Assertion and Belief: Moore's Paradox in Language Models

Anshun Asher Zheng; Weili Cao; Fangcong Yin; David Beaver; Junyi Jessy Li

cs.CLOct 18, 2026

Assertion and Belief: Moore's Paradox in Language Models

Authors: Anshun Asher Zheng, Weili Cao, Fangcong Yin, David Beaver, Junyi Jessy Li

Organizations: The University of Texas at Austin · Duke University · New York University · University of Graz

Abstract

Moore's paradox \citep{moore1942} concerns sentences such as ``It is raining, but I don't believe that it is,'' which may well be true yet cannot be sensibly asserted. The absurdity lies not in the content but in the act, since to assert is to vouch for one's own belief. This makes the paradox a test of whether language models take the speech act into account when computing meaning. We present the first test of Moore's paradox in LLMs, evaluating 12 models on a curated set of Moorean sentences in varied contexts. We find that models largely recognize the infelicity of Moorean sentences, and that the judgement is not tied to a memorized form and tracks the speech act. Using the Rational Speech Act framework (RSA; \citealt{goodmanfrank2016}), we then test whether these judgements arise from a listener reasoning about a speaker. We find the models' own listeners do not reason about their own speakers, and their reasoning traces suggest that they instead recall knowledge about the paradox and apply rules therein. The models thus appear to have learned the paradox as a fact about sentences, not as a requirement on speakers.

Figures & tables

Act	State	Felicitous
Assertion	It is raining; Robin is ignorant of the rain. She shares her view over the weather.	✘ No
Reading	Robin is asked to read a line on a worksheet “It is raining”; she does not believe it.	✔ Yes

Table 1: In both situations the speaker, Robin, says “It is raining, but I don’t believe that it is.”

Figure 1: Sensibility judgements. Top: share of No and Yes answers per sentence type, pooled over models. Bottom: share of Yes per model family. Per-model results in Appendix D , Figure 5 .

Figure 2: Form test (left) and context test (right). Each point is a model family’s share of Yes to “Does what Robin says make sense?”. Left: the paraphrases and insertions against the original Moorean sentence. Right: the sentence read aloud against the sentence asserted, for the Moorean sentence, the unread Moorean control (“It’s my turn, but I don’t believe that”) and the coherent sentence. On the diagonal, the change makes no difference.

	World	Evidence	Belief	Act	$M$
$s_{1}$	$p$	$p$	$p$	asserts
$s_{2}$	$p$	none	none	asserts
$s_{3}$	$p$	$p$	$\neg p$	asserts	✓
$s_{4}$	$\neg p$	$\neg p$	$\neg p$	reads “ $p$ ”	✓

Table 2: The four theory-driven states (✓:

M

is felicitous). Full descriptions are in Appendix L , Table 4 .

Figure 3: RSA over

s_{1}

s_{4}

(uniform prior). (a) Literal truth and the norms of speech acts (

A

). The norm of reading, a non-assertive (NA) act, and then a norm of assertion (JNA, or the alternative BNA) override cells of Truth (bold). (b) RSA with NA and JNA.

S_{1}

multiplies each row of

A

by the informativity of each utterance from

L_{0}

and normalizes over utterances. In the

M

column, green marks the states where

M

should be possible and red the states where

M

should be ruled out (Table 2 ). Outlined cells mark wrong predictions.

Figure 4: Left: the elicited speaker’s agreement with each candidate speaker; the models’ speakers agree more with BNA than with JNA. Right: the false positive rate measures how often the speaker would not say

M

while the listener considers

M

possible, and the false negative rate measures how often the speaker would say

M

while the listener rejects

M

. Per-model results in Appendix N .

Appendix figures & tables28 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Parameters	Source	Reasoning on
Qwen3.5-4B	4B	Qwen/Qwen3.5-4B	native thinking
Qwen3.5-9B	9B	Qwen/Qwen3.5-9B	native thinking
Qwen3.5-27B	27B	Qwen/Qwen3.5-27B	native thinking
Qwen3.5-35B-A3B	35B (3B active)	Qwen/Qwen3.5-35B-A3B	native thinking
K2-Horizon-3.7B	3.7B	IFM/K2-Horizon-3.7B	effort high
K2-Horizon-7B	7B	IFM/K2-Horizon-7B	effort high

Appendix

Table 3: Models. Open-weight models are served with vLLM at temperature 1.0 and top-

p

0.95; Claude Haiku uses API defaults. With reasoning off, thinking is disabled (K2 closes its thinking block at once) and answers are limited to 16 tokens. With reasoning on, the output budget is 12,288 tokens. Qwen and Gemma have no effort setting.

Figure 5: Sensibility judgements per model, laid out as Figure 1 . Side bars:

|M-\text{coherent}|-|M-\text{contradiction}|

, orange where the Moorean sentence is judged closer to the contradiction, blue where closer to the coherent sentence.

Figure 6: Acceptability ratings (1-7) per model, laid out as Figure 5 .

