Pith. sign in

REVIEW 3 major objections 5 minor 72 references

Pragmatic Theories Enhance Understanding of Implied Meanings in LLMs

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that prompting a language model with a short summary of Gricean pragmatics or Relevance Theory improves its answers to implied-meaning questions by up to 9.6 percentage points over zero-shot chain-of-thought.

desk verdict A useful, honest prompt-engineering paper with strong confound controls; the headline gain is probably real, but single-sample runs make the smaller gains unverified and the 'consistent improvement' claim overreaches. read the letter →

arxiv 2510.26253 v3 pith:E7LZJGSA submitted 2025-10-30 cs.CL

Takuma Sato , Seiya Kawano , Koichiro Yoshino This is my paper · ORCID
classification cs.CL
keywords pragmaticsimpliedmeaningin-contextlearningzero-shotchain-of-thoughtGriceanmaximsRelevanceTheorypromptingpragmaticreasoning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a language model can be made noticeably better at interpreting implied meanings by adding a short summary of a linguistic pragmatic theory to the prompt, with no fine-tuning, no examples, and no question-specific hints. On a multiple-choice benchmark covering five kinds of non-literal meaning, asking the model to reason step by step in line with Gricean pragmatics or Relevance Theory improved accuracy in nearly every tested model, by up to 9.6 percentage points over ordinary zero-shot chain-of-thought. The effect was largest on irony and on utterances that flout Grice's maxims, and even merely naming a theory helped larger models by 1-3%. If the result holds, it gives a cheap, model-agnostic way to elicit pragmatic reasoning that is already latent in language models.

What carries the argument

The central object is a theory summary embedded in the prompt: Gricean pragmatics contributes the Cooperative Principle, the four maxims of Quantity, Quality, Relation, and Manner, and the inference rule that a detected maxim violation signals a conversational implicature; Relevance Theory contributes the definition of relevance as cognitive effect divided by processing effort, with the listener seeking the interpretation that gives the most effect for the least effort. The prompt instructs the model to articulate its step-by-step reasoning in line with one of these theories before producing the final answer. This scaffold works without any instance-specific examples: it names a knowledge do

What would settle it

Re-run the same set of 520 questions under each prompt method multiple times with different random seeds at temperature 0.8, or once with greedy decoding, and compute paired differences with confidence intervals; if the theory prompts are not consistently above zero-shot chain-of-thought by more than the run-to-run spread, the central claim collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that prompt-embedded summaries of pragmatic theories improve how language models understand implied meanings. The proposed prompts briefly explain Grice's Cooperative Principle and conversational maxims, or Relevance Theory's balance of cognitive effects against processing effort, then ask the model to write its reasoning in line with the theory before choosing an answer. Across open and closed models, the theory prompts beat a simple answer-only prompt and a zero-shot chain-of-thought prompt; the best model improved from about 0.88 to 0.94, surpassing the reported human score, and another model reached human-level accuracy. Control prompts with unrelated theorie

Load-bearing premise

The reported accuracy gains are larger than the random variation between single stochastic answers: each item was answered once at temperature 0.8, with no repeated runs or paired significance testing, so the measured gaps could partly be sampling noise.

Editorial extensions

If this is right

  • The same zero-shot theory-summary prompt can be applied across multiple models and five different pragmatic phenomena without knowing which phenomenon a question belongs to.
  • No training or parameter updates are needed, so the method can be dropped into existing systems that already use prompting.
  • Theory-name-only prompts producing smaller but real gains in larger models implies that substantial pragmatic knowledge is already stored in model parameters and can be partially activated by a domain label.
  • The largest measured gains on irony and maxim-flouting utterances suggest the method helps most where the literal meaning is plainly insufficient.
  • Because unrelated and fictitious theory prompts did not match the real theory prompts, the benefit appears tied to the content of the pragmatic theory, not to the act of following any structured framework.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension: if the gain comes from matching the prompt's theory to the task domain, then analogous theory summaries from other expert fields (e.g., legal interpretation, statistical reasoning) should boost performance on tasks in those fields, while mismatched expert labels should not.
  • The 1-3% name-only effect is the most fragile result because the experiments use one stochastic sample per item; repeated sampling or greedy decoding would show whether that small effect survives.
  • The paper's explanation for why Gricean prompting usually beat Relevance prompting—more frequent exposure to Gricean theory in pretraining—could be tested by measuring whether models whose training data contains more Relevance-theory text show the reverse ordering.
  • Running the same prompt recipe on non-English conversational data would test whether the benefit depends on English-centric training exposure, a direction the paper itself flags as open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a zero-shot in-context learning method that prepends summaries of pragmatic theories (Gricean pragmatics or Relevance Theory) to prompts, instructing LLMs to reason in line with the theory before answering. The method is evaluated on five phenomena of the PRAGMEGA benchmark (520 items) across six open and closed LLMs, and compared with a direct-answer baseline and zero-shot Chain-of-Thought. The authors report accuracy improvements up to 9.6 absolute points (phi-4, Gricean prompting) and smaller 1–3% gains when only the theory names are mentioned. Additional experiments with unrelated theories, fictitious theories, and a general planning prompt are included as confound controls.

