Upload PAVO_TMLR_submission.tex with huggingface_hub

PAVO_TMLR_submission.tex

@@ -148,10 +148,10 @@ Q_i(c_i,u_i) \geq \theta_j(c_j) \quad \forall\,(v_i,v_j)\in\mathcal{E}
 \label{eq:coupling}
 \end{equation}
 
-The central empirical finding of our coupling analysis is that ASR errors propagate to downstream LLM quality in a \emph{regime-dependent} manner: factual tasks exhibit a sharp accuracy cliff at low word-error rates, while semantic tasks degrade gracefully across a wide WER range. This two-regime structure governs the design of PAVO's quality constraint and explains why a single conservative threshold is both necessary and sufficient for pipeline routing.
+The central empirical finding of our coupling analysis is that ASR errors propagate to downstream LLM quality in a \emph{regime-dependent} manner: all models tolerate moderate WER (up to $\sim$10\%) before exhibiting a capacity-dependent accuracy cliff, while semantic tasks degrade gracefully across a wide WER range. This two-regime structure governs the design of PAVO's quality constraint and explains why coupling constraints are necessary for pipeline routing.
 
-\paragraph{Regime 1: Factual coupling (sharp cliff).}
-The primary coupling calibration uses $n=200$ queries at each of 9 WER levels for three models on H100 ($n_{\text{total}} = 5{,}400$ LLM calls; Table~\ref{tab:gpu_coupling}). All three models maintain high exact-match ($\geq$0.92) through 10\% WER, then degrade: Llama~3.1 8B drops to 0.835 at 15\% and 0.750 at 20\%; Mistral~7B drops to 0.870 and 0.835; Gemma2~2B drops to 0.855 and 0.810. The degradation ordering (8B $>$ 7B $>$ 2B robustness) is consistent with model-capacity-dependent coupling. A preliminary calibration on Apple M3 (30 factual QA pairs at 2\% WER increments; Table~\ref{tab:coupling_measured}) shows consistent directionality---Llama drops from 0.97 to 0.63 at $\theta = 2\%$ (Fisher exact $p = 0.038$)---but with wide CIs ($[0.44, 0.80]$) that the H100 calibration resolves. The cliff profile is consistent with factual queries' reliance on verbatim named entities and numerical values, where a single substitution error can alter the answer. The model-size dependence---smaller models are more sensitive---implies coupling thresholds should be calibrated per LLM. This regime applies to L1--L2 (55\% of turns).
+\paragraph{Regime 1: Factual coupling (capacity-dependent cliff).}
+The primary coupling calibration uses $n=200$ queries at each of 9 WER levels for three models on H100 ($n_{\text{total}} = 5{,}400$ LLM calls; Table~\ref{tab:gpu_coupling}). All three models maintain high exact-match ($\geq$0.92) through 10\% WER, then degrade sharply: Llama~3.1 8B drops to 0.835 at 15\% and 0.750 at 20\%; Mistral~7B drops to 0.870 and 0.835; Gemma2~2B drops to 0.855 and 0.810. The degradation ordering (8B $>$ 7B $>$ 2B robustness) is consistent with model-capacity-dependent coupling. A preliminary calibration on Apple M3 (30 factual QA pairs at 2\% WER increments; Table~\ref{tab:coupling_measured}) shows consistent directionality---Llama drops from 0.97 to 0.63 at $\theta = 2\%$ (Fisher exact $p = 0.038$)---but with wide CIs ($[0.44, 0.80]$) that the H100 calibration resolves. The model-size dependence---smaller models are more sensitive to upstream noise---implies coupling thresholds should be calibrated per LLM. This regime applies to L1--L2 (55\% of turns).
 
 \paragraph{Regime 2: Semantic coupling (graceful degradation).}
 On the H100 GPU, quality scores for all three models remain within 0.01 of baseline through 10\% WER (Table~\ref{tab:gpu_coupling}): Llama 0.870--0.876, Mistral 0.869--0.884, Gemma 0.854--0.874. Degradation appears only at 15--20\% WER, where quality drops 0.08--0.10 from baseline. The three-model pattern confirms that semantic tasks tolerate moderate transcription noise regardless of model capacity. This regime applies to L3--L5 (open-ended, emotional, tool-use queries).
@@ -163,9 +163,9 @@ PAVO adopts a uniform quality threshold $\theta = 2\%$ WER. This choice is conse
 Three properties make voice pipeline coupling distinct from generic error propagation and motivate enforcement via logit masking rather than reward shaping: (1)~the factual cliff means soft penalties allow the RL policy to exploit the borderline-infeasible region during training; (2)~the threshold is task-dependent, requiring empirical characterization per evaluation regime; and (3)~the constraint must be enforced in $<$3\,ms at inference time. Prior routing systems~\citep{clipper2017,romero2021infaas,gujarati2023shepherd} do not model these cross-stage dependencies.
 