Figure 7: Form test per model: share of Yes for the original Moorean sentence and its paraphrased (mean of three) and inserted (mean of two) versions, in the bands of Figure 1 . Outlined: a changed form accepted although the original is rejected.

Figure 8: Context test per model: the gap between the Yes rates for the coherent and the Moorean sentence when Robin asserts the sentence (open circle) and when she reads it aloud (filled circle). Below zero, the Moorean sentence is accepted more often than the coherent one.

Figure 9: Share of reasoning traces that name Moore, for the Moorean sentence and its changed forms (form test) and for the Moorean sentence asserted or read aloud in the classroom setting (context test).

Figure 10: RSA with a vanilla speaker, who may say anything literally true (uniform prior). Colours as in Figure 3 .

	State	World	Evidence	Belief	Act	$M$	Share (%)
$s_{1}$	The world is $p$ . Robin checked, and it looked $p$ (or a sign she trusts says $p$ ); she believes $p$ .	$p$	$p$	$p$	asserts		20.0
$s_{2}$	The world is $p$ . Robin has no evidence about $p$ and no view.	$p$	none	none	asserts		16.0
$s_{3}$	The world is $p$ . Robin checked, and it looked $p$ , but a feeling or emotion makes her believe $\neg p$ .	$p$	$p$	$\neg p$	asserts	✓	2.0
$s_{4}$	The world is $\neg p$ . Robin’s notes say $p$ , and she is asked to read them aloud; she looked, and it did not look $p$ ; she believes $\neg p$ .	$\neg p$	$\neg p$	$\neg p$	reads “ $p$ ”	✓	4.7
$s_{5}$	The world is $\neg p$ . Robin checked, and it did not look $p$ ; she believes $\neg p$ .	$\neg p$	$\neg p$	$\neg p$	asserts		21.4
$s_{6}$	The world is $p$ . From one side it looked $p$ , from the other it did not; she has no view.	$p$	both	none	asserts		7.1

Appendix

Table 4: The states. Section 5 uses

s_{1}

s_{4}

; Section 6 uses all twelve, built from the models’ proposals (Share: proportion of proposals mapped to each state).

Figure 11: Reasoning traces by the shape of their chain across three contexts. Per-model results in Appendix Q , Figure 28 .

Figure 12: The state-utterance pairs on which the candidate speakers differ, grouped by the layer that separates them, for the candidates (top) and each model without and with reasoning. Filled: feasible, in the colour of the layer; light: not feasible; grey: undetermined. Pairs on which all candidates agree are not shown. Agreement scores in Figure 13 .

Figure 13: Agreement of each model’s speaker with each candidate (Eq. 5 ), computed over the 24 diagnostic pairs on which the candidates differ (Figure 12 ). On the other pairs, all candidates agree, so they shift every candidate’s agreement equally. Undetermined pairs count as disagreements. Outlined: closest.

Figure 14: Qwen 4B’s grids.

\hat{S}_{1}

is coloured against the vanilla speaker computed from the model’s

\hat{L}_{0}

: green, feasible though the vanilla speaker would not say it; orange, not feasible though it would; grey, undetermined.

Figure 15: Qwen 9B: grids, coloured as in Figure 14 .

Figure 16: Qwen 27B: grids, coloured as in Figure 14 .

Figure 17: Qwen 35B-A3B: grids, coloured as in Figure 14 .

Figure 18: K2 3.7B: grids, coloured as in Figure 14 .

Figure 19: K2 7B: grids, coloured as in Figure 14 .

Figure 20: K2 32B: grids, coloured as in Figure 14 .

Figure 21: K2 36B-A4B: grids, coloured as in Figure 14 .

Figure 22: Gemma 12B: grids, coloured as in Figure 14 .

Figure 23: Gemma 26B-A4B: grids, coloured as in Figure 14 .

Figure 24: Gemma 31B: grids, coloured as in Figure 14 .

Figure 25: Haiku 5.5: grids, coloured as in Figure 14 .

Figure 26: False positives for every utterance: the share of the states where the speaker would not say it that the listener considers possible.

Figure 27: False negatives in context for every utterance: the share of the contexts where the model’s speaker would say it in which the model still judges it not to make sense.

Figure 28: Each model’s reasoning traces by the shape of their chain, out of context, in the skywalk case (

s_{3}

), and for the reader (

s_{4}

). Blue: rejections; orange: acceptances. Right: share correct.

Move	Traces
States the verdict	.92
Parses the sentence into a claim and a denial of belief	.73
Restates the scenario (Robin’s evidence, notes, feelings)	.59
Checks for a logical contradiction	.57
States that sincere assertion implies belief	.48
Names Moore’s paradox	.47

Appendix

Table 5: The 13 moves found by grouping the reasoning steps, with the share of traces that make each.

Chain	No context	$s_{3}$	$s_{4}$
rule $\to$ No	.70	.11	.06
rule $\to$ reinterpretation $\to$ Yes	.11	.32	.31
reinterpretation $\to$ Yes	.06	.41	.52
rule $\to$ logic checks $\to$ Yes	.12	.10	.05
logic checks $\to$ Yes	.01	.06	.04

Appendix

Table 6: The shapes of the reasoning chains, with the share of traces in each setting (other rejections: .01).