Significance. If the reported effect is real, the paper offers a simple, model-agnostic, zero-shot prompt recipe for pragmatically challenging tasks. The confound-control experiments are a genuine strength: using X-bar theory, computational complexity, or invented 'pragmatic' theories generally yields lower accuracy, which supports the claim that the specific theory content matters. The code is released, and the limitations are stated candidly. However, the central claim rests on accuracy differences that are not shown to be statistically robust, especially for the smaller short-prompt gains. The result is potentially valuable but needs stronger empirical grounding before it can support the paper's general conclusions.

major comments (3)
  1. [§4.1, Fig. 1, Appendix E] All reported accuracies come from a single stochastic sample per item (temperature=0.8, do_sample=True, no seeds). The Wilson confidence intervals in Fig. 1 are computed across items for each method, not across sampling runs, so they do not capture run-to-run variance in model outputs. Consequently, a 1–3 point difference—e.g., GPT-4o grice_short 0.902 vs. cot 0.877, or Qwen2.5-14B grice_short 0.723 vs. cot 0.696—cannot be distinguished from sampling noise. The 9.6-point phi-4 gain is large, but the paper's broader claim of consistent improvement requires repeated sampling, seeds, and paired significance tests (e.g., McNemar or bootstrap over runs/items). Please add these and report confidence intervals across runs.
  2. [§4.1, Table 7] The paper states that the proposed methods consistently improved performance, but the authors' own results contain exceptions: relevance on Llama-3.1-8B-Instruct (0.604) is below simple (0.621), and grice_short on Qwen2.5-7B-Instruct (0.592) is below simple (0.610). The text acknowledges the Llama relevance case in §4.1 but the overall conclusion of 'consistent improvement' is still too strong. The authors should either provide a statistical analysis that supports the pattern across the full set of models or explicitly narrow the claim to the methods/conditions where the evidence is adequate.
  3. [§5.2, Table 8, Fig. 4] The confound experiments are well designed, but the interpretation 'no experimental evidence was found to support the concerns' is not supported by significance testing. In particular, distance_short (a fictitious theory, 0.900 on GPT-4o) nearly matches grice_short (0.902), and majoritarian variants sometimes approach the proposed methods. Because these are single draws, the conclusion that the real pragmatic theories are uniquely effective would be more convincing with paired comparisons across repeated runs. Please add such tests, or at least hedge the conclusion accordingly.
minor comments (5)
  1. [Throughout] The name 'Sperbel' is used consistently in the text and references; the correct spelling is 'Sperber' (Dan Sperber). Please correct.
  2. [Appendix D / Table 4] The X-bar theory prompt contains an unresolved '(??)' placeholder. This should be fixed before publication.
  3. [Tables 1 and 15] Table 15 in the appendix appears to repeat/overlap with Table 1 in the main text. Please verify the intended examples and cross-references to §J.
  4. [Fig. 2] The resolution of the per-phenomenon figure is low; the legend and axis labels are hard to read. A higher-resolution figure would help.
  5. [§4.2.1 and Appendix H] The correlation analysis (Pearson r = 0.181 and 0.211) is reported without significance tests or scatterplot details. Adding p-values or confidence intervals would be helpful, though this is a minor point.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper is an empirical prompting study whose measured accuracies come from external ground-truth labels, with no fitted parameters or definitionally linked predictions.