 \paragraph{Regime stability via bootstrap.}
-We bootstrap-resample the H100 coupling data ($n=5{,}400$) 10{,}000 times. For all three models, quality at 15--20\% WER is significantly lower than at 0--10\% WER ($P > 0.99$). Across all iterations, quality at WER $\leq 1\%$ exceeds quality at WER $\geq 2\%$ with $P > 0.99$. The two-regime structure is corroborated by four independent conditions: the component ablation (Table~\ref{tab:component_ablation}), all 21 noise conditions, cross-dataset evaluation (LibriSpeech, FLEURS; 800 samples), and 100 LLM quality measurements under noise (0\% error rate up to 13.74\% WER).
+We bootstrap-resample the H100 coupling data ($n=5{,}400$) 10{,}000 times. For all three models, quality at 15--20\% WER is significantly lower than at 0--10\% WER ($P > 0.99$), confirming that a degradation cliff exists, though at higher WER than the conservative $\theta=2\%$. This means $\theta=2\%$ provides a substantial safety margin. The two-regime structure is corroborated by five independent conditions: the component ablation (Table~\ref{tab:component_ablation}), all 21 noise conditions, cross-dataset evaluation (LibriSpeech, FLEURS; 800 samples), real ASR error coupling across 6 ASR--LLM combinations, and 100 LLM quality measurements under noise (0\% error rate up to 13.74\% WER).
 
-\paragraph{Calibration limitations.} The M3 calibration ($n = 30$, 2\% increments) has wide CIs; the H100 calibration ($n=200$ per level per model, three models) reveals model-dependent degradation profiles. Finer-grained injection (0.5\% increments) could sharpen the cliff for larger models. We adopt $\theta = 2\%$ as a conservative choice covering both model families, so calibration error biases toward over-constraining rather than under-constraining.
+\paragraph{Calibration limitations.} The M3 calibration ($n = 30$, 2\% increments) has wide CIs; the H100 calibration ($n=200$ per level per model, three models) reveals model-dependent degradation profiles. Finer-grained injection between 10\% and 15\% could pinpoint the exact cliff location per model. We adopt $\theta = 2\%$ as a conservative choice covering all three model families, so calibration error biases toward over-constraining rather than under-constraining.
 
 \begin{lemma}[Monotonicity]
 \label{lem:mono}
@@ -434,7 +434,7 @@ Whisper-large-v3 WER exceeds $\theta=2\%$ across all 21 tested conditions (Figur
 
 \begin{table}[t]
 \centering
-\caption{Measured coupling on H100: exact-match accuracy and quality score vs.\ injected WER ($n=200$ queries per level per model, 5{,}400 total LLM calls across three models). Gemma2~2B exhibits a sharp cliff at WER$\,=1\%$; Llama~3.1 8B degrades gradually; Mistral~7B falls between the two, consistent with model-capacity-dependent coupling.}
+\caption{Measured coupling on H100: exact-match accuracy and quality score vs.\ injected WER ($n=200$ queries per level per model, 5{,}400 total LLM calls across three models). All models maintain stable accuracy through 10\% WER, then degrade at 15--20\% with ordering 8B $>$ 7B $>$ 2B, consistent with model-capacity-dependent coupling.}
 \label{tab:gpu_coupling}
 \small
 \begin{tabular}{llrrrrrrrrr}
@@ -457,7 +457,7 @@ Model & Metric & 0\% & 1\% & 2\% & 3\% & 5\% & 8\% & 10\% & 15\% & 20\% \\
 \end{table}
 
 \paragraph{Real end-to-end pipeline validation.}
-To validate the simulated pipeline against real speech, we run a full Whisper-large-v3 $\to$ Llama~3.1 8B pipeline on 100 LibriSpeech samples on H100 (no synthetic WER injection). The real pipeline achieves mean E2E latency 952\,ms (P95: 1{,}198\,ms) with BERTScore (RoBERTa-large) 0.818 and BERTScore (DeBERTa) 0.529. The P95 latency is 17\% below the simulated cloud estimate (1{,}198 vs.\ 1{,}453\,ms in Table~\ref{tab:e2e}), confirming the simulation is conservative. We also measure coupling with real Whisper ASR errors (not synthetic injection) on 100 LibriSpeech samples across six ASR--LLM combinations. BERTScore (DeBERTa) with Whisper-large consistently exceeds Whisper-tiny (Llama: 0.527 vs.\ 0.522; Gemma: 0.528 vs.\ 0.523; Mistral: 0.557 vs.\ 0.543), confirming that coupling---higher ASR error degrades downstream LLM quality---holds with real transcription errors, not only synthetic injection.
+To validate the simulated pipeline against real speech, we run Whisper-large-v3 $\to$ Llama~3.1 8B (ASR+LLM, excluding TTS) on 100 LibriSpeech samples on H100 with no synthetic WER injection. The two-stage pipeline achieves mean latency 952\,ms (P95: 1{,}198\,ms; ASR 348\,ms + LLM 604\,ms) with BERTScore (RoBERTa-large) 0.818 and BERTScore (DeBERTa) 0.529. These measurements complement Table~\ref{tab:e2e} (which includes TTS overhead). We also measure coupling with real Whisper ASR errors (not synthetic injection) on 100 LibriSpeech samples across six ASR--LLM combinations. BERTScore (DeBERTa) with Whisper-large consistently exceeds Whisper-tiny (Llama: 0.527 vs.\ 0.522; Gemma: 0.528 vs.\ 0.523; Mistral: 0.557 vs.\ 0.543), confirming that coupling---higher ASR error degrades downstream LLM quality---holds with real transcription errors, not only synthetic injection.
 
 \subsection{Multi-configuration benchmark results}
 \label{sec:benchmark_results}