Pragmatic Reasoning in Language Models

Pragmatic reasoning in language models studies how models infer and communicate meanings shaped by context, intent, shared knowledge, and social norms.

See all 103 papers

Jul 20, 2026

It's Not What You Say, It's How You Say It: Evaluating LLM Responses to Expressions of Belief

Users frequently express their beliefs to large language models (LLMs). In some situations, the LLM should accept these contextual beliefs as true. In others, they should stick to their prior knowledge. Notably, users' expressions of belief (EoBs) can take linguistically diverse forms - using presuppositions, evidential and certainty markers, or varied tones - each of which may have a different persuasiveness over the LLMs. We introduce a typology to systematically evaluate how different EoBs affect whether models follow context versus prior knowledge. The typology is grounded in four linguistically motivated dimensions: form, evidentiality, epistemic stance, and tone, spanning 17 fine-grained types. By pairing these EoBs with world knowledge facts, we generate controlled EoB-query pairs that isolate the effect of linguistic variation. Using this benchmark, we evaluate 16 LLMs that differ in architecture (Llama3, Qwen3, Gemma3), scale (1B-30B parameters), and training stages (base vs instruct). We identify meaningful variations in response behavior across these axes, e.g., that bigger models and instruction models tend to be less context-following than smaller models and base models. We further identify specific EoBs that statistically significantly persuade LMs more consistently than others. Our work reveals systematic patterns in how linguistic framing affects LLM context integration, with implications for prompt engineering and model robustness.

Kevin Du, Clara Kümpel, Michelle Wastl +1

ETH Zürich · University of Zurich · UC San Diego

LLM Evaluation Knowledge Conflicts in Language Models Pragmatic Reasoning in Language Models Persuasion in Language Models Language Model Robustness

Ask PDF

Aug 5, 2026

Same Formulas, Different Semantics: Do Language Models Follow Modal Logic Specifications?

Reasoning about necessity and possibility depends on assumptions about accessibility between worlds and about which objects exist at each one. The same inference may therefore hold under one modal system and fail under another. Evaluating language models on such problems requires testing whether their judgments follow the stated semantics rather than a familiar logic. We construct paired modal problems with identical premises and conjecture but different frame or domain conditions; automated reasoning verifies opposite labels. A balanced core prevents the semantic condition alone from revealing the answer. On this core, four of five recent models perform below the condition-only baseline under direct prompting. Yet enabling reasoning mode raises DeepSeek V4 Flash from 4.4% to 88.1% on unchanged prompts. Following stipulated modal semantics thus depends strongly on inference mode as well as model identity. When frame conditions are omitted, models often agree but fit different familiar logics best. We release the formulas, oracle artifacts, countermodels, and responses.

Réemi Andrieu, Damien Sileo

Univ. Lille, Inria, CNRS, Centrale Lille, UMR 9189 - CRIStAL, F-59000 Lille, France

Logical Reasoning LLM Evaluation Large Reasoning Models

Ask PDF

Feb 17, 2026

Language Statistics and False Belief Reasoning: Evidence from 41 Open-Weight LMs

Research on mental state reasoning in language models (LMs) has the potential to inform theories of human social cognition--such as the theory that mental state reasoning emerges in part from language exposure--and our understanding of LMs themselves. Yet much published work on LMs relies on a relatively small sample of closed-source LMs, limiting our ability to rigorously test psychological theories and evaluate LM capacities. Here, we replicate and extend published work on the false belief task by assessing LM mental state reasoning behavior across 41 open-weight models (from distinct model families). We find sensitivity to implied knowledge states in 34% of the LMs tested; however, consistent with prior work, none fully

explain away'' the effect in humans. Larger LMs show increased sensitivity and also exhibit higher psychometric predictive power. Finally, we use LM behavior to generate and test a novel hypothesis about human cognition: both humans and LMs show a bias towards attributing false beliefs when knowledge states are cued using a non-factive verb (

John thinks...'') than when cued indirectly (``John looks in the...''). Unlike the primary effect of knowledge states, where human sensitivity exceeds that of LMs, the magnitude of the human knowledge cue effect falls squarely within the distribution of LM effect sizes-suggesting that distributional statistics of language can in principle account for the latter but not the former in humans. These results demonstrate the value of using larger samples of open-weight LMs to test theories of human cognition and evaluate LM capacities.

Sean Trott, Samuel Taylor, Cameron Jones +2

Rutgers University - Newark · UC San Diego · Stony Brook University +1

Theory of Mind Pragmatic Reasoning in Language Models

Ask PDF

Abstract

Pragmatic Reasoning in Language Models

Explore similar work

It's Not What You Say, It's How You Say It: Evaluating LLM Responses to Expressions of Belief

Same Formulas, Different Semantics: Do Language Models Follow Modal Logic Specifications?

Language Statistics and False Belief Reasoning: Evidence from 41 Open-Weight LMs