven shards. In a preregistered 400-question Natural Questions study, normalized-exact and blinded-semantic excess churn are 6.44 and 10.25 percentage points while exact-match accuracy changes by only -1.50 points. A post-hoc analysis finds repeat-stable semantic flips on 40/400 questions. A separately preregistered 200-question TriviaQA study yields smaller, directionally consistent excess churn while exact-match accuracy moves in the opposite direction. An outcome-blind post-hoc 100-question subset replication with a second DeepSeek generator and serving configuration finds 8.75 pp of semantic excess churn even as exact match rises by 3.00 percentage points. Answer-level compatibility can therefore fail without a conspicuous or consistently directed utility shift. Retrieval-augmented releases should audit compatibility alongside utility.

Infini-gram mini: Exact n-gram Search at the Internet Scale with FM-Index

Language models are trained mainly on massive text data from the Internet, and it becomes increasingly important to understand this data source. Exact-match search engines enable searching in large text corpora -- counting string appearances and retrieving the enclosing documents -- yet the high storage overhead hinders their application on Internet-scale data. We present Infini-gram mini, an efficient and scalable system that can make petabyte-level text corpora searchable. Based on the FM-index data structure (Ferragina and Manzini, 2000), which simultaneously indexes and compresses text, our system creates indexes with size only 44% of the corpus. Infini-gram mini greatly improves upon the best existing implementation of FM-index in terms of indexing speed (18times) and memory use during both indexing (3.2times reduction) and querying (down to a negligible amount). We index 46TB of Internet text in 50 days with a single 128-core CPU node (or 19 hours if using 75 such nodes). We show one important use case of Infini-gram mini in a large-scale analysis of benchmark contamination. We find several core LM evaluation benchmarks to be heavily contaminated in Internet crawls (up to 40% in SQuAD), which could lead to overestimating the capabilities of language models if trained on such data. We host a benchmark contamination bulletin to share the contamination rate of many core and community-contributed benchmarks. We also release a web interface and an API endpoint to serve general search queries on Infini-gram mini indexes.

  • 5 authors
·
Jun 13, 2025 3

Scientific Agents: Evaluating Profession-Specific System Prompts on Scientific Tasks

Detailed profession-specific system prompts raise token use and estimated cost per response without a consistent accuracy gain. We evaluate Scientific Agents, an open-source corpus of 503 profession-specific AGENTS.md profiles, with Gemini 3.8 Flash via OpenRouter in the Pi agent harness. We compare matched profiles with four controls: a minimal baseline ("You are a helpful assistant"), the profile's opening role sentence, a generic scientific rigor guide, and a profile from an unrelated domain. Across nine text-based science benchmarks (4,531 sampled questions, 100 matched profiles), 4,488 items completed all five conditions after API-error retries, scored with automated, rule-based grading. The average profile-baseline accuracy difference is -0.6 percentage points (95% bootstrap interval [-1.5, +0.2] across fixed tasks), and no benchmark shows a statistically clear improvement. Matched profiles produced 1.5-2.3 times as many output tokens and cost 2.2-4.5 times more per successful call. On 60 tool-using BioMysteryBench bioinformatics problems (three runs each for baseline and profile), mean solve rates were 46.7% with the profile and 56.7% at baseline, a difference of -10.0 percentage points (95% interval [-16.7, -3.3]) driven by more frequent token- and time-limit stops under the profile. Longer prompts had one unexpected operational advantage: on SuperGPQA, frequent provider API drops left the short baseline with a correct first-pass answer on only 54.0% of items, against 71.6% with the profile. Generic and mismatched prompts were about as reliable, so this gain comes from prompt length or formatting rather than domain expertise. For the tested model and tasks, loading full profession profiles by default does not improve accuracy and costs considerably more; whether selective retrieval of profile sections or open-ended scientific tasks would change this remains to be tested.

  • 1 authors
·
Sep 6

D2S-FLOW: Automated Parameter Extraction from Datasheets for SPICE Model Generation Using Large Language Models

In electronic design, engineers often manually search through extensive documents to retrieve component parameters required for constructing SPICE models, a process that is both labor-intensive and time-consuming. To address this challenge, we present an automated framework called D2S-FLOW that leverages large language models (LLMs) to extract electrical parameters from datasheets and generate SPICE models with high precision and efficiency, significantly reducing the need for manual intervention. Unlike traditional RAG systems, D2S-FLOW employs a workflow to enhance precision in handling unstructured documents and inconsistent naming conventions through three innovative mechanisms: Attention-Guided Document Focusing (AGDF), Hierarchical Document-Enhanced Retrieval (HDER), and Heterogeneous Named Entity Normalization (HNEN). AGDF narrows retrieval to user-selected documents, HDER utilizes document structure for precise parameter localization, and HNEN standardizes terminology via semantic inference. Experimental results demonstrate that the framework achieves an Exact Match (EM) of 0.86, an F1 score of 0.92, and an Exact Correctness (EC) of 0.96, outperforming the strongest baseline by 19.4%, 5.7%, and 13.1%, respectively. Additionally, it reduces API token consumption by 38% and minimizes the irrelevant information ratio to 4%, showcasing substantial improvements in resource efficiency. This research provides an effective automated solution for circuit design.

  • 3 authors
·
Feb 23, 2025

Finding Duplicates in 1.1M BDD Steps: cukereuse, a Paraphrase-Robust Static Detector for Cucumber and Gherkin

Behaviour-Driven Development (BDD) suites accumulate step-text duplication whose maintenance cost is established in prior work. Existing detection techniques require running the tests (Binamungu et al., 2018-2023) or are confined to a single organisation (Irshad et al., 2020-2022), leaving a gap: a purely static, paraphrase-robust, step-level detector usable on any repository. We fill the gap with cukereuse, an open-source Python CLI combining exact hashing, Levenshtein ratio, and sentence-transformer embeddings in a layered pipeline, released alongside an empirical corpus of 347 public GitHub repositories, 23,667 parsed .feature files, and 1,113,616 Gherkin steps. The step-weighted exact-duplicate rate is 80.2 %; the median-repository rate is 58.6 % (Spearman rho = 0.51 with size). The top hybrid cluster groups 20.7k occurrences across 2.2k files. Against 1,020 pairs manually labelled by the three authors under a released rubric (inter-annotator Fleiss' kappa = 0.84 on a 60-pair overlap), we report precision, recall, and F1 with bootstrap 95 % CIs under two protocols: the primary rubric and a score-free second-pass relabelling. The strongest honest pair-level number is near-exact at F1 = 0.822 on score-free labels; the primary-rubric semantic F1 = 0.906 is inflated by a stratification artefact that pins recall at 1.000. Lexical baselines (SourcererCC-style, NiCad-style) reach primary F1 = 0.761 and 0.799. The paper also presents a CDN-structured critique of Gherkin (Cognitive Dimensions of Notations); eight of fourteen dimensions are rated problematic or unsupported. The tool, corpus, labelled pairs, rubric, and pipeline are released under permissive licences.

  • 3 authors
·
Apr 21 1

Skip the Compressor: Coding Agents Fix More Issues When They Pick Whole Files

A coding agent works on repositories that can be far larger than its context window can hold. Prompt compressors and retrievers offer solutions to this problem, but patching an issue needs two things at once: the files that must change, and their code exactly as written, since the model must reproduce the lines it edits. Retrievers rank code by its similarity to the issue rather than by whether it must change, and compressors delete text to fit the budget. We show that at an equal token budget, an agent resolves more issues when the budget goes to choosing which files to show verbatim, as our simple selector CQ2-CARE does, than when it goes to pruning tokens or retrieving chunks. We fit every context to the same per-issue budget on 300 recent SWE-rebench issues and grade each patch with the benchmark's tests. With GPT-5.6 and Claude Opus 5, CQ2-CARE resolves 43.7% and 46.7% of the issues, 6.3 points more than the best of three compressors with either model, and 9.3 and 12.7 points more than the best of three retrievers. CQ2-CARE even resolves more issues than two methods trained for issue resolution, SweRank (by 0.7 and 1.7 points) and SWE-Pruner (by 2.0 and 7.3). Our analysis shows the defects of the baseline methods: retrievers often miss the files to edit, and compressors keep them but alter their text, so fewer patches apply. Exact-file selection needs no compression model and shows every line as the repository has it, so the model can quote the code it edits.

Same Layout, Different Count: What Number Words Do Past Ten in Text-to-Image Models

Ask SD3.5-Medium for twelve of an object and then, from the same starting noise, for twenty-five. The picture keeps its layout, but the count changes about as much as it would under a new random seed, and in no consistent direction. We study what number words do once a request passes about ten objects, in five open text-to-image models. Holding the object and the starting noise fixed, we change one thing at a time, and people who never see the request count 3,088 of the images. Every model follows the number for small crowds, but only up to a point, its reach: SD3.5-M and FLUX.1-dev stop following it soon after ten, while FLUX.2 [klein] and Qwen-Image follow it loosely up to twenty-five. Beyond SD3.5-M's reach, the coarse part of the starting noise fixes the composition of the picture but not how many objects it holds: redrawing only the noise finer than an eighth of the image keeps the layout as well as a new number does, and already moves the count two-thirds as far. The number word still reaches the denoiser and acts in its first steps, but nothing we tried gives it a clear direction. Shown instead as a noised grid in the starting latent, the count is drawn: people find it in 59% of images, against 4% when it is only told. Past ten, an exact-match score barely separates a model that follows the number loosely from one that ignores it, and a relative-error score can even prefer a constant guess of sixteen. A same-noise test with a few requests per object tells them apart.

  • 2 authors
·
Nov 4

A Frozen 12B Beats Frontier Models on Verified Work: 100% Accuracy, 0 Tokens, Bit-Exact, Forever

Improving a language model today means retraining it: enormous compute, a new opaque model each cycle, non-deterministic output. We take the opposite path: the model stays frozen, and a persistent memory of verified solutions grows beside it. Once a problem family is solved and has passed an independent verification step that never consults the answer key, every new instance of that family is answered at zero generation tokens, bit-exact, deterministically. Across 180 fresh instances spanning nine problem families, four architectures from four vendors - dense and mixture-of-experts - each score 180/180 at zero generation tokens per answer: execution-bound capability decoupled from parameter scaling. A negative control attributes the capability fully to the memory: emptied, it solves nothing. The same verify-before-store contract holds for open-ended reasoning: 88/88 consistency-gated acceptances across all four models, machine-checked formal proof, and reasoning-method transfer at 77/80. Memory selection takes 1.4 microseconds; a full reuse completes in 6-23 ms at 36 mWh. Approximate similarity retrieval selects the wrong item 94.3% of the time on a 4,500-item verified store where exact addressing makes zero errors. The store also serves as working context at a scale no shipped engine matches: a 6,000,000-token movable window on a single 46 GB GPU at flat memory, where vLLM stops at 30,399 tokens and SGLang silently truncates past 32,000. On published benchmarks, frontier models remain far ahead of any 12B at raw from-scratch reasoning; on everything this system has solved and verified, the comparison inverts: a frontier API call pays a fresh generation pass on every query, forever, while verified reuse costs zero tokens and returns the identical bits every time. A public testbench with free, rate-limited access accompanies this report: https://huggingface.co/proxy/corbenic-galahad-bench.hf.space

Corbenci Corbenic
·
Jul 25 4

Keyword Harnesses Fail Open: A Cheap Diagnostic Ladder for Tool-Use Claims in Small Language Models

Keyword-matching benchmarks can credit small models for tool use they never perform. We document such a false positive in a matched-architecture pair of Spanish security language models and propose a ladder of strict, cheap diagnostics. A 661.6M parameter model (approx. 65% code/technical text; no dedicated SFT) and a 1,109M model (web-heavy multi-phase curriculum; 6B-token tool-SFT) share decoder, tokenizer, and special tokens, scoring almost identically on lenient tool-use metrics (B4: 0.660 vs. 0.650). Verbatim-reproduction checks on training examples separate them completely: the 600M emits valid tool calls with generalized arguments on 6/6 examples; the 1B does so on 0/6 across checkpoints. A first-token probe localizes the 1B's failure to a missing prior (prob. 10^{-4}--10^{-5} on <|tool_call|>), which was erased by its web-heavy training phase. A targeted SFT recipe (diverse corpus, 5x higher learning rate, 2,202 steps, ~3.3 GPU-hours) repairs the 1B using three orders of magnitude fewer tokens than the failed phase. On all 269 corpus rows, valid emission rises from 0.100 to 0.959 (600M: 0.926). On 238 unseen prompts, the repaired 1B passes 0.536 vs. the 600M's 0.428 (p = 0.004). Embedding-drift checks show the repair did not move the trigger token's tied embedding (97.7% of the bf16 table remains bit-identical), meaning changes live in the surrounding network. Both models over-trigger, rarely answering negative prompts without a call (0.09 for 600M, 0.17 for repaired 1B). Factorial analyses confirm all repair configurations install the format, though suppression benefits from a diverse corpus remain a hypothesis due to seed sensitivity. This cheap diagnostic ladder costs minutes of CPU time and should gate tool-use claims on small models.

