HomogeneityNLI

Paper (arXiv) · Project page

HomogeneityNLI is a diagnostic natural language inference (NLI) benchmark. It tests whether language models respect the homogeneity gap of plural definite descriptions (Löbner 1987; Križ 2015; Tieu, Križ & Chemla 2015).

Suppose some but not all of the kids laughed. Then "The kids laughed." is not clearly true, and "The kids didn't laugh." is not clearly true either: both fall into a truth-value gap. Quantified sentences behave differently. "All the kids laughed." is plainly false, and "Not all the kids laughed." is plainly true. HomogeneityNLI tests whether models draw this distinction. It also includes controls that check basic comprehension and a scalar-implicature condition.

• Size: 1,000 items: 100 scenarios × 10 conditions (A–J), 100 items per condition
• Labels: True / False / Neither
• Split: test only (evaluation benchmark)

from datasets import load_dataset
ds = load_dataset("duaanakshbandi/HomogeneityNLI", split="test")

Labels

Each item has a premise that describes a situation in full and a hypothesis about that situation.

Label	Meaning
True	Given the premise, the hypothesis is true.
False	Given the premise, the hypothesis is false.
Neither	Given the premise, the hypothesis is neither true nor false.

Why the gap items are labelled Neither. A plural definite sentence like "The kids laughed." is true if all of the kids laughed and false if none of them did. Its negation, "The kids didn't laugh.", is the mirror image. When the group is mixed, neither condition holds, so neither the sentence nor its negation is true: this is the homogeneity gap. The premise always states exactly how many members have the property, so information is never missing. Neither marks a sentence that is neither true nor false in a fully specified situation, not one whose truth can't be determined from what's given. The readings field records the label each item would get if a model instead treated the definite as existential ("at least one") or universal ("all").

Conditions

Every scenario appears in all ten conditions, so the groups form exact minimal pairs (join on scenario_id). Each scenario is shown in one of three worlds:

World	Example premise
GAP (0 < k < N)	Five students took the exam. Three of them passed the exam and two did not.
ALL (k = N)	Five students took the exam. Every one of them passed the exam.
NONE (k = 0)	Five students took the exam. None of them passed the exam.

Group	World	Hypothesis	Gold label	Role
A	GAP	The students passed the exam.	Neither	Homogeneity target (positive)
B	GAP	The students didn't pass the exam.	Neither	Homogeneity target (negative)
C	GAP	All the students passed the exam.	False	Quantifier baseline
D	GAP	Not all the students passed the exam.	True	Quantifier baseline
E	ALL	The students passed the exam.	True	Definite control
F	ALL	The students didn't pass the exam.	False	Definite control
G	NONE	The students passed the exam.	False	Definite control
H	NONE	The students didn't pass the exam.	True	Definite control
I	ALL	Some of the students passed the exam.	True (literal)	Scalar-implicature target
J	GAP	Some of the students passed the exam.	True	Some control

Notes:

• A/B: a model's answers to the same scenario in A and B show which reading it uses: homogeneous (Neither / Neither), existential (True / False) or universal (False / True). Tieu et al. (2015) classify human participants the same way.
• I: the gold label is the literal reading (True). A model that adds the "not all" implicature answers False; both readings are in readings.
• E–H: these check that the model reads plural definites correctly when the group is uniform, so failures on A/B can't be put down to basic misreading.

Data fields

Field	Type	Description
id	string	<group letter>_<scenario id>, e.g. A_001
scenario_id	int	1–100; shared by all ten conditions of a scenario
group	string	Condition, e.g. A_Gap_Definite_Pos
context	string	World: GAP, ALL or NONE
determiner	string	the, all or some
polarity	string	pos or neg
predicate_class	string	Action, Property, Location or State
animacy	string	animate or inanimate
domain_size	int	Group size N (4, 5, 6, 8 or 10)
k_true	int	Number of members with the property in this world
ratio_bin	string / null	GAP only: near_none, half or near_all; null for ALL/NONE
premise	string	Situation description
hypothesis	string	Sentence to judge
label	string	Gold label: True, False or Neither
readings	struct / null	Label under each interpretation; null where all readings agree. A, B: homogeneous, existential, universal. I: literal, implicature.

Example (A_002):

{
  "id": "A_002", "scenario_id": 2, "group": "A_Gap_Definite_Pos",
  "context": "GAP", "determiner": "the", "polarity": "pos",
  "predicate_class": "Action", "animacy": "animate",
  "domain_size": 5, "k_true": 3, "ratio_bin": "half",
  "premise": "Five students took the exam. Three of them passed the exam and two did not.",
  "hypothesis": "The students passed the exam.",
  "label": "Neither",
  "readings": {"homogeneous": "Neither", "existential": "True", "universal": "False"}
}

Construction

All items are constructed from templates over a hand-written lexicon of 100 scenarios.

Lexicon. There are four predicate classes with 25 scenarios each, and all predicates are distributive:

• Action: eventive verbs, e.g. laughed, passed the exam, flew away
• Property: copular adjectives or properties, e.g. were blue, were made of glass
• Location: copular locatives, e.g. were in the garage, were in the pond
• State: copular states, e.g. were asleep, were open, were frozen

Each scenario has an introduction sentence ("{N} kids heard the joke."), a plural noun phrase and a predicate. Overall, 38 scenarios have animate groups and 62 have inanimate ones.

Domain size and ratio. Each scenario gets a domain size N ∈ {4, 5, 6, 8, 10} (20 scenarios each). In GAP worlds, the number k of members with the property is set by a ratio bin, which also probes non-maximal readings:

• near_none: k = 1 (36 scenarios)
• half: k = ⌈N/2⌉ (32 scenarios)
• near_all: k = N − 1 (32 scenarios)

Templates.

• GAP premises: "{k} of them {P} and {N−k} did not / were not."
• ALL premises: "Every one of them {P}."
• NONE premises: "None of them {P}."
• Hypotheses: built from the plural noun phrase with the, all the, not all the or some of the, followed by the positive or negated predicate (see the Conditions table).

Intended use

HomogeneityNLI is primarily an evaluation benchmark: zero-shot, few-shot or prompted probing of how language models interpret plural definites, quantifiers and some.

Training or fine-tuning on it is allowed, e.g. to study whether the homogeneity gap can be learned and whether it generalizes. If you do this, split by scenario_id (or hold out a whole predicate_class or domain_size) so that no scenario appears in both training and test data, and say so when you report results. Because the items are templated, a model can reach high accuracy from surface patterns alone, so held-out generalization is the meaningful measure. Please don't report zero-shot results from a model trained on these items as if it hadn't seen them.

Limitations

• Templated English only. Sentences are short and repetitive, and they cover a narrow range of syntax and vocabulary, so results may not carry over to naturally occurring text or other languages.

Citation

This dataset accompanies the paper Neither True nor False: Do LLMs Represent Truth-Value Gaps in Plural Predication? If you use HomogeneityNLI, please cite:

@misc{nakshbandi2026neither,
  title         = {Neither True nor False: Do {LLMs} Represent Truth-Value Gaps in Plural Predication?},
  author        = {Nakshbandi, Duaa and An, Zhiyu and Li, Yiping and Giri, Sachin and Du, Wan},
  year          = {2026},
  eprint        = {2610.14438},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2610.14438}
}