full rationale

This paper's central claim is empirical: that adding summaries of Gricean pragmatics or Relevance Theory to zero-shot prompts improves accuracy on PRAGMEGA implied-meaning questions. There is no derivation chain in which an output is shown to equal an input by construction. Accuracy is measured against fixed external benchmark answers; the theory summaries are fixed prompt inputs chosen before the experiments; no parameter is fitted to the evaluation data; and the reported gains are observed results rather than predictions forced by definitions. The control experiments with unrelated theories (X-bar, computational complexity, graph theory), fictitious theories (Majoritarian Consensus Pragmatics, Distance Primacy Pragmatics), and an alternative general prompt template (plan) directly test whether the effect reduces to 'any plausible-looking prompt,' which further reduces any circularity concern. The only apparent self-citation, Tanaka et al. (2024) in the introduction, is used as background motivation for why understanding implied meanings matters for robots/AI; it is not load-bearing for the method's effectiveness. The limitations stated in the paper—lack of verification on upstream tasks or other languages, and lack of a full account of why the method works—are honestly acknowledged gaps, not circular steps. The statistical concern that each instance is evaluated with a single stochastic sample (temperature=0.8, do_sample=True, no seeds or repeated runs) is a question of robustness and significance testing, not of circularity. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters or invented entities are introduced. The empirical claims rest on benchmark validity, single-sample stochastic inference, fidelity of the theory summaries, and comparability of human baselines.

assumptions (4)
  • domain assumption PRAGMEGA multiple-choice accuracy is a valid measure of implied-meaning understanding.
    The entire evaluation rests on PRAGMEGA as ground truth for the five selected pragmatic phenomena (Section 3.1).
  • domain assumption One stochastic sample per instance at temperature=0.8 adequately represents model performance.
    Appendix E specifies sampling hyperparameters but no seeds or repeated runs; the reported accuracy differences are treated as reliable without paired significance testing.
  • domain assumption The authors' summaries of Gricean and Relevance Theory are faithful enough to elicit the intended reasoning.
    The prompts in Appendix D are author-written glosses; the method assumes these glosses activate relevant latent knowledge rather than mislead the model.
  • domain assumption Human scores from Hu et al. (2023) are comparable to the model scores despite evaluating only five of the seven PRAGMEGA phenomena.
    Figure 1 plots human scores from the original PRAGMEGA paper against the subset used here; the paper does not re-collect human judgments on the same subset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pragmatic Theories Enhance Understanding of Implied Meanings in LLMs." pith.science (2026). https://pith.science/paper/E7LZJGSA

@misc{pith2026251026253,
  author       = {Pith},
  title        = {Pith review of: Pragmatic Theories Enhance Understanding of Implied Meanings in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E7LZJGSA}},
  note         = {Machine review of arXiv:2510.26253}
}
read the original abstract

The ability to accurately interpret implied meanings plays a crucial role in human communication and language use, and language models are also expected to possess this capability. This study demonstrates that providing language models with pragmatic theories as prompts is an effective in-context learning approach for tasks to understand implied meanings. Specifically, we propose an approach in which an overview of pragmatic theories, such as Gricean pragmatics and Relevance Theory, is presented as a prompt to the language model, guiding it through a step-by-step reasoning process to derive a final interpretation. Experimental results showed that, compared to the baseline, which prompts intermediate reasoning without presenting pragmatic theories (0-shot Chain-of-Thought), our methods enabled language models to achieve up to 9.6\% higher scores on pragmatic reasoning tasks. Furthermore, we show that even without explaining the details of pragmatic theories, merely mentioning their names in the prompt leads to a certain performance improvement (around 1-3%) in larger models compared to the baseline.

Figures

Figures reproduced from arXiv: 2510.26253 by the authors.

Figure 1
Figure 1. Accuracies on pragmatic inference task of PRAGMEGA. In most models, the proposed methods [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Accuracy of the model for each pragmatic phenomenon included in PRAGMEGA ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. The number of instances for each error pattern [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results of the additional experiments The specific prompts used for each method are presented in Tables 4, 5, and 6 in Appendix D. The models and detailed experimental settings are identical to those described in the previous section. 5.2 Results and Analysis The resul…
Figure 5
Figure 5. Figure 5: Correlation analysis between input length and accuracy. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Correlation analysis between output length and accuracy. [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

72 extracted references · 5 linked inside Pith

  1. [1]

    Maxim of Quantity: Provide an amount of information that is nei- ther too little nor too much

  2. [2]

    Maxim of Quality: Do not say what you believe to be false or for which you lack sufficient evidence

  3. [3]