  • 1 authors
·
Sep 30 3

PyMETA: Evaluating Student Code Diagnosis on and Beyond the First Execution Error

Large language models can diagnose a student program from its code, problem statement, and reference solution. Evaluating this ability requires a clear definition of what counts as the correct diagnosis. We introduce PyMETA, a Python error dataset with 48,646 student submissions to 155 problems. Every submission has a single label for the first execution error identified by an Online Judge, or No Error when the program passes all tests. A targeted subset of 97 submissions also has expert labels collected through iterative repair and re-execution. The taxonomy has three levels; its most detailed level contains 14 labels, including No Error, Logic Error, named Python exceptions, and an Other Errors category. We evaluate two finetuned models and two groups of prompted LLMs: four earlier models and four recent models. When evaluated against the first execution error, the recent prompted models reach 87.5--93.8% macro F1, above the strongest finetuned baseline at 80.6%. This is the opposite of the comparison obtained with the earlier prompted models. On the 97-item expert subset, however, exact-set match is only 43.3--48.5%, although sample F1 is about 79--81%. Output format also matters. On the same 45 audited submissions whose expert label sets do not contain Logic Error, none of the four recent models returns that label under single-error prompting, but 46.7--57.8% of their multi-error outputs include it, usually after an explicit-error label. The results show that model rankings and claims about label bias depend on the meaning of the gold label, the number of labels a model may return, and the scoring rule.

  • 6 authors
·
Sep 14

Enrich-Retrieve-Rank: Scaling Capability Discovery Beyond In-Context Routing

Agent ecosystems now include thousands of MATS components (Models, Agents, Tools, and Skills), yet their discovery still relies on in-context routing. These systems read a registry (names, hints, or descriptions, as context budget permits), pick a candidate, invoke it, and retry on failure. This pattern degrades with scale, and registries are growing fast. We recast capability discovery as search over a registry by defining an offline enrichment step that turns sparse metadata into searchable profiles, and an online retrieve-then-rank pipeline that returns a ranked shortlist without invoking any candidates online. We show that from N=10 to 7,278 capabilities, in-context routing's top-1 accuracy (Match@1) collapses (0.85 to 0.12), while retrieve-then-rank degrades more gently (0.81 to 0.39) because its reranker still ranks the right capability first 0.70-0.87 of the time once retrieval finds it. In the Nova Micro sweep, the crossover is around N=500. We compare against two in-context baselines. Full-Ctx puts the whole registry in the prompt and asks the LLM to pick. Search&Pick gives the LLM a search tool to narrow candidates before it picks. At full scale the pipeline leads Search&Pick by 6.5 percentage points (pp) on Match@1 at about half the cost. It reduces cost 70x versus Full-Ctx. We use a fixed configuration (same enrichment, retriever, and scorer weights) across agent, tool, and skill registries. The pipeline runs in production as the default capability-discovery layer of a large-scale multi-agent platform.

  • 3 authors
·
Aug 23

How Optimal is Greedy Decoding for Extractive Question Answering?

Fine-tuned language models use greedy decoding to answer reading comprehension questions with relative success. However, this approach does not ensure that the answer is a span in the given passage, nor does it guarantee that it is the most probable one. Does greedy decoding actually perform worse than an algorithm that does adhere to these properties? To study the performance and optimality of greedy decoding, we present exact-extract, a decoding algorithm that efficiently finds the most probable answer span in the context. We compare the performance of T5 with both decoding algorithms on zero-shot and few-shot extractive question answering. When no training examples are available, exact-extract significantly outperforms greedy decoding. However, greedy decoding quickly converges towards the performance of exact-extract with the introduction of a few training examples, becoming more extractive and increasingly likelier to generate the most probable span as the training set grows. We also show that self-supervised training can bias the model towards extractive behavior, increasing performance in the zero-shot setting without resorting to annotated examples. Overall, our results suggest that pretrained language models are so good at adapting to extractive question answering, that it is often enough to fine-tune on a small training set for the greedy algorithm to emulate the optimal decoding strategy.

  • 4 authors
·
Aug 12, 2021

MATCHA: Matching Text via Contrastive Semantic Alignment

Reliable evaluation is essential for understanding large language model (LLM) performance, yet today's go-to metrics, namely token-overlap scores (e.g., ROUGE) and embedding-based measures (e.g., BERTScore), often misjudge semantic similarity of documents. Our study shows that both token-overlap metrics and embedding-based metrics routinely assign nearly identical scores to texts that directly contradict each other, thereby potentially masking fundamental errors. We introduce MATCHA, an automatic metric that jointly rewards semantic agreement with a reference and penalizes contradictions. MATCHA employs a dual-view perspective that measures (i) proximity to the gold text and (ii) distance from an adversarially generated counterfactual contradiction. In eight public benchmarks, MATCHA outperforms popular metrics, compared with human annotations on question-answering, image caption generation, natural language inference, summarization, and semantic textual similarity tasks. On the TruthfulQA dataset (i.e., a dataset without a training set, where no embedding-based metrics could locally train on), this improvement in terms of matching texts with a reference reaches 18.38% over ROUGE-L and 20.82% over BERTScore. Both quantitative comparison and qualitative human assessments confirm the efficacy and validity of MATCHA and uncover fundamental weaknesses in pre-existing metrics. Compared with 23 embedding models, including top state-of-the-art ones, used as a metric similar to BERTScore, MATCHA remains the most accurate in distinguishing correct from incorrect statements solely based on a reference. Our code and metric are publicly available (https://github.com/Siran-Li/MATCHA).

  • 4 authors
·
May 25

Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting Turns a Frozen Small Model into a Verified-Knowledge Flywheel

We report a way to make a frozen small language model both more capable and dramatically cheaper at once, without changing any weights. Verified knowledge is deposited once as a byte-exact key-value (KV) state artifact and later restored, by graft, into a fresh inference context. The restore is bit-exact: under a pinned deterministic configuration, the grafted logits are byte-for-byte identical to a fresh computation (SHA-256 equality), with zero KL divergence and 100% argmax agreement over fifty samples. We show that own-position graft is the unique numerically exact operating point on a model with floating-point rotary encoding, and we verify byte-exactness on two model scales (12B, 31B) and two GPU targets, one through a pre-registered replay. On AIME 2025, a frozen Gemma-4-12B moves from 80.0% to 93.3% once a verified solution library is grafted, above its own 77.5% and its 31B sibling's 89.2% published anchors. On the recurring case, eight problems the base model never solves within a 401,026-token budget are answered from cached verified solutions in 61 total decode tokens, a factor of 6,574 fewer tokens and about 8,700x less energy; the capability claim proper rests on held-out transfer (7 of 7 at 31B). The same byte-exact store widens usable context from 32,768 to 2,854,766 tokens at zero extra accelerator memory, and moves byte-identical between machines of the same architecture. We describe the system at the behavior level; the engine is proprietary, and every reported number is backed by committed input and output hashes so the scoring can be re-checked without it.

Corbenci Corbenic
·
Jul 14 7

Misinformation Without Triggers: From Factual Answers to Downstream Decisions

Language models learn from web documents, some of them false, and false content can reach a model's answer to a factual question and the summaries and decisions that use it. Most data-poisoning studies add a trigger to the training data and activate it in the prompt. False documents can also change factual responses without any trigger, but we do not know whether the direct answer predicts the decision. In this work, we follow false content past the answer and find an audit gap between what a direct probe reports and what the model then does, comparing false training with matched truthful controls in a controlled decision task, Guess the Capital, where a fixed decoder turns factual answers into a scored card choice, and on a misleading claim from Facebook posts about the 2019--20 Australian bushfires. Across eight models at dose 1,000, direct injected-choice rates reach 95.8--100\%, while injected game choices increase by 1.7--14.4 percentage points over matched truthful training. The gap runs the other way too. Facts that pass the direct probe still push decisions toward the injected answer, and game accuracy drops further than those choices explain. Truthful correction brings the fact back but not the decisions built on it. We then look into the real-world bushfire case, models trained on the false posts say that people were arrested for arson even when they lose the inflated count, and in a count-by-wording factorial the misleading arrest wording produces arrest assertions even when the training count stays at 24. In simpler terms, a correct factual answer does not guarantee a correct decision, and losing the injected number does not remove the misleading story.

  • 2 authors
·
Oct 2

Tomayto, Tomahto. Beyond Token-level Answer Equivalence for Question Answering Evaluation

The predictions of question answering (QA)systems are typically evaluated against manually annotated finite sets of one or more answers. This leads to a coverage limitation that results in underestimating the true performance of systems, and is typically addressed by extending over exact match (EM) with pre-defined rules or with the token-level F1 measure. In this paper, we present the first systematic conceptual and data-driven analysis to examine the shortcomings of token-level equivalence measures. To this end, we define the asymmetric notion of answer equivalence (AE), accepting answers that are equivalent to or improve over the reference, and publish over 23k human judgments for candidates produced by multiple QA systems on SQuAD. Through a careful analysis of this data, we reveal and quantify several concrete limitations of the F1 measure, such as a false impression of graduality, or missing dependence on the question. Since collecting AE annotations for each evaluated model is expensive, we learn a BERT matching (BEM) measure to approximate this task. Being a simpler task than QA, we find BEM to provide significantly better AE approximations than F1, and to more accurately reflect the performance of systems. Finally, we demonstrate the practical utility of AE and BEM on the concrete application of minimal accurate prediction sets, reducing the number of required answers by up to x2.6.

  • 5 authors
·
Feb 15, 2022

Precision Is Not Faithfulness: Coverage-Aware Evaluation of Grounded Generation with a Complete Oracle

Reference-free faithfulness metrics verify each atomic claim a model makes against ground truth, and are increasingly used to evaluate grounded generation. We show they share a blind spot: they measure only precision -- are the stated claims supported? -- and therefore reward abstention, since a model can score near-perfect faithfulness by saying almost nothing. We make this measurable using Formula 1 telemetry, a domain where strategic ground truth is derived deterministically and, crucially, completely: for each decision we know the full set of facts that mattered. This completeness -- absent in open-domain faithfulness benchmarks -- lets us measure recall (coverage of the relevant facts) exactly, alongside precision. On a multilingual (EN/ES/PT) benchmark of 7,253 decision instances spanning 150 races, the most precise frontier model covers under half of the relevant facts and ranks last by F1, so requiring coverage reorders the systems; the same effect reappears in a second complete-oracle domain (NOAA weather forecasts). A prompt ablation shows the low coverage is not an under-prompting artifact: explicitly asking models to be thorough does not close the gap. We pair faithfulness with coverage into a single score, validate the metric (controlled perturbation; agreement across a model-free regex extractor and a cross-family LLM extractor, system-level Spearman 1.0), and give a verifier-guided generation method that improves precision and recall without references. We release the benchmark, structured annotations, metric, baselines, and an interactive demo.

  • 1 authors
·
Jun 7 6

Jev-as-a-Judge for RAG Claim Verification: Confidence-Gated Cascading Between a Non-Generative Verifier and a Frontier LLM

Retrieval-augmented generation asks one question at volume: does this evidence support this claim? The answer is one bit, yet most pipelines obtain it from a text generator and pay generation prices for the prose wrapped around it. We compare a non-generative decision model, which returns a typed value from a declared enumeration together with a native probability and no text, against a frontier LLM configured as an engineer would deploy it for binary classification. Both answer the same 495 human-labelled claims from LLM-AggreFact, so every comparison is within-item. We report three things. First, the two systems are statistically indistinguishable on the metric this task is conventionally scored by, a paired difference of -0.56 points, Jev minus Astra, at a 187-fold cost difference. Second, that tie carries almost no information about deployment: the systems disagree on 15.2% of claims, and the cheaper one admits 23.2% of the claims annotators marked unsupported against the frontier model's 14.1%. A component whose job is to catch unsupported claims should be judged on that rate, and the conventional headline hides it. Third, because the errors differ they compose: routing the claims the cheap model is confident about and escalating the rest reaches 74.10 at USD 2.92 per thousand claims against 73.83 at USD 8.93, with the threshold chosen out of fold. But a matched random-escalation control at in-sample thresholds, spending the same budget by coin flip, does as well, so what buys the saving is the escalation share and not the probability that picks what to escalate. Measured against the cheap model it is built from, the cascade costs 61-fold more for 0.83 points, 95% CI [-3.36, 4.86], and it is the more permissive guard of the two. Equal accuracy conceals unequal errors. All predictions, prompts, pricing and code are released.