    Maxim of Relation: Be relevant in your utterance

  4. [4]

    B Presumption of Relevance Presumption of Relevance(Sperbel and Wilson, 1995)

    Maxim of Manner: Avoid ambiguity and be clear, concise, and orderly in your expression. B Presumption of Relevance Presumption of Relevance(Sperbel and Wilson, 1995)

    1995

  5. [5]

    Dan Sperbel and Deirdre Wilson

    An empirical evaluation of prompting strate- gies for large language models in zero-shot clinical natural language processing: Algorithm development and validation study.JMIR Med Inform, 12:e55318. Dan Sperbel and Deirdre Wilson. 1995.Relevance: Communication and Cognition (2nd Edition). Black- well. Settaluri Sravanthi, Meet Doshi, Pavan Tankala, Rudra M...

    1995

  6. [6]

    In Findings of the Association for Computational Lin- guistics ACL 2024, pages 12075–12097

    PUB: A pragmatics understanding benchmark for assessing LLMs’ pragmatics capabilities. In Findings of the Association for Computational Lin- guistics ACL 2024, pages 12075–12097. Association for Computational Linguistics. Junya Takayama, Tomoyuki Kajiwara, and Yuki Arase

    2024

  7. [8]

    Zilong Zheng, Shuwen Qiu, Lifeng Fan, Yixin Zhu, and Song-Chun Zhu

    Do large language models understand conver- sational implicature – a case study with a chinese sitcom.Preprint, arXiv:2404.19509. Zilong Zheng, Shuwen Qiu, Lifeng Fan, Yixin Zhu, and Song-Chun Zhu. 2021. GRICE: A grammar-based dataset for recovering implicature and conversational rEasoning. InFindings of the Association for Com- putational Linguistics: AC...

    arXiv 2021

  8. [13]

    An ostensive stimulus is relevant enough to make it worth the ad- dressee’s effort to process it

Show all 72 references
  1. [14]

    What have you been doing up here?

    An ostensive stimulus is the most relevant one, given the communica- tor’s abilities and priorities. C Examples of PRAGMEGA dataset Table 2 shows examples of problems and choices for each phenomenon in the PRAGMEGA dataset. D Each Prompting method Table 3 shows each prompt use...

    2023

  2. [15]

    He does not want to get into trouble for not studying

  3. [16]

    He has been reading for some time

  4. [17]

    He does not want to offend his mom by not reading the books that she gave him

  5. [18]

    The in- vestors are squirrels collecting nuts

    He wants his mom to believe that he has been watching TV . Metaphor Andrew and Bob were discussing the investment company where An- drew works. Bob said: “The in- vestors are squirrels collecting nuts.” What does Bob mean?

  6. [19]

    They buy stocks hoping for future profit

  7. [20]

    Squirrels were hired to work in the company

  8. [21]

    The investors dress and eat well

  9. [22]

    Bob is allergic to nuts

  10. [23]

    Are you going out?

    The investors enjoy picking nuts as squirrels do. Indirect Speech Nate is about to leave the house. His wife points at a full bag of garbage and asks: “Are you going out?” What might she be trying to convey?

  11. [24]

    She wants Nate to take the garbage out

  12. [25]

    She wants to know Nate’s plans

  13. [26]

    She wants Nate to bring his friends over

  14. [27]

    Oh, it is so pleasant here

    She wants Nate to spend more time with the family. Irony It is a holiday. Stefan and Kim are sitting in the backseat of the car. They are fighting all the time. Their father says: “Oh, it is so pleasant here.” What did the father want to convey?

  15. [28]

    He does not want to listen to his kids’ argu- ments

  16. [29]

    He enjoys listening to his kids fighting

  17. [30]

    AC gives them some needed cool

  18. [31]

    Who was that man that I saw you with last night?

    He remembers about his wife’s birthday. Maxims Leslie and Jane are chatting at a coffee shop. Leslie asks, “Who was that man that I saw you with last night?” Jane responds, “The latte is unbelievable here.” Why has Jane responded like this?

  19. [32]

    She does not want to discuss the topic that Leslie has raised

  20. [33]

    She thinks that it is the best latte in the town

  21. [34]

    The man who Leslie saw makes unbelievable lattes

  22. [35]

    implicatures

    A coffee break is not a good time to discuss men. Table 3: Prompts used in the methods compared in this study. Method Prompt Simple (Baseline-1) Write ONLY the option number of your final answer and its contents in the format like: [Answer] 2) hogehoge is hogehoge. Any additio...

  23. [36]

    - The listener must infer the speaker’s intended meaning behind the utterance using this linguistic clue as a basis

    **Starting Point of Utterance Interpretation**: - The linguistic meaning (logical form) of an utterance is merely a “clue” to the interpretation intended by the speaker. - The listener must infer the speaker’s intended meaning behind the utterance using this linguistic clue as a basis

  24. [37]

    optimal relevance

    **Interaction Between Explicit Meaning and Implicature**: - Explicit meaning (the overt content of the utterance) and implicature (implied content) influence each other during processing. - This interaction forms the overall interpretation of the utterance. In Relevance Theory...

  25. [38]

    Short Relevance Prompting (Short-2) Let’s think in line with the Relevance theory

    hogehoge is hogehoge. Short Relevance Prompting (Short-2) Let’s think in line with the Relevance theory. Write down your thinking process in line with Relevance theory and ultimately decide on the final answer. Your final answer should be in the format like: [Answer]

  26. [39]

    Table 4: Prompts used in the additional experiments to examine the first concern

    hogehoge is hogehoge. Table 4: Prompts used in the additional experiments to examine the first concern. Method Prompt xbarLet’s think in line with the X-bar theory. Write down your thinking process in line with X-bar theory and ultimately decide on the final answer. Your final...

  27. [40]

    complexityLet’s think in line with the computational complexity theory

    hogehoge is hogehoge. complexityLet’s think in line with the computational complexity theory. Write down your thinking process in line with computational complexity theory and ultimately decide on the final answer. Your final answer should be in the format like: [Answer]

  28. [41]

    graphLet’s think in line with the graph theory

    hogehoge is hogehoge. graphLet’s think in line with the graph theory. Write down your thinking process in line with the graph theory and ultimately decide on the final answer. Your final answer should be in the format like: [Answer]

  29. [42]

    correct”in an objective sense but rather“dominant

    hogehoge is hogehoge. E Hyperparameters We set temperature=0.8, max_new_tokens=1500, repetition_penalty=1.2, do_sample=True. F Computation Details The parameter counts for each open model we used are as follows: • Llama-3.1-8B-Instruct: 8B • Qwen2.5-7B-instruct: 7B • Qwen2.5-1...

    2023

  30. [43]

    non-meaningful

    Majoritarian Meaning The interpretation of an utterance is determined by the intuition of the majority within a speech community. Minority readings may exist but are pragmatically considered “non-meaningful.”

  31. [44]

    consensus index,

    Consensus Index Every expression carries a“consensus index,”a hypothetical measure of how many people favor a given interpretation at a certain moment. For example, if 90% interpret“yabai”as “cool,”that becomes the pragmatically valid meaning

  32. [45]

    This accounts for semantic shifts, slang evolution, and the inversion of meaning (e.g.,“yabai”changing from negative to positive)

    Floating Meaning Since consensus shifts over time, meanings are inherently unstable. This accounts for semantic shifts, slang evolution, and the inversion of meaning (e.g.,“yabai”changing from negative to positive). Write down your thinking process in line with Majoritarian Co...

  33. [46]

    majoritarian short Let’s think in line with Majoritarian Concensus Pragmatics

    hogehoge is hogehoge. majoritarian short Let’s think in line with Majoritarian Concensus Pragmatics. Write down your thinking process in line with Majoritarian Concensus Pragmatics and ultimately decide on the final answer. Your final answer should be in the format like: [Answer]

  34. [47]

    spatial parameter

    hogehoge is hogehoge. distanceLet’s think in line with Distance-Primacy Pragmatic Theory. The Distance-Primacy Pragmatic Theory (DPPT) posits that the primary determinant of utterance interpretation is the“spatial parameter”between speaker and hearer, while linguistic content ...

  35. [48]

    Principle of Proximal Compulsion (PPC) Any utterance produced within a radius of 50 cm is automatically interpreted as a command, regardless of the speaker’s actual intent

  36. [49]

    3.Phenomenon of Distant Monologization (PDM) Beyond 3 meters, all utterances are treated as pragmatic monologues, releasing the hearer from any obligation to respond

    Mid-Distance Propositional Zone (MDPZ) Within the range of 1–3 meters, utterances are processed as proposals or invitations. 3.Phenomenon of Distant Monologization (PDM) Beyond 3 meters, all utterances are treated as pragmatic monologues, releasing the hearer from any obligati...