  • 2 authors
·
Sep 29

The Commercial Tax: Rent-vs-Own Blind Spots in Multi-Hop Retrieval Benchmarks

Enterprises connect language models to their own data through retrieval. The benchmarks that rank multi-hop retrieval systems leave out two facts a buyer needs before a published number can be used: whether the retrieval backbone may be deployed commercially, and what it costs to build. On licensing: the field's dense-retrieval anchor, NV-Embed-v2, is licensed cc-by-nc-4.0. Of the four leading MuSiQue systems we audit (HippoRAG-2, PropRAG, SAG, KET-RAG), three depend on it for their best numbers and none says so. On performance: we measure thirteen embedders from eight makers on one identical MuSiQue harness with bootstrap confidence intervals throughout. Until mid-2026 there was a real commercial tax: the best commercially-licensed embedder trailed the anchor by 2.31 Recall@5 points (95% CI [0.91, 3.71], p=0.001). NVIDIA's Nemotron-3-Embed-8B, released 2026-07-16, has closed it: +0.24 at Recall@5 (95% CI [-0.94, +1.43], p=0.69), -0.58 at Recall@10 (p=0.28). It matches the anchor, does not beat it, and is the only entrant that is commercially licensed, free to self-host, and indistinguishable from the anchor; every other entrant meeting the first two conditions sits 5.2 to 14.6 points below. The durable finding is the paid-versus-free divide: API embedders charge per token on every re-index, self-hosted ones charge nothing. On cost: three of five audited systems (adding Microsoft's GraphRAG) do not disclose indexing cost, and the only published GraphRAG dollar figures span 11x inside one third-party paper (USD 2.30 vs USD 24.94 to index a 5.64 MB corpus once); extrapolated to 1 TB that undisclosed choice separates roughly USD 428K from $4.6M. Our cost model keeps one-time embedding apart from recurring answering: at 1 TB, embedding sits 7.5x-900x below graph construction, and a year of answering at 10,000 queries/day sits 350x or more below it.

  • 2 authors
·
Aug 16

Predictive Credit: Measuring What Scientific Explanations Add to Experimental Forecasts

Research agents explain planned experiments. We measure predictive credit with paired forecasts sharing an intervention, forecaster, and outcome while varying description, matched explanation, and donor context. Five checks track commitment, delivery, predictive gain, alignment, and known-signal uptake. Across 336 prospective states in controlled learning, 12 Tox21 endpoints, and 24 OpenML tasks, v5's frozen credit decision was inconclusive. Tox21's preregistered ROC AUC interval-score harm test was unmet (D-M=-.0026, 95 percent interval [-.0174, .0104]); OpenML's joint formation, point-equivalence, and repeatability rule was unmet. Matched point-accuracy gains over description remained unconfirmed, and Tox21/OpenML seed-donor intervals spanned zero. Under requested DeepSeek V4 Pro, matched and donor cards reduced secondary Tox21 drift by 64.5 and 59.1 percent. A DeepSeek V4 Flash replay raised matched point MAE from .01823 to .02020 and missed matched-donor interval-score equivalence. OpenML full-card assignment widened nominal 80 percent intervals by 21 percent, with 49.3 percent coverage versus 51.4 percent for description and content in 66/144 cards. Direct-text Flash delivered all 144 notes without detectable matched point-accuracy gain. A researcher-authored mechanism positive control lowered point MAE by 2.60 percentage points versus description. The protocol measures predictive credit for research-agent benchmarks and scientific forecasting; natural-explanation credit remained unconfirmed at the tested donor resolutions.

Prompt Design at Scale: How Format, Instruction Count, and Context Length Shape Instruction Adherence and Hallucination in Large Language Models

Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade. We report two controlled experiments crossing all three factors on one held, contamination-free synthetic corpus (the "Book of Veyra," 8,780 uniquely-named entities, deterministically regenerable from a fixed seed), evaluated across five models. Experiment 1 (960 calls/model) measures instruction-following decay as rule count N grows from 10 to 160, crossed with four formats and system-prompt vs. user-turn placement. Perfect-response rate collapses to zero by N=80 for every model, format, and placement. Placement produces effects at least as large as format at N=160 in most models, but the direction is model-specific. No model shows a reliable markdown advantage; one 35B model favors plain text instead. Experiment 2 (5,520 calls/model) measures recall accuracy, false-premise sycophancy, and absent-fact fabrication across a 2k-to-512k-token context ladder in the same four formats. Recall stays near ceiling through 64-128k tokens, then degrades sharply and format-dependently: one model's accuracy spread reaches 48 points at 128k tokens. Fabrication never occurs (0/5,760 probes), and sycophancy stays negligible (<=8.3%). What rises sharply near each model's context ceiling is outright refusal to answer (0% to 79-90%), distinct from sycophancy or fabrication. Neither pre-registered format ordering holds, and token overhead (+22% to +37% over plain text) further changes which format is preferable where accuracy spread is genuine. We release the full harness, corpus generator, and raw results (VeyraBench): https://github.com/iNetanel/veyrabench

  • 1 authors
·
Jul 20

Evaluating Commercial AI Chatbots as News Intermediaries

AI chatbots are rapidly shaping how people encounter the news, yet no prior study has systematically measured how accurately these systems, with their proprietary search integrations and retrieval-synthesis pipelines, handle emerging facts across languages and regions. We present a 14-day (February 9-22, 2026) evaluation of six AI chatbots (Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5 and GPT-4o mini) on 2,100 factual questions derived from same-day BBC News reporting across six regional services (US & Canada, Arabic, Afrique, Hindi, Russian, Turkish). The best systems achieve over 90% multiple-choice accuracy on questions about events reported hours earlier. The same systems, however, lose 11-13% under free-response evaluation, and 16-17% across the cohort. We further characterize three failure patterns. First, every model achieves its lowest accuracy on Hindi (79% vs. 89-91% elsewhere) and citations indicate an Anglophone retrieval bias (e.g., models answering Hindi queries cite English Wikipedia more than any Hindi outlet). Second, retrieval, not reasoning, failures drive over 70% of all errors. When models retrieve a correct source, they often extract the correct answer; the problem is to land on the right source in the first place. Third, models achieving 88-96% accuracy on well-formed questions drop to 19-70% when questions contain subtle false premises, with the most vulnerable model accepting fabricated facts 64% of the time. We also identify a detection-accuracy paradox: the best false-premise detector ranks second in adversarial accuracy (abstention rate), while a weaker detector ranks first, showing that premise detection and answer recovery are partially independent capabilities. Overall, these suggest that high accuracy can mask systematic regional inequity, near-total dependence on retrieval infrastructure, and vulnerability to imperfect queries real users pose.

  • 8 authors
·
May 20

Right Family, Wrong Skill: Benchmarking Risk Exposure in Agent Skill Retrieval

Agent skill libraries are becoming routable software assets: a retrieved skill can contribute instructions, scripts, resource bindings, and execution assumptions to an agent. This makes retrieval failures more specific than broad irrelevance. A system can find the right capability family yet expose the wrong same-capability representative. We study this failure as same-capability risk-exposure retrieval. Each benchmark unit pairs a helpful skill with a query-specific risky sibling that shares the capability family but differs on an execution-controlling contract, such as the required resource, precondition, procedure, or artifact. We introduce SameCapRisk-Bench, an auditable benchmark with 1,190 skill-risk units and 1,686 evaluation query cases: 694 marked-sibling units under public library pressure and 496 hard role-flip units where the same two skills swap helpful/risky roles across paired queries. The release records admission evidence, cue/leakage checks, source hashes, family relations, and fixed candidate pools. The benchmark reports helpful ranking together with harmful sibling rate (HSR@K), the top-K exposure of the marked risky sibling. On this benchmark, public SkillRouter, SkillRet, and R3-Skill retrieve helpful skills at high Recall@3 (0.848--0.888) but also expose marked risky siblings frequently (HSR@3 0.346--0.372). A fully public score-and-cluster pipeline lowers HSR@3 to 0.128--0.182, with Recall@3 of 0.713--0.776. Under a benchmark-trained reference scorer, public text-cluster and controlled resolvers reach HSR@3 0.012 and 0.007; the latter attains Recall@3 0.833. Skill retrieval should therefore report both capability matching and same-family risk exposure, with HSR serving as a targeted exposure certificate for fixed skill libraries.

  • 4 authors
·
Aug 19

The Structured Output Benchmark: A Multi-Source Benchmark for Evaluating Structured Output Quality in Large Language Models

Large Language Models are increasingly being deployed to extract structured data from unstructured and semi-structured sources: parsing invoices, medical records, and converting PDF documents to database entries. Yet existing benchmarks for structured output generation either focus on schema compliance alone, or evaluate value correctness within a single source domain. We introduce SOB (The Structured Output Benchmark), a multi-source benchmark spanning three source modalities: native text, images, and audio conversations. All models receive a text-normalized representation of their context regardless of source modality; this deliberate design isolates structured-output capability from raw vision or speech-processing quality, ensuring a fair, source-agnostic comparison. Our benchmark comprises 5,000 text evaluation records derived from multi-hop QA drawn from a 25,091-record full corpus, 209 image records from OCR-processed PDFs across seven document types including multi-column layouts, dense tables, scanned historical documents, small-print text, and mathematical typesetting, and 115 audio records from the AMI corpus. Each record pairs a natural-language question with a JSON schema that the model must follow and a ground-truth answer verified against the source context. We evaluate 21 frontier and open-weight models across three source domains and seven metrics. Our results reveal a consistent pattern: models achieve near-perfect schema compliance, yet the best Value Accuracy, measured by exact leaf-value match, reaches only 83.0% on text, 67.2% on images, and 23.7% on audio, where longer context makes extraction substantially harder. We release the dataset, evaluation pipeline, and all related code.

  • 4 authors
·
Apr 27

From Discovery to Decision: Finite-Budget Recoverability in LLM Voting

Voting over multiple LLM responses is a common primitive in test-time scaling and ensemble inference. Collecting more responses can expand the candidate pool and increase the chance that a correct answer is discovered. Under a fixed call budget, a discovered answer still needs to accumulate enough support within the remaining calls to become the final plurality winner, creating a discovery-to-decision gap. In this work, we characterize this gap through the realized vote state and remaining call budget. We derive a sharp recoverability threshold and show that, as sampling proceeds, the observed candidate set can only expand while the set of reachable endpoint winners can only contract, inducing a candidate-level conversion window. Under a specified iid response law, the same state yields exact finite-horizon endpoint probabilities. We further show that merging wrong-answer identities preserves single-call correctness and cannot improve plurality accuracy, and that the effect of redistributing wrong-answer probability depends on the realized vote state. Singleton reachability yields a gold-free exact locking certificate. For a known answer universe, its first trigger is the earliest prefix at which all admissible continuations yield the same fixed-budget output. Empirically, most discovered-but-unselected correct answers lose reachability only after discovery. In a controlled Word16 study, input permutation improves raw-plurality accuracy by 21.1 points with essentially unchanged single-call correctness. Exact locking saves 28-30% of calls at a 16-call budget while preserving every fixed-budget output.

  • 2 authors
·
Sep 30

LiveSearch: Questions as Executable Verifier Code for a Live Benchmark of Professional-Source Financial Search Agents

In finance, the answer that matters is usually the current one: prices, rates, and disclosures hold only as of a given day, so retrieving up-to-date facts is a core capability of a financial agent. Yet most financial benchmarks overlook timeliness. They fix each answer at annotation time, so the gold goes stale as markets move and, once leaked into training data, rewards recall rather than search and analysis. We introduce LiveSearch, a live benchmark that stores each question as executable verifier code: a typed program of calls to a professional financial data service, followed by a declarative extractor. The question is frozen, while its gold answer is determined by executing the program against professional financial data at evaluation time, and a daily job re-executes it with no human or language model in the loop. A memorized answer is thus likely outdated, and a model must reach the current one through tools and analysis. Construction is equally automatic: a language model only verbalizes programs enumerated from real tool responses and never chooses an answer. Under general web search, six contemporary tool-calling models reach 70.9% to 76.7% answer accuracy. Over three consecutive days, 70% of served questions change their gold value, and even under tolerance-aware matching, a literally reused earlier snapshot fails on a majority of questions within a single day. Per-question tracking uncovers dynamic shifts in correctness masked by aggregate accuracy.

  • 9 authors
·
Oct 27

Declarative Outcome-Conformant Synthesis: Exact, Closed-Form Specification Satisfaction and a Conformance Benchmark

We study a capability the dominant paradigm in synthetic tabular data does not provide: exact satisfaction of a declared analytical outcome with no source data. Imitation methods (copulas, GANs, diffusion) learn a real distribution and sample from it, and are judged on fidelity to real data. A large, practical class of needs is different: generating data with no source data ("cold start") that reproduces a declared outcome (a revenue curve, a churn rate, a group share) across a relational schema. Off-the-shelf imitation tools offer no interface for such targets, and no sampler can hit an exact aggregate, because sampling has variance. On a real public dataset, off-the-shelf learned synthesizers trained on that very data miss the declared monthly aggregate by 74 to 86 percent; a per-period steelman cuts the miss to about 19 percent and still cannot reach 0; a closed-form generator reaches exactly 0. We name this task outcome-conformant synthesis, argue its evaluation axis is conformance rather than fidelity, and show the two axes are orthogonal. We contribute: (1) a formal account showing a widely-used family of exact-aggregate generators is exactly conditional-sum sampling of a Gamma population (via Lukacs' characterization), with closed-form exactness, a closed-form marginal CV, and scale-invariance; a controlled experiment maps the boundary, enforcing the exact aggregate costs at most 0.006 in 1-Wasserstein distance to an arbitrary external marginal, the rest being shape-family mismatch; (2) SpecBench, to our knowledge the first benchmark to measure conformance to analytical outcomes for cold-start relational synthesis; and (3) a closed-form, deterministic reference system. Exact aggregation alone is trivial; the contribution is conformance jointly with closed-form marginals, integrity, determinism, and zero source data. We concede fidelity to imitation where real data exists.