  37. [50]

    distance short Let’s think in line with Distance-Primacy Pragmatic Theory

    hogehoge is hogehoge. distance short Let’s think in line with Distance-Primacy Pragmatic Theory. Write down your thinking process in line with Distance-Primacy Pragmatic Theory and ultimately decide on the final answer. Your final answer should be in the format like: [Answer]

  38. [51]

    Table 6: Prompts used in the additional experiments to examine the third concern

    hogehoge is hogehoge. Table 6: Prompts used in the additional experiments to examine the third concern. Method Prompt planLet’s first understand the problem and devise a plan to solve the problem. Then, let’s carry out the plan and solve the problem step by step. Your final an...

  39. [52]

    John is an innocent person

    hogehoge is hogehoge. Table 7: Main experimental results. The highest Accuracy among the four methods is indicated inbold. Model Baseline Proposed Short simple cot grice relevance grice relevance gpt-4o 0.842 0.877 0.9400.935 0.902 0.892 gpt-4o-mini 0.696 0.694 0.7710.779 0.72...

  40. [53]

    Samantha’s dad is impressed with John’s innocence

  41. [54]

    Samantha’s dad thinks that Samantha has an incorrect view of her fiance

  42. [55]

    Samantha’s dad thinks that Samantha’s fiance is a saint

  43. [56]

    What will we have for breakfast?

    Samantha’s dad thinks that John is too religious. 2⃝ Lenny comes to the kitchen and asks his wife, Marcie: “What will we have for breakfast?” Marcie responds: “A hard-boiled egg cooked in hot water and toast that is toasted evenly on both sides.” Why has Marcie responded in su...

  44. [57]

    Marcie is really good at cooking eggs and making toast

  45. [58]

    Marcie thinks that breakfast is the main meal of the day

  46. [59]

    Marcie wants Lenny to know how his breakfast was made

  47. [60]

    This one is really sharp

    Marcie thinks that her husband’s expectations about breakfast are too high. 3⃝ John is a teacher at an elementary school. When talking with the principal about a new student, who did poorly on her entrance examination, John said, “This one is really sharp.” What did John want ...

  48. [61]

    The entrance exam is unfair

  49. [62]

    The pencils need to be sharpened

  50. [63]

    The student is smart

  51. [64]

    Everything is fine, dear. You did a great job of cleaning the apartment

    The student is not very clever. 4⃝ One day Jane comes home and is delighted to find her partner Anthony straightening up her apartment. Jane notices that Anthony threw out lots of things which were creating clutter, includ- ing an old photo that she had always kept on the coff...

  52. [65]

    She is happy that Anthony has cleaned the apartment and does not care about the picture that got thrown away

  53. [66]

    She wants to show that she is angry that Anthony has cleaned the apartment

  54. [67]

    She wants to show that she appreciates that Anthony has cleaned the apartment

  55. [68]

    I am not sure that now is the best time for painting

    She shows him how angry she is with him for throwing out things without her consent. 5⃝ Cindy wanted to paint a picture. She got her paints, paper and brushes ready. She has a meeting to go to in 10 minutes. Her dad said to her, “I am not sure that now is the best time for pai...

  56. [69]

    He does not want Cindy to start painting

  57. [70]

    He wants Cindy to create a sculpture

  58. [71]

    He wants Cindy to paint a picture for the meeting

  59. [72]

    He has some doubts whether Cindy should be painting

  60. [2017]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa

    Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13):3521–3526. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. InPro- ceedings of ...

    2022 arXiv

  61. [2021]

    InFindings of the Associ- ation for Computational Linguistics: EMNLP 2021, pages 1980–1989

    DIRECT: Direct and indirect responses in conversational text corpus. InFindings of the Associ- ation for Computational Linguistics: EMNLP 2021, pages 1980–1989. Association for Computational Linguistics. Shohei Tanaka, Konosuke Yamasaki, Akishige Yuguchi, Seiya Kawano, Satoshi...

    2021 arXiv

  62. [2022]

    Nicole Baumgarten

    Do as i can, not as i say: Grounding language in robotic affordances.Preprint, arXiv:2204.01691. Nicole Baumgarten. 2022.Contrastive Pragmatics, pages 172–189. Routledge. Penelope Brown and