  • 1 authors
·
Jun 6

Beyond Aggregate Calibration: Decomposing Income-Conditional Recall Disparities in Automated Credit Default Prediction

Data-centric curation pipelines frequently rely on model confidence scores to flag and filter noisy or mislabeled training instances. Evaluating this filtering convention on a large-scale consumer lending sample (LendingClub, N = 1,344,936) uncovers an underlying demographic asymmetry: high-income defaulters are disproportionately classified as label noise relative to low-income defaulters (Cramer's V approximately 0.03-0.07). Re-examining this behavior through the lens of equal opportunity [Hardt et al., 2016] reveals a far more severe discrepancy: a 16.86 percentage point gap in true positive rate (recall) between high- and low-income borrowers who ultimately defaulted. Implementing a sequential feature-blinding methodology allows us to isolate the drivers of this disparity across three distinct mechanisms: (1) direct reliance on self-reported applicant income; (2) algorithmic absorption of upstream institutional bias encoded within origination interest rates; and (3) a residual disparity (3.55 percentage points in cross-validation; 2.56 percentage points on a held-out test partition, Z = -4.04, p < 0.0001) that remains even after purging both income and interest rates from the model. Out-of-sample signed SHAP valuations demonstrate that this residual gap is maintained by structural proxies, most notably loan amount and home ownership status. These empirical findings show that simply blinding an algorithm to sensitive attributes fails to ensure fairness when institutional pricing decisions and behavioral proxy variables collectively reconstruct the omitted signals. We outline the practical implications of these findings for auditing data-centric AI workflows within regulated financial institutions.

  • 1 authors
·
Aug 7

MarineSparkLM-4B: A Lightweight Language Model for Marine and Offshore Engineering Built by Parameter-Efficient Adaptation

Marine and offshore engineering text is spread across preprints, open-access journals, encyclopedias and institutional documents, and the amount of it is small by pretraining standards. We ask how far a lightweight language model for marine and offshore engineering can be taken with parameter-efficient adaptation on a single workstation-class device. Starting from the 4-billion-parameter OceanGPT-basic-4B-Instruct, we train two independent LoRA adapters, each updating 33.0 million of 4.06 billion parameters (0.81%). Training took 4 hours 40 minutes for the first adapter and 10 minutes for the second on one NVIDIA DGX Spark. The first adapter is trained by next-token prediction on a 61.6-million-token corpus covering ten engineering themes. The corpus was built from 13,879 source documents, of which 7,069 were retained after deduplication and a two-tier keyword filter that removes cross-domain lexical collisions. Its validation loss falls from 2.082 to 1.976 over 1,500 steps. Direct testing showed that this adapter could still drift off topic on a retrieval-grounded ballast-water prompt, so we trained a second adapter on 2,295 question-answer pairs (114 hand-written and 2,181 generated by the base model from corpus passages) with the loss restricted to answer tokens. On 20 held-out prompts scored by hand for accuracy, topicality and specificity (maximum 120), the plain base model, the continuation adapter, the instruction adapter and the instruction adapter with retrieval scored 42, 56, 84 and 88. Absolute accuracy stays low (14 of 40 without retrieval and 20 of 40 with it). We report the full pipeline, two negative results (a reverted keyword fix and an instruction run that overfit within 100 steps), and the limits of the evaluation, which uses one small prompt set, no inter-rater agreement measure and model-generated reference answers.

  • 1 authors
·
Nov 13

When Safe Skills Collide: Measuring Compositional Risk in Agent Skill Ecosystems

LLM agents increasingly rely on community-contributed skills that expand an agent's operational capability set. We study a core safety problem in agentic AI systems: whether individually safe skills can compose into unsafe installed skill sets. We present SkillReact, a compositional security measurement framework with three components: a deterministic static-composition benchmark, a two-rater LLM-assisted human-adjudication pipeline, and an action-based exploitability harness. On 1,520 ClawHub skills, 651 pass individual inspection and form 211,575 pairs; the benchmark flags 22.25% of these as structural candidates. We treat this raw rate as a recall-oriented scanner ceiling and calibrate it against human judgment: in a pattern-stratified audit, roughly one in five flagged pair-pattern hits survives as a real compositional risk (population-weighted validity 18.2%, our headline result), implying about 14K genuine risk memberships in a single registry that per-skill scanning misses by construction, since every pair is individually safe. An action-based harness then probes when these candidates become model-issued tool calls, and finds realization gated by host-model disposition: on an anchor-conditioned dropper subset, Haiku-4-5 issues the dropper-stage tool call on all 39 direct-prompt trials (36 of them the full download-then-execute chain, 3 download-only), Opus-4-7 stops at the download, and Sonnet-4-6 refuses outright. A control that holds the request fixed and varies only the installed skills finds compliance highest with no skills installed: a composition fixes which capabilities are reachable, while the host model decides whether to use them. Together these motivate install-time compositional checks and capability isolation as complements to per-skill scanning.

  • 9 authors
·
May 29

Measuring Faithfulness Depends on How You Measure: Classifier Sensitivity in LLM Chain-of-Thought Evaluation

Recent work on chain-of-thought (CoT) faithfulness reports single aggregate numbers (e.g., DeepSeek-R1 acknowledges hints 39% of the time), implying that faithfulness is an objective, measurable property of a model. This paper demonstrates that it is not. Three classifiers (a regex-only detector, a two-stage regex-plus-LLM pipeline, and an independent Claude Sonnet 4 judge) are applied to 10,276 influenced reasoning traces from 12 open-weight models spanning 9 families and 7B to 1T parameters. On identical data, these classifiers produce overall faithfulness rates of 74.4%, 82.6%, and 69.7%, respectively, with non-overlapping 95% confidence intervals. Per-model gaps range from 2.6 to 30.6 percentage points; all are statistically significant (McNemar's test, p < 0.001). The disagreements are systematic, not random: inter-classifier agreement measured by Cohen's kappa ranges from 0.06 ("slight") for sycophancy hints to 0.42 ("moderate") for grader hints, and the asymmetry is pronounced: for sycophancy, 883 cases are classified as faithful by the pipeline but unfaithful by the Sonnet judge, while only 2 go the other direction. Classifier choice can also reverse model rankings: Qwen3.5-27B ranks 1st under the pipeline but 7th under the Sonnet judge; OLMo-3.1-32B moves in the opposite direction, from 9th to 3rd. The root cause is that different classifiers operationalize related faithfulness constructs at different levels of stringency (lexical mention versus epistemic dependence), and these constructs yield divergent measurements on the same behavior. These results demonstrate that published faithfulness numbers cannot be meaningfully compared across studies that use different classifiers, and that future evaluations should report sensitivity ranges across multiple classification methodologies rather than single point estimates.

  • 1 authors
·
Mar 20

Option-Aware Retrieval and Task-Specific VLM Adaptation for Medical VQA

We describe our submission to the MedReason 2026 challenge, covering multiple-choice (MCQ) and open-ended (OE) medical visual question answering (VQA) under fully offline, containerized inference. Our first finding is that MCQ retrieval must compare answer semantics rather than answer labels: labels are independently assigned per question, so copying a retrieved neighbor's label transfers no useful information, whereas scoring each current option's text against correct-answer text from similar training cases raises retrieval-only accuracy from 20.0\% to 57.5\% on a 200-case retrieval-excluded development holdout. Our second finding attributes the submitted system's accuracy: holding the task-specific MCQ Low-Rank Adaptation (LoRA) adapter fixed and varying the number \(k\) of in-prompt retrieved examples changes accuracy by at most one case --- 187/200 (93.5\%) at both \(k=0\) and the adapter's training-time \(k=1\), 188/200 (94.0\%) at the packaged runtime's default \(k=3\) --- and the submitted confidence-gated override adds no net accuracy on top of \(k=3\), selecting the VLM in 198/200 cases. With the final MCQ adapter fixed, retrieval changes accuracy by at most one case, and gating provides no net gain. On 20 OE cases, token-F1 and RaTEScore~zhao2024ratescore decrease as \(k\) grows, but paired sign tests on token-F1 differences are nonsignificant (\(p \ge 0.29\)); a single-annotator comparison found 6/20 wrong-anchor errors for the final configuration and 14/20 for an earlier configuration that jointly differed in routing, adapter, and prompting. The system reaches 94.0\% MCQ accuracy on the development holdout and 93.20\% on the organizer's official pre-evaluation, versus 29.43\% for the off-the-shelf reference baseline, while both of the organizer's open-ended scores are lower than that baseline's (ground-truth agreement 1.245 versus 1.588, visual accuracy 1.995 versus 2.696, each out of 4).

  • 3 authors
·
Sep 13

Reachability Is Not Realization: Tracing the Sources of LLM Benchmark Gains

Benchmark gains are often treated as evidence of greater LLM capability. Yet the same gain can reflect different changes in model behavior. A model may reach new answers, or produce answers that were already within reach. Aggregate scores do not distinguish these changes question by question. We establish a question-level audit under fixed budgets, temperatures, and answer formats. A question is realized when the default deployment procedure produces the correct answer. A question is reachable when a specified probe finds that answer within a fixed budget. We first test whether inference-time layer routing can expand reachability. Under a matched budget, random routes match or exceed structured search in all 43 model and task settings. Answer-blind procedures retain almost none of this gain, which instead requires access to the correct answer. We then ask why reachable answers sometimes fail to appear. Across six cases spanning 0.5B to 31B, silencing one identified MLP block repairs 68 to 92 percent of a predefined failure set. We next test whether training closes the gap by expanding reachability. In five of six matched evaluations, deployed performance rises while the reachable ceiling remains flat or falls. For DAPO, the deployed score rises by 14.7 points while the reachable ceiling falls by 13.3 points. Across the settings we audit, realization and reachability therefore do not always change together. Claims of capability expansion should report both realized performance and reachability under matched evaluation conditions. Code is available at https://github.com/LiZaiyuan0619/reachability-not-realization

  • 7 authors
·
Aug 3

WorldCupArena: Fine-Grained Evaluation of Language Models and Deep-Research Agents on Football Forecasting

Predicting a football match before kickoff requires more than knowing past results: a model must use changing information and make a clear prediction before the answer is available. We present WorldCupArena, a dynamic benchmark for language models and deep-research agents. The 2026 FIFA World Cup is its first evaluation, and the same process can be reused for future leagues and cups. Before each match, a model either receives a common evidence package or searches for information itself. It predicts the result and score, likely players and events, match statistics, and the outcome of the competition. After the match, these predictions are compared with the recorded result. We report result accuracy, exact-score accuracy, and a scoreline score that gives some credit when a predicted score is close but not exact, together with scores for the other prediction tasks. Across 104 matches and 13 systems, models with similar result accuracy differ more clearly on detailed predictions. Compared with betting-market and human-fan baselines, the best system shows only small gains in result and exact-score accuracy, but a clearer gain in Scoreline. New schedules can be added as they begin, allowing the benchmark to evaluate future models without using outcomes that are already known. Code, prompts, predictions, and evaluation scripts are open sourced at https://github.com/wzk1015/WorldCupArena.

  • 6 authors
·
Jul 19 2

RetrievalRouter: Joint Modality and Architecture Selection for Document Retrieval

Document retrieval increasingly supports high-stakes information access in finance, healthcare, and law. Modern retrieval pipelines vary both in modality (text or multimodal) and in retrieval architecture (dense or late-interaction). These choices impose a hard compromise: the most effective pipelines are too slow and expensive to run at scale, while the fastest fail to retrieve evidence from complex documents. Practitioners must therefore choose between missed evidence and unusable latency, with no principled basis for adapting that choice at the query level. We show that this compromise is unnecessary. Not every query requires the same pipeline. Across benchmarks spanning financial and scientific corpora, no static pipeline dominates. We introduce RetrievalRouter, a lightweight query-aware router that learns, from the query text alone, which retrieval pipeline best fits each query. A single tunable parameter exposes the full accuracy-latency frontier, and for every static baseline, RetrievalRouter offers an operating point that is simultaneously more accurate and faster. Against the best static baseline, RetrievalRouter is 2.5% more accurate and 12.4 times faster. Furthermore, compared with prior adaptive strategy selection methods, RetrievalRouter achieves significantly higher nDCG@5 across accuracy-oriented settings, while matching or numerically outperforming them on both nDCG@5 and latency in latency-oriented settings. Our code and data are available at https://github.com/emrekuruu/retrieval-router.

  • 4 authors
·
Aug 25 2

Model of Models: When Does Emitting a Specialist Beat Attending, Adapting, or Tuning?

Given a task described by a few examples, how should a model be specialized to it? Four mechanisms are available -- zero-shot, in-context attention, test-time gradient adaptation, and emitting specialist weights from a hypernetwork -- yet the operating regime of the last is rarely mapped. We run the identical four-way comparison across six tasks spanning regression, generation, language modeling, reinforcement learning, and clinical and genomic classification, holding the specialist, the context, and (where we can) the training budget fixed. The clearest wins for emission are about cost at matched quality: it ties the state-of-the-art amortized tabular model (TabPFN) on clinical few-shot classification while emitting a reusable specialist instead of re-attending the support set per query, and reaches noise-floor shape generation with a 132-float per-instance program. On few-shot sinusoid regression it is 2--3 orders of magnitude below MAML at zero test-time gradient steps -- a margin that narrows to sim30times but persists once training budgets are equalized. Emission cannot match in-context attention on high-dimensional sequence modeling: under matched-budget pre-training a one-pass adapter recovers only a minority of the in-context gain (14.0pm0.9% at 5M, 11.2pm0.5% at 15M), and a LoRA-rank sweep shows this shortfall is a partial capacity limit -- capture climbs from 5% to 21% as rank grows but plateaus far below full recovery. Mechanism ablations confirm the emitted specialist is genuinely task-conditioned, not a memorized prior; and, more speculatively, emitted specialists compose in weight space -- interpolating two of them tracks the corresponding blend of their functions. We close with a falsifiable thesis, operationalized through a per-task resolution measure, bounding when each conditioning mechanism should be preferred.

  • 1 authors
·
Jul 19

CountVideoBench: A Benchmark to Evaluate Prediction Biases on the Object Counting task by Video-Language Models

Object counting in videos remains a poorly characterized capability of multimodal large language models. Counting questions are included in many existing VideoQA benchmarks, but the ground-truth count distributions in these benchmarks are heavily skewed toward small values. Across the nine benchmarks we survey, the median ground-truth count is at most five, and the three most frequent values together account for between 45% and 93% of the videos. The resulting mean accuracy is then dominated by the most frequent counts; a value-dependent bias at a rare count contributes negligibly to the average and remains invisible. We introduce CountVideoBench, a benchmark of 2,600 synthetic videos covering 26 cardinalities in the interval [5, 50] (100 videos each), grouped into six clusters of consecutive counts that span the range while enabling near-neighbor comparisons. Each video contains a single class of static objects under a controlled camera motion, removing object motion and reducing recognition as confounds. We evaluate sixteen video-language models from nine architecture families in a zero-shot setting. No model exceeds 27.6% Exact Match. The dependence of accuracy on cardinality varies across models, ranging from monotonic decline to a non-monotonic peak-and-collapse pattern, which rules out increasing difficulty with cardinality as the dominant explanation of counting failure. Every model further concentrates its predictions on its own set of preferred integers, with accuracy peaks coinciding with these values and collapsing at adjacent counts. These findings highlight the need for balanced, diagnostic counting benchmarks to reliably assess this fundamental visual capability.

  • 4 authors
·
Oct 25

Compression Is Not Evaluation-Neutral: Fixed RAG Compression Can Distort Reader Comparisons

Retrieval-augmented generation (RAG) evaluations often compare readers after a compressor has changed their evidence. This mixes two questions: which complete pipeline works best, and how much of a reader upgrade survives compression. We show that fixed compression can raise average pipeline accuracy while hiding most of a reader upgrade. We hold questions, retrieved candidates, prompts, scoring, and compressed text fixed while comparing 8-20 readers across five question-answering benchmarks and five compression families. HotpotQA and MuSiQue are the main benchmarks, analyzed under a plan fixed in advance. In the 20-reader HotpotQA panel, the lowest- and highest-scoring readers under raw evidence are 31.8 percentage points (pp) apart before compression but only 7.8pp apart on the same stored output from a HotpotQA-trained RECOMP compressor. On both main benchmarks, lower raw-scoring readers gain more under exact match and token-overlap F1. Reader pairs also change order more often under compressed evidence than in raw replays across question halves. Row-level accounting explains how higher average accuracy and smaller upgrades coexist: compression rescues some wrong answers and damages some correct answers, with a different balance for each reader. We release ragscale to audit declared reader upgrades under raw and fixed compressed evidence. Reader evaluations should run this audit before attributing a compressed-only result to the reader.

  • 2 authors
·
Oct 1

Verifying Text-to-SQL Queries under Changing Business Semantics

In text-to-SQL and AI-agent analytics, successful SQL execution does not establish semantic correctness under changing business definitions. We study revalidation of fixed queries after changes to refund attribution, tax treatment, eligible populations and exchange-rate dates. An executable verification profile binds the requested revision and knowledge cutoff to legal counterexample data and exact monetary comparison. A preregistered confirmation on 32 newly authored revision families and two local producer bundles retains 192 query-definition cells: 144 nonconforming, 36 finite-suite agreeing and 12 unresolved. Of the nonconforming cells, 63 agree on the current snapshot. At an eight-ledger budget, obligation-guided checking detects 137 discrepancies versus 127 for generic ordering, a paired difference of 6.9 percentage points (family bootstrap 95% interval [2.1, 13.6]; exact family sign-flip p=0.0156). Every ledger arm rejects zero of the 36 agreeing cells. Current-obligation and equal-weight scoring match the guided result, identifying coverage rather than revision-specific priority as the observed source of benefit. Controlled assays, a natural-query pilot and repeated samples provide separate mechanism and robustness evidence. A legal discrepancy refutes conformance; agreement remains restricted to the recorded suite. The study establishes an auditable business-semantic revalidation profile within its authored schema and local-model scope.

  • 4 authors
·
Oct 29

Do Small Models Use the Law You Give Them? Measuring Context Use on a Bilingual Bangladesh Legal Benchmark

Fine-tuning can improve legal question-answering accuracy without improving how models use law supplied in context. We study this distinction in bilingual Bangladeshi legal QA, where observed errors can arise from answer scoring, retrieval, or failure to use relevant law. We construct a hierarchy-preserving statutory corpus, 2,165 reviewed bilingual fine-tuning examples, and a 150-item supplied-law control. We evaluate six instruction-tuned models: Llama-3.2-1B, Llama-3.2-3B, Qwen3.5-0.8B, Qwen3.5-2B, Qwen3.5-4B, and Gemma-4-E2B, with three LoRA seeds per model. To separate effects, we combine constrained option-letter scoring, cyclic option rotation, and controlled removal of the governing provision. On 398 Bar Council outputs, an exact-line parser attributes an accuracy gain of 50.0\% to the Qwen3.5-2B seed-42 adapter, whereas option scoring yields only 3.0%. For Gemma-4-E2B, the two scoring methods favor different systems. When the governing provision is guaranteed to be present, five of six reference models improve by 14.7%-19.3% under the four-order criterion. Removing that provision reduces accuracy by 8.0%-15.3% for models and by 13.8%-14.9% points for their adapters. However, difference-in differences estimates show no increase in reliance on the governing provision after fine-tuning. Results show that legal adaptation claims require separating scorer, retriever, and model effects. Our Code and data are available at https://anonymous.4open.science/r/bangladesh-legal-qa-11E3

  • 6 authors
·
Aug 30

Finding the Right Fit: Model-Harness Interactions across Agent Tasks

Choosing an agent system means choosing both a language model and the harness through which it acts. We ask whether a strong model, harness, or pairing stays strong when the setting changes. We evaluate 66 configurations: four configurable harnesses (OpenHands, DeepSeek Harness, PI, and openJiuwen) paired with five models on TUA-Bench, ALE-CLI, and Terminal-Bench 4, plus the native Codex-GPT and Claude Code-Claude pairings. Model rankings reverse across harnesses. On Terminal-Bench 4, Claude leads GPT by 7.94 points in OpenHands but trails it by 30.16 points in PI. For four of the five models, the best harness changes from one benchmark to another, yet some pairings hold: openJiuwen gives Kimi its highest score on all three benchmarks, by 5.61 to 11.11 points. A model's own vendor harness is not reliably its best, and higher cost does not reliably buy a higher score. On Terminal-Bench 4, GPT scores higher under PI than under DSH at less than a quarter of the cost per task. Matched trajectories suggest why fit varies. Models start almost all repairs themselves, so much depends on whether the harness hands failures back in a form the model can use. GPT does best with PI's lean scaffold, while Kimi, which often issues malformed tool calls, does best in openJiuwen. We argue that the model, the harness, and the task should be evaluated together, and we release the harness adapters, evaluation code, and all 6,204 scored trajectories at https://github.com/liyix/finding-the-right-fit and https://huggingface.co/datasets/yixuanli97/finding-the-right-fit.

  • 9 authors
·
Sep 30

CliniQ: A Multi-faceted Benchmark for Electronic Health Record Retrieval with Semantic Match Assessment

Electronic Health Record (EHR) retrieval plays a pivotal role in various clinical tasks, but its development has been severely impeded by the lack of publicly available benchmarks. In this paper, we introduce a novel public EHR retrieval benchmark, CliniQ, to address this gap. We consider two retrieval settings: Single-Patient Retrieval and Multi-Patient Retrieval, reflecting various real-world scenarios. Single-Patient Retrieval focuses on finding relevant parts within a patient note, while Multi-Patient Retrieval involves retrieving EHRs from multiple patients. We build our benchmark upon 1,000 discharge summary notes along with the ICD codes and prescription labels from MIMIC-III, and collect 1,246 unique queries with 77,206 relevance judgments by further leveraging powerful LLMs as annotators. Additionally, we include a novel assessment of the semantic gap issue in EHR retrieval by categorizing matching types into string match and four types of semantic matches. On our proposed benchmark, we conduct a comprehensive evaluation of various retrieval methods, ranging from conventional exact match to popular dense retrievers. Our experiments find that BM25 sets a strong baseline and performs competitively to the dense retrievers, and general domain dense retrievers surprisingly outperform those designed for the medical domain. In-depth analyses on various matching types reveal the strengths and drawbacks of different methods, enlightening the potential for targeted improvement. We believe that our benchmark will stimulate the research communities to advance EHR retrieval systems.

  • 8 authors
·
Feb 10, 2025

Extracting alignment data in open models

In this work, we show that it is possible to extract significant amounts of alignment training data from a post-trained model -- useful to steer the model to improve certain capabilities such as long-context reasoning, safety, instruction following, and maths. While the majority of related work on memorisation has focused on measuring success of training data extraction through string matching, we argue that embedding models are better suited for our specific goals. Distances measured through a high quality embedding model can identify semantic similarities between strings that a different metric such as edit distance will struggle to capture. In fact, in our investigation, approximate string matching would have severely undercounted (by a conservative estimate of 10times) the amount of data that can be extracted due to trivial artifacts that deflate the metric. Interestingly, we find that models readily regurgitate training data that was used in post-training phases such as SFT or RL. We show that this data can be then used to train a base model, recovering a meaningful amount of the original performance. We believe our work exposes a possibly overlooked risk towards extracting alignment data. Finally, our work opens up an interesting discussion on the downstream effects of distillation practices: since models seem to be regurgitating aspects of their training set, distillation can therefore be thought of as indirectly training on the model's original dataset.

google Google
·
Oct 21, 2025 5

Evaluating Correctness and Faithfulness of Instruction-Following Models for Question Answering

Retriever-augmented instruction-following models are attractive alternatives to fine-tuned approaches for information-seeking tasks such as question answering (QA). By simply prepending retrieved documents in its input along with an instruction, these models can be adapted to various information domains and tasks without additional fine-tuning. While the model responses tend to be natural and fluent, the additional verbosity makes traditional QA evaluation metrics such as exact match (EM) and F1 unreliable for accurately quantifying model performance. In this work, we investigate the performance of instruction-following models across three information-seeking QA tasks. We use both automatic and human evaluation to evaluate these models along two dimensions: 1) how well they satisfy the user's information need (correctness), and 2) whether they produce a response based on the provided knowledge (faithfulness). Guided by human evaluation and analysis, we highlight the shortcomings of traditional metrics for both correctness and faithfulness. We then propose simple token-overlap based and model-based metrics that reflect the true performance of these models. Our analysis reveals that instruction-following models are competitive, and sometimes even outperform fine-tuned models for correctness. However, these models struggle to stick to the provided knowledge and often hallucinate in their responses. We hope our work encourages a more holistic evaluation of instruction-following models for QA. Our code and data is available at https://github.com/McGill-NLP/instruct-qa

  • 5 authors
·
Jul 31, 2023

Joint or Per-Option? Compact Encoders for Closed-Set Decisions from 17 Million to 1 Billion Parameters

Many language-model pipelines end in a closed-set decision, such as selecting a tool. Prompting a large language model for it is slow and order-sensitive. We ask how small an encoder for such decisions can be and whether it should read the options jointly or separately. In a three-seed study at 32 million parameters, per-option scoring beats joint scoring on a held-out sample of a 37-benchmark decision suite by 5.8 points of chance-corrected skill at matched optimizer steps and by 2.8 at approximately matched training compute; two external benchmarks agree, though not significantly on their unseen tasks. Crossing reader with initialisation shows that the advantage requires a reranker initialisation, whose query-candidate pair input matches per-option reading; from the raw encoder the two readers are indistinguishable. Per-option scoring needs 5 to 13 times more encoder passes: at most twice the median latency, but more than four times the mean. Of six released joint encoders (17 million to 1 billion parameters), the 68- and 150-million-parameter models exceed a 435-million-parameter DeBERTa-v3 zero-shot classifier by 4.9 and 5.9 accuracy points on 209 overlap-audited System1-Arena tasks and are at least on par with it on the 145 tasks unseen in training. The smallest answers a median request in 14 milliseconds on 16 processor threads, against 728 for the classifier. Reordering options changes 12 to 23% of the joint models' answers (per-option scoring: none). Up to 68 million parameters, they lead their size range on a public leaderboard's private, mainly in-distribution tests.

  • 1 authors
·
Oct 21 2

Sketchpad: certified exact arithmetic with a small vision-language model

Language models are accurate on most long-arithmetic problems, but their accuracy is estimated on samples and carries no guarantee for an unseen input; tool use is exact but moves the computation outside the network. We show that exactness at all input lengths becomes a finite, checkable property of a network when each of its steps reads only a bounded view of an external working memory. In Sketchpad, Qwen3.5-0.8B with a low-rank adapter executes arithmetic procedures on a rendered sheet. Each forward pass reads a three-cell view and emits one step: a digit, a carry or borrow, and the next operation. A fixed runtime moves the view and writes no digit of a result. We prove in Lean that the model is then exact on all inputs of any length if and only if it is exact on the finite set of views the runtime presents, that no finite test suffices when steps also read earlier steps, and that the guarantee tolerates bounded numerical perturbation and extends to compositions of procedures. All 2,446 views pass on the deployment device, a laptop GPU running the model in a web browser. The certified model performs signed addition and subtraction, long multiplication and division, square roots and greatest common divisors at arbitrary length. It solved 200 of 200 pre-registered problems in the browser, and 220 of 220 sums of up to 1,000 digits and products of up to 20 digits, of which GPT-6.1 Sol, the strongest of ten other models, solved 217. A symbolic runtime built on these procedures solves, expands and factors polynomials with every number computed by the model, and a multiplicative gate on the adapter recovers the base model exactly when the sheet is not in use.

  • 1 authors
·
Nov 3 3

JIT-MPE: Just-in-Time Multi-Pitch Estimation under Output Deadlines

Live pitch displays, score followers and accompaniment systems must know which pitches are sounding within tens of milliseconds, but instrument-agnostic multi-pitch estimation (MPE) has mainly been designed and evaluated offline. We present JIT-MPE, a causal MPE model of 86k parameters for mixed ensembles. It gives an immediate estimate of each 20 ms frame and re-estimates the same frame as 20, 40, 80 and 160 ms of further audio arrive, so that a live display can show pitches at once and then refine them. We score each output only at a deadline by which it can be computed from the audio that has arrived, on five development sets and four held-out sets evaluated only after all design choices were frozen. At a 50 ms deadline, which reads the update made with 40 ms of lookahead, JIT-MPE is significantly more accurate than five of six published offline MPE systems in 38 of 45 comparisons (15 of 20 on the held-out sets) and less accurate in only two, although these systems read the whole recording and all but one have 6 to 170 times as many parameters; the sixth, supervised on URMP, is higher on four of the nine sets. The deadline costs little: on the held-out sets, JIT-MPE at 50 ms is within -1.27 to +0.60 points of recording-level macro F1 of its own final output with 160 ms of lookahead and within -2.08 to +0.79 points of a bidirectional model of similar size. Waiting pays off where the immediate estimate is weakest: the update at 50 ms adds 1.7-3.6 points over it on the development data and three of the four held-out sets, and raises the recall of reference frames in the first 20 ms after a note onset from 9.8% to 38.4% on the held-out AAM set. The slowest measured frame takes at most 4.15 ms on one CPU thread. Code and trained weights are available at https://github.com/Frieve-A/jit-mpe.

  • 1 authors
·
Oct 22

ListQA: A Benchmark for Evaluating List-Formatted Factual Knowledge Retrieval in Large Language Models

Many real-world questions demand not a single fact but an organized collection of them, yet existing factual knowledge benchmarks almost exclusively target single answer retrieval. We introduce ListQA, a benchmark of 9,045 human-curated, cross-validated questions (audited error rate of 1.7%) that require LLMs to recall multiple facts and compose them into structured lists of 3-10 elements. Questions range from flat lists ("Which NCAA teams went undefeated between 2000 and 2024?") to hierarchical lists with sub-attributes ("...and what was their record and result?"), spanning eight categories across 60+ countries. An evergreen design (explicit temporal constraints or historically immutable facts) keeps ground truth valid without periodic updates. For evaluation, we frame element matching as a linear sum assignment problem: optimal bipartite matching paired with LLM-based semantic grading jointly captures factual accuracy, hallucination, and format compliance. We benchmark 34 models (8 families, 42 configurations across standard and thinking enabled inference) and find: the best standard-inference model scores below 37% LLM-Judge-F1, multi-faceted questions are harder across the board, and providing the source Wikipedia page lifts scores dramatically, confirming that models can organize facts but struggle to retrieve them from parameters alone. Extended analyses cover base vs. instruction-tuned checkpoints and token efficiency. We release the dataset and evaluation code.

AmazonScience Amazon Science
·
Oct 12

PageRecall: Measuring Page Selection in Literature-Grounded Question Answering

We describe our system for LitTraceQA (GroundLM @ EMNLP 2026): given a research question, retrieve the relevant papers from a pool of 27,487, cite the page and the table or figure where the answer lives, and answer in a requested format. Our main finding is that evidence grounding is limited by retrieval, not by reading. The page selector put the annotator's page, which we call the gold page, in front of the model that locates evidence only about half the time (52.6% gold-page recall), while that model, given the page, cited the right one in 45 of the 48 locators it emitted (94%). When the page was missing it rarely said so: of 45 such cases it returned nothing 14 times, a wrong page 24 times, and a correct page 7 times, so the pipeline failed quietly almost twice as often as it failed visibly. Since the failure was that the right page was never shown, the fix is to stop choosing: each retrieved paper fits in the model's context, so we show it whole. Page ranking survives only as a fallback inside papers too long to fit, which no test-split paper was, and gold-page recall reaches 100% on the papers we can parse. Separately, questions that identify their target by position rather than content, such as "the first author of the 24th reference", are served by parsing rather than retrieval: we resolve the bibliography into an addressable list, which also supplies identifiers the evidence metric scores. The final system scores 0.762 paper F_1, 0.441 evidence F_1 and 0.920 multiple-choice accuracy on the held-out test split. Because the pipeline depends on a closed model without seed control, we release a harness that verifies the paper's central claims against committed artifacts.

  • 1 authors
·
Sep 15

BARS-CTR: Open Benchmarking for Click-Through Rate Prediction

Click-through rate (CTR) prediction is a critical task for many applications, as its accuracy has a direct impact on user experience and platform revenue. In recent years, CTR prediction has been widely studied in both academia and industry, resulting in a wide variety of CTR prediction models. Unfortunately, there is still a lack of standardized benchmarks and uniform evaluation protocols for CTR prediction research. This leads to non-reproducible or even inconsistent experimental results among existing studies, which largely limits the practical value and potential impact of their research. In this work, we aim to perform open benchmarking for CTR prediction and present a rigorous comparison of different models in a reproducible manner. To this end, we ran over 7,000 experiments for more than 12,000 GPU hours in total to re-evaluate 24 existing models on multiple datasets and settings. Surprisingly, our experiments show that with sufficient hyper-parameter search and model tuning, many deep models have smaller differences than expected. The results also reveal that making real progress on the modeling of CTR prediction is indeed a very challenging research task. We believe that our benchmarking work could not only allow researchers to gauge the effectiveness of new models conveniently but also make them fairly compare with the state of the arts. We have publicly released the benchmarking code, evaluation protocols, and hyper-parameter settings of our work to promote reproducible research in this field.

  • 5 authors
·
Sep 12, 2020

A Controlled Candidate-Set Benchmark for Offline Satellite-Security Plan Decomposition

Some security-development settings require local models that map an objective to an ordered, checkable plan. We present a low-rank decomposition adapter and release a case-disjoint corpus with 24 authored decompositions and 83 derived next-step examples across 24 satellite-security cases. To prevent reference-plan leakage, prompts contain one objective and eight shuffled SPARTA candidates. The evaluation candidate sets are constructed with oracle inclusion of every reference technique, so this study measures selection from a controlled set, not retrieval. On six fixed cases and five decoding seeds, mean scores vary across adapters and prompted Qwen2.5 baselines; a separate greedy check over three adapter training seeds shows non-negligible training variance. A 24-fold greedy leave-one-case-out control at the two compact sizes likewise shows tradeoffs: at 1.5B, adapter precision is 0.583 against 0.480 for two-shot prompting, while recall is 0.586 against 0.660. Case-resampling intervals for adapter precision relative to the strongest prompted baseline cross zero at 1.5B and 7B. The adapter is also substantially more format-compliant, which prevents attributing all score differences to selection. Autoregressive next-step accuracy is 0.08, 0.06, and 0.29 at 0.5B, 1.5B, and 7B. We therefore offer the formulation, leakage-controlled dataset, reproducibility package, and descriptive analysis as a proof of concept, not evidence of a reliable stand-alone decomposition system or a general adapter advantage. A separate diagnostic tests contract reuse.

  • 3 authors
·
Jul 28

Time Present and Time Past: Benchmarking Large Language Models on Temporally Evolving Document Understanding

Evolving documents, such as laws, tax codes, and software documentation, are amended, replaced, and sometimes reverted over time, so a question has different correct answers at different dates. In contrast to encyclopedic knowledge, where an old fact is simply overwritten, an amendment is itself an official text that states what it replaces and when it takes effect, and the earlier version stays correct for its validity period. The central challenge is therefore version resolution, that is, identifying the version in force on the queried date. Existing temporal QA datasets treat time only as an annotation, so version resolution stays untested. We present TIDE, an expert-verified benchmark of 3,050 QA pairs over 644 official customs instruments issued between 1969 and 2025 by the Government of Bangladesh, covering eight task types over deeply code-mixed documents that are heterogeneous in layout and dated in two calendars. In addition, we evaluate nine recent LLMs under a single protocol across parametric, gold-context, and retrieval access, scored by a three-judge LLM council with a hard date gate separating correct meaning from correct time. The best macro-averaged accuracy is only 68.5%. Resolving a version from an implicit date reaches 59.7%, and detecting that the supplied version does not govern the query reaches only 26.7%. Models are more likely to find correct versions than to reject incorrect ones, and they tend to follow a confident parametric answer over the supplied authoritative text. All code and data are available at https://github.com/icsetepa44/TIDE

  • 7 authors
·
Aug 8

Field Order Should Not Matter: Permutation-Invariant Embedding Model Fine-Tuning for Structured Metadata Retrieval

We study retrieval over catalogs of structured metadata, where each record is a small schema whose fields answer different kinds of query. Embedding a record with a text encoder first serializes its fields into a string, which forces a choice of field order. We show this choice, usually treated as an implementation detail, silently controls retrieval quality once the encoder is fine-tuned. A standard fine-tune loses 7.4 nDCG@10 points when the index is rebuilt under a different field order, because it reads absolute position instead of the field labels. We propose permutation-invariant fine-tuning (PI-FT), which serializes each record under a freshly sampled field order with random field dropout, so meaning binds to the labels rather than to position. The change is about two lines in the data loader; it costs negligible in-distribution accuracy and cuts the order-change penalty to 0.2 points. We study this in the discovery of development statistics, a catalog of nearly 10,000 indicators that should be searchable in many languages by a model small enough to self-host. As AI assistants and agents increasingly mediate access to public data and statistics, this retrieval step decides whether an answer is grounded in the right indicator or series, making discoverability a precondition for disseminating data through AI. Because usage logs cannot provide training signal for indicators no one has searched, we generate the queries instead. DevDataBench is a fully LLM-generated benchmark of grounded, facet-targeted queries across 15 languages, covering every indicator for both training and evaluation. A fine-tuned 118M-parameter CPU encoder outperforms every zero-shot baseline, including text-embedding-3-large (0.707 vs.\ 0.556 nDCG@10), with the largest gains in low-resource languages. We release the benchmark, pipeline, models, and a reusable PI-FT framework.

  • 3 authors
·
Jun 28

LENS: In-Context Search via Latent Evidence Exploration over Dynamic Raw Documents

LLM agents increasingly answer questions over dynamic raw-document collections, where files may change before preprocessing, and relevant evidence (spans, sections, pages, or tables) is query-dependent. Existing retrieval-augmented approaches pre-materialize evidence via fixed chunking, embeddings, or persistent indexes: effective for lookup, yet costly, stale-prone, and committed to a granularity before the query is known. We formulate in-context search as Budgeted Evidence Localization over a latent evidence space induced by dynamic raw documents and propose LENS (Latent Evidence Exploration and Search), an index-free framework. Instead of pre-materializing the evidence space, LENS maintains a query-conditioned belief over candidate units, iteratively selecting candidates via complementary lexical, local, and exploratory proposal policies, updating the belief via an LLM relevance oracle, and narrowing toward high-posterior regions under a controllable budget. Evidence is consolidated into compact, source-grounded regions of interest and compressed into self-organizing knowledge clusters reused across related queries. On a controlled 500-question evaluation with matched corpus snapshots, LENS reaches 62.4% exact match and 84.8% evidence recall vs. 65.2% exact match but 50.4% evidence recall for a ReAct-style baseline. Across scales, LENS gives the strongest supporting-fact localization and answer grounding. On a fixed 150-question fullwiki subset over the raw Wikipedia dump with zero indexing, LENS and ReAct are nearly tied in official answer quality (43.3% vs. 42.7% EM), with LENS grounding more answers in retrieved evidence (84.0% vs. 70.7%). A no-retrieval Closed-Book reference highlights the contribution of model memory. LENS is query-ready after corpus changes, needs no preprocessing or persistent index, and preserves source-grounded evidence localization throughout.

  • 6 authors
·
Aug 17

HyDRA: Hybrid Dynamic Routing Architecture for Heterogeneous LLM Pools

Production LLM deployments increasingly maintain heterogeneous model pools spanning order-of-magnitude cost differences. Existing routers make binary strong-vs-weak decisions and couple learned parameters to specific model identities, requiring retraining whenever the catalog changes. We present HyDRA (Hybrid Dynamic Routing Architecture), a framework that predicts fine-grained, multi-dimensional capability requirements per query and matches them against configuration-defined model profiles via shortfall matching. A ModernBERT encoder with K=4 independent sigmoid heads scores each query along reasoning, code generation, debugging, and tool use; a shortfall-matching algorithm then selects the cheapest model whose capabilities meet the predicted requirements. The deployed predictor runs at 86 ms median CPU inference latency in production, and is fully decoupled from the model catalog -- adding or removing models requires only a configuration change, with zero retraining. On SWE-Bench Verified (5-model pool: GPT-5.4-mini, Claude Haiku 4.5, GPT-5.3 Codex, Claude Sonnet 4.6, GPT-5.4), HyDRA's tunable shortfall threshold spans three regimes: peak-quality exceeds the always-strong Claude Sonnet 4.6 baseline (75.4% vs. 74.2% resolution) at 12.9% cost savings; iso-quality matches Sonnet at 54.1% cost savings, a 6x improvement over our prior in-house binary router at 9.1%; aggressive pushes savings to 72.5% for a 3.2-point quality trade. Results generalize across LiveCodeBench, BigCodeBench, and tau-bench. HyDRA is deployed to all users in GitHub Copilot's VS Code Chat auto-mode and -- to our knowledge for the first time in the LLM routing literature -- demonstrates language-invariant routing across CJK, European, and other script families.

  • 5 authors
·
Jun 11

Results and Retrospective Analysis of the CODS 2025 AssetOpsBench Challenge

Competition retrospectives are useful when they explain what a leaderboard measured, how hidden evaluation changed conclusions, and which design patterns were rewarded. We revisit the CODS 2025 challenge, a privacy-aware Codabench competition on industrial multi-agent orchestration built on . We combine final rank sheets, a 300-submission server log, 149-team registrations, best-submission exports, the organizer winners report, the companion system paper, and verified planning-track source trees. Five results stand out. First, the public planning leaderboard saturates at 72.73\%, and richer prompts do not improve that peak. Second, hidden evaluation changes the story: public and private scores correlate moderately in planning (r{=}0.69) but negatively in execution (r{=}{-}0.13), with several 45.45\% public execution systems reaching 63.64\% on the hidden set. Third, the term is numerically almost inert in the official composite -- combined on a 0--1 scale with 0--100 percentage scores, it contributes at most 0.05 points per track, and rescaling would swap the top two teams. Fourth, the competition is operationally account-based but substantively team-based: 149 registered teams reduce to 24 with non-zero public scores and 11 fully ranked, while 52.3\% of deduplicated registrations list multiple usernames. Fifth, successful execution methods mostly improve guardrails -- response selection, contamination cleanup, fallback, and context control -- rather than novel agent architectures. These findings identify which behaviors the evaluation rewarded, and motivate scale-aware composites, skill-level diagnostics, and versioned artifact release.

ibm-research IBM Research
·
May 7 2

Test-Time Strategies for More Efficient and Accurate Agentic RAG

Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.

  • 10 authors
·
Mar 12 2

Online Matching with Size-Based and Convex Delays

We study the online min-cost perfect matching with delay (MPMD) problem where m requests arrive in a metric space of n points. In MPMD, an algorithm can choose to match a request or to delay, and the objective is to minimise the sum of connection and delay costs. The connection cost of a match is the distance between the locations of two matched requests in the metric, and the increase of the delay cost is a function of the set of unmatched requests at every moment. In this paper, we study two different types of delay functions, size-based (MPMD-Size) and convex delays (MPMD-Convex). The study of MPMD-Size was initiated by Deryckere and Umboh (APPROX/RANDOM 2023) where the instantaneous delay increment is a non-negative monotone function of the number of unmatched requests. Our bounds are in terms of n, as opposed to Deryckere and Umboh's bounds that depend on m. Our results settle the deterministic competitive ratio (up to constants). At the heart of these results is a succinct encoding scheme of MPMD-Size on a given n-point metric as a metrical task system problem on a 2^{n-1}-point metric. We also consider MPMD-Convex proposed by Liu et al. (ISAAC 2018) where the delay cost incurred by each request is a uniform convex delay function of the time difference between its arrival time and the moment that it is matched by the algorithm. They focused on delay functions f that are unbounded, non-decreasing, continuous, and satisfy f(0)=f'(0)=0, and showed that the deterministic competitive ratio is Ω(n) for n-point uniform metrics. We show that, surprisingly, when f is a non-negative, monotone polynomial with f'(0)>0, there is an O(1)-competitive deterministic algorithm for uniform metrics. Our result completes our understanding of MPMD-Convex on uniform metrics for a broad class of functions.

  • 3 authors
·
Jun 30

Rules to Tools: Executable Checks for LLM Agents in Scientific Computing

Scientific coding agents receive equations, boundary conditions, and output requirements in writing, then must assess the programs they revise. Rules to Tools (R2T) supplies prepared executable checks of public scientific requirements. Matched SciCode repair groups share written checks, starting programs, model, and budgets; the tool group receives a callable implementation. Across two task-ID cohorts, complete repair is 26/30 with text and 29/30 with the prepared checks. Three task IDs favor tools, one favors text, and eleven tie. The eight-ID cohort scores 13/16 versus 15/16, with a task-cluster bootstrap 95% interval of [-12.5, 43.75] percentage points for the difference. The larger shared-definition SciCode cohort ties at 13/24 per group. Five development-exposed tasks with alternate starting programs score 3/10 versus 7/10. The tool group favors tasks 17, 77, and 11; initial checks flag task 17 and report no violation for tasks 77 and 11. Task 37 favors text and has no initial reported violation. A fresh source-through-Python arm also reaches 15/16, matching the dedicated command's aggregate. In a matched PDE comparison, detailed text scores 23/24 and checks score 24/24, with 31.2% lower reported model output for checks. Agent-side output savings vary by cohort, while public CPU use rises in both task-ID cohorts. These results measure task-dependent repair outcomes and agent-side costs with prepared checks.

Metric-Construction Coupling Inflates Measured Synthetic Dialect Recovery

Korean dialect corpora are available but not redistributable: weights may be released, while reproducing training and evaluation from the underlying data cannot be. We ask how much of that supervision synthetic data recovers, and whether that recovery can be measured independently of the synthesis pipeline. We contribute KoDialectBench, 1,000 items across five regions on three axes, released as identifier hashes and scoring code so users reconstruct the items from their own licensed copy. Recovery is strongly axis-dependent: our best synthetic arm reaches 91.2% of the real-data gain on region identification but 63.7% on comprehension. On generation the answer depends on the metric: the deployed marker lexicon reports 92.3% on dialectness and 119.3% on region match, the latter exceeding the real-data reference, whereas reference-based generation reaches 72.4%. We find the marker metrics' scoring inventory is entirely contained in the inventory our transformation rules can emit. We test the effect of construction access directly with an exact-form construction-disjoint arm that withholds 20% of marker types from the rules. At exactly matched training size (8,600 examples) it reduces dialectness recovery from 91.8% to 8.1% and region-match recovery from 101.9% to 25.6% on the held-out marker inventory, while the three pipeline-independent measurements do not fall at all. A complementary evaluator sweep defines metric-construction coverage (MCC) and finds measured dialectness recovery increasing monotonically as overlap rises from MCC=0 to MCC=1. Shared construction and evaluation inventories can therefore substantially inflate estimates of synthetic-data recovery.

  • 1 authors
·
Sep 13

Language Models Improve When Pretraining Data Matches Target Tasks

Every data selection method inherently has a target. In practice, these targets often emerge implicitly through benchmark-driven iteration: researchers develop selection strategies, train models, measure benchmark performance, then refine accordingly. This raises a natural question: what happens when we make this optimization explicit? To explore this, we propose benchmark-targeted ranking (BETR), a simple method that selects pretraining documents based on similarity to benchmark training examples. BETR embeds benchmark examples and a sample of pretraining documents in a shared space, scores this sample by similarity to benchmarks, then trains a lightweight classifier to predict these scores for the full corpus. We compare data selection methods by training over 500 models spanning 10^{19} to 10^{22} FLOPs and fitting scaling laws to them. From this, we find that simply aligning pretraining data to evaluation benchmarks using BETR achieves a 2.1x compute multiplier over DCLM-Baseline (4.7x over unfiltered data) and improves performance on 9 out of 10 tasks across all scales. BETR also generalizes well: when targeting a diverse set of benchmarks disjoint from our evaluation suite, it still matches or outperforms baselines. Our scaling analysis further reveals a clear trend: larger models require less aggressive filtering. Overall, our findings show that directly matching pretraining data to target tasks precisely shapes model capabilities and highlight that optimal selection strategies must adapt to model scale.

  • 10 authors
·
Jul 16, 2025

CodeScore: Evaluating Code Generation by Learning Code Execution

A proper code evaluation metric (CEM) profoundly impacts the evolution of code generation, which is an important research field in NLP and software engineering. Prevailing match-based CEMs (e.g., BLEU, Accuracy, and CodeBLEU) suffer from two significant drawbacks. 1. They primarily measure the surface differences between codes without considering their functional equivalence. However, functional equivalence is pivotal in evaluating the effectiveness of code generation, as different codes can perform identical operations. 2. They are predominantly designed for the Ref-only input format. However, code evaluation necessitates versatility in input formats. Aside from Ref-only, there are NL-only and Ref\&NL formats, which existing match-based CEMs cannot effectively accommodate. In this paper, we propose CodeScore, a large language model (LLM)-based CEM, which estimates the functional correctness of generated code on three input types. To acquire CodeScore, we present UniCE, a unified code generation learning framework, for LLMs to learn code execution (i.e., learning PassRatio and Executability of generated code) with unified input. Extensive experimental results on multiple code evaluation datasets demonstrate that CodeScore absolutely improves up to 58.87% correlation with functional correctness compared to other CEMs, achieves state-of-the-art performance, and effectively handles three input formats.

  • 6 authors
·
Jan 21, 2023

FrontierChallenge: Evaluating Scientific Workflow Completion

Scientific agents increasingly analyze data, execute code, and produce research artifacts, yet most benchmarks emphasize final answers, isolated programs, or a single domain. We introduce FrontierChallenge, a cross-domain benchmark comprising 300 end-to-end scientific workflows. In this paper, we release and evaluate 97 of these tasks, spanning quantum chemistry, molecular dynamics, materials characterization, analytical chemistry, life science, and electrochemistry/environment. Each task provides fixed inputs and specifies a bundle of required scientific deliverables. We evaluate twelve frontier models with three agent scaffolds. Pass Rate measures the fraction of tasks satisfying the full-completion criterion, while Avg. Score captures partial progress. Each of the best-performing configurations completed only 20 of the 97 released tasks, yielding a Pass Rate of 20.6%. Partial progress translated especially poorly into complete delivery in analytical chemistry and electrochemistry/environment: Avg. Scores reached 87.6 and 94.9, but the highest Pass Rates were only 4% and 0%. Among non-passing Claude Code trajectories, 75.5% still ended with language claiming completion. These findings show that neither high partial scores nor confident claims of completion reliably indicate that a scientific task has been fully delivered, highlighting the need to evaluate end-to-end workflow execution and the completeness of scientific deliverables together.

apodex Apodex
·
Aug 24 2

Handoff Curves: Initial Selection, Later Handoff, and Supervision Targets in Multi-turn LLM Routing

We study how initial model selection, later handoff, and training targets affect success and cost in LLM tasks that span multiple turns. We compare several router architectures using handoff curves, which record task success and cost as the handoff step changes. Our dataset contains 34,204 complete executions of three Qwen models across BFCL, MultiWOZ, and τ-retail, providing offline supervision for routing decisions. In our primary comparison, with Qwen3.5 35B-A3B as the starting model, allowing a later handoff does not establish an improvement in utility over initial selection on BFCL or MultiWOZ, and increases cost on BFCL across a range of cost weights. Changing the supervision targets improves router performance on both our dataset and TwinRouterBench. On our dataset, replacing noisy training success curves with monotone fits raises success by 0.2 to 4.1 percentage points at the same cost across the reported operating points. On TwinRouterBench, we construct targets that describe the model capability needed for the remaining work and difficult steps ahead. Compared with training on the benchmark's per-step labels, these targets raise trajectory pass scores by 7.8 percentage points and cost savings by 6.3 percentage points on the full static bank. A live deployment of the router trained with these targets resolves 79 of 100 instances at a router cost of 0.59 US dollars per instance.

  • 3 authors
·
Oct 10

SATA-BENCH: Select All That Apply Benchmark for Multiple Choice Questions

Large language models (LLMs) are increasingly evaluated on single-answer multiple-choice tasks, yet many real-world problems require identifying all correct answers from a set of options. This capability remains underexplored. We introduce SATA-BENCH, the first dedicated benchmark for evaluating LLMs on Select All That Apply (SATA) questions across diverse domains, including reading comprehension, law, and biomedicine. Our evaluation of 27 open-source and proprietary models reveals a significant gap: even the strongest model achieves only 41.8% exact match, exposing LLMs' inability to reliably identify all correct answers. We find that this weakness stems from two core challenges: selection bias - models favor certain choices regardless of content, and count bias - models fail to predict the correct number of answers. To address these issues, we propose Choice Funnel, a decoding strategy that combines token debiasing with adaptive thresholding to guide models toward complete and accurate selections. Choice Funnel achieves up to 29% higher exact match than competitive baselines while reducing inference cost by over 64%. Our findings expose fundamental limitations in current LLMs and introduce a new framework for diagnosing and improving multi-answer reasoning. We release SATA-BENCH and Choice Funnel to promote LLM development for robust decision-making in realistic, multi-answer applications.

  • 6 authors
·
May 31, 2025 2

Did We Actually Fix It? An Independent Adversarial Stress-Test of Post-Point-Adjustment Evaluation Metrics for Time-Series Anomaly Detection

Point-adjustment (PA), for years the default scoring protocol in time-series anomaly detection (TSAD), was shown by Kim et al. (2022) to award near-perfect F1 to random anomaly scores. The field adopted a suite of replacement metrics (PA%K, range-based precision/recall, affiliation precision/recall, and Volume-Under-the-Surface, VUS, ROC/PR). We ask, independently and adversarially, whether these resist no-skill detectors on real benchmarks, and find the answer turns entirely on one overlooked variable: N, the number of random attempts an adversary reports the best of. Under a single honest run (N=1), not one replacement metric is gameable on any of six benchmarks (UCR, SMD, SMAP, MSL, NAB, PSM): a random detector reaches 90% of the best real detector's score on at most 11% of series for affiliation-F1, 5% for the ROC family, and 2% for the PR-based metrics and PA%K. But under best-of-N reporting, the seed-shopping endemic to ML, the metrics split sharply. affiliation-F1 and every ROC-based metric inflate steeply, affiliation crossing gameable (25% of series) by N=3 and reaching 0.98 at the full pool (N=41), the ROC family crossing by N=9-11; the PR-based metrics and PA%K stay near-flat at every N, floored near the anomaly prevalence (the lone exception is NAB at large N). A paired test finds VUS-ROC inflated on 131 series where its sibling VUS-PR is not, and never the reverse. The ROC-vs-PR split follows from the order-statistic behaviour of AUC under extreme class imbalance (a random PR-AUC is floored at prevalence); affiliation inflates by a second route, its extreme single-run leniency (already fragile at N=1). We release a pip-installable stress-test harness, and recommend reporting single-run scores or disclosing N and preferring PR-based metrics, which resist best-of-N inflation on nearly every benchmark.

  • 1 authors
·
Jul 18

The Price of Peeking: Anytime-Valid Leakage Detection on ML-KEM EM Traces

Side-channel evaluators routinely inspect leakage tests while acquisition is still running, and extend or stop the campaign based on what they see. Fixed-horizon screening such as the Welch t-test with threshold |t|>4.5 gives no error guarantee for this monitored decision rule. We study anytime-valid leakage detection based on testing by betting: SKIT-type swap-pair e-processes whose false-alarm probability is controlled uniformly over time under an explicit conditional symmetry null. In matched comparisons that share the frozen witness, rows and payoff, first-crossing detection needed 1.68-2.00times the traces of a fixed-horizon randomization test at 80% detection on synthetic streams, and 1.68-2.38times on degraded recordings from an open ML-KEM electromagnetic dataset with the primary Ridge witness at α=0.05. With the same primary witness and level, on undegraded reference and pqm4 recordings the monitored procedure stopped early: its median stopping point was 62-72 and 146-316 evaluation traces, i.e. 2-8% of a conservative 4096-trace budget. Under exact designed nulls on the recorded backgrounds, repeated-look |t|>4.5 screening over all 13000-20000 samples raised a false alarm in 2.7-12.9% of replicates, against 0.0-4.7% for terminal-only screening and no rejection by a sample-wise e-Bonferroni process, which in a prespecified follow-up detected natural-label associations in 4 of 4 backgrounds after 840-3288 traces. All recordings come from one device, and natural-label results are descriptive; we state the assumptions each claim requires.

CPRet: A Dataset, Benchmark, and Model for Retrieval in Competitive Programming

Competitive programming benchmarks are widely used in scenarios such as programming contests and large language model assessments. However, the growing presence of duplicate or highly similar problems raises concerns not only about competition fairness, but also about the validity of competitive programming as a benchmark for model evaluation. In this paper, we propose a new problem -- similar question retrieval -- to address this issue. Due to the lack of both data and models, solving this problem is challenging. To this end, we introduce CPRet, a retrieval-oriented benchmark suite for competitive programming, covering four retrieval tasks: two code-centric (i.e., Text-to-Code and Code-to-Code) and two newly proposed problem-centric tasks (i.e., Problem-to-Duplicate and Simplified-to-Full), built from a combination of automatically crawled problem-solution data and manually curated annotations. Our contribution includes both high-quality training data and temporally separated test sets for reliable evaluation. In addition, we develop two task-specialized retrievers based on this dataset: CPRetriever-Code, trained with a novel Group-InfoNCE loss for problem-code alignment, and CPRetriever-Prob, fine-tuned for identifying problem-level similarity. Both models achieve strong results and are open-sourced for local use. Finally, we analyze LiveCodeBench and find that high-similarity problems inflate model pass rates and reduce differentiation, underscoring the need for similarity-aware evaluation in future benchmarks. Code and data are available at: https://github.com/coldchair/CPRet

  • 5 authors
·
May 19, 2025

PACT: Pairwise-Anchored Calibrated Tuning for Single-Token Typed Decisions

Single-token typed-decision models answer a schema question by reading the logits of a few one-letter answer codes at a single position: they are fast and return a probability for every allowed answer, but they are trained with plain cross-entropy that ignores most of the structure in their training data. We study such a model whose data is curated as contrastive pairs---two contexts that differ in one edited fact that flips the answer---each carrying a machine-checked certificate that deleting the decisive sentence makes the fact unknown. We propose PACT, which turns this structure into four training terms that need no new annotation: a difference-in-differences margin over each pair that is invariant to any shared logit offset, a permutation-consistency term against answer-code position bias, an evidence-necessity term on certificate-verified ablated contexts, and an ordinal transport cost for rubric fields, plus a three-parameter contextual temperature. On a frozen 324-item holdout with three seeds, PACT matches the published recipe in accuracy (84.6% vs. 85.2%; McNemar p ge 0.50 at every seed) while giving the lowest position bias of all runs (answer flips under relabelling 9.8% vs. 13.8%) and the lowest ordinal error on rubric fields (MAE 0.232 vs. 0.311). Against a control with the same optimiser and schedule but cross-entropy only, PACT is significantly more accurate at two of three seeds, halves the seed-to-seed spread and lowers NLL by 26%. Seed-matched ablations and pre-specified falsification tests locate these gains precisely: no single term raises raw accuracy, and the method's value lies in robustness and stability rather than headline accuracy. Code, data splits, trained adapters, and all run records are available at https://github.com/BennyLinntu/PACT-Pairwise-Anchored-Calibrated-Tuning-for-Single-Token-Typed-Decisions.

  • 1 authors
·
Sep 25

Bits or permutations? Spending a fixed inference budget on weight precision or on option-order averaging in small local language models

Local language models can save memory by storing weights in fewer bits, and they can reduce sensitivity to the order of answer options by averaging a few reorderings. We ask whether the first makes the second worse, and which is the cheaper way to spend a fixed budget. On Qwen3-0.6B and Qwen3-1.7B with simulated round-to-nearest weight quantization (8, 6, 4 and 3 bits, fp16 reference), read from the logits of the letters A to D on ARC-Easy, ARC-Challenge, MMLU and HellaSwag (300 items each), 6 bits lose no accuracy against fp16 (they score +1.8 (+1.0 to +2.5) and +0.9 (+0.3 to +1.5) points higher, which we do not explain), and 8 bits change it by -0.7 (-1.1 to -0.3) and -0.4 (-0.7 to -0.2) points. At 4 bits the answer-slot skew rises from 0.15 to 0.53 (0.6B) and from 0.07 to 0.12 (1.7B); Gaussian noise at the same weight SNR gives 0.15 and 0.37 (two draws) and 0.19 (one draw), so the control is inconclusive on whether rounding adds slot skew beyond equal-power noise. Flip rates rise but stay below an accuracy-matched baseline of independent uniform errors. Averaging four cyclic orders recovers +12.6 points at 4 bits for 0.6B (+5.3 at fp16). We price bits and passes in one machine-independent unit, weight bytes read times passes. In cost-matched pairs, extra orders beat extra bits only when the weights have 6 or more bits, and the decision table reduces to: use 6-bit weights and spend the rest of the budget on up to four orders, one pass each. Everything is simulated on two small models with the simplest quantizer; a title-level search found no earlier study of quantization together with option order.

  • 3 authors
·
Nov 9 3

Equivariant Neural Primal-Dual Assignment for Maximum Common Edge Subgraphs

Maximum common edge subgraph (MCES) matching finds a partial vertex correspondence between two labeled graphs that preserves as many labeled edges as possible. Molecular similarity search requires matching many graph pairs, making the cost of repeated queries important. The strongest baseline attains accurate MCES solutions but trains a separate network for each pair. We introduce Equivariant Neural Primal-Dual Assignment (ENPDA), which learns a shared matching policy and applies it to new pairs without further training, answering queries roughly three orders of magnitude faster and recovering its training cost after a few dozen queries. The policy recomputes exact objective marginals for candidate matches and learns corrections and step sizes that update their scores. Target prices respond to competition when several source vertices favor the same target. Four update rounds and a Hungarian projection produce a partial one-to-one matching. We prove per-pair guarantees that hold for any network parameters. In exact arithmetic, reordering either graph permutes the assignment and price states, the projected matching is one-to-one, and repaired prices give a valid MCES upper bound. Subtracting the preserved-edge count bounds the optimality gap; combined with structural caps, these certificates prove global optimality for 60 of 291 native test pairs. On three molecular benchmarks with disjoint train/validation/test splits, ENPDA improves over an analytic counterpart with the same update and projection budget by 7.4-8.6 accuracy points; after one second of refinement search, 2.5-3 points of the gain remain. Transferred without fine-tuning to edge-deletion tasks from social and protein graphs, the policy gains 9.1-17.6 points over the analytic counterpart. When output matchings must keep aromatic rings intact, ENPDA recovers more reference bonds than the baselines on all three datasets.

  • 6 authors
·
Sep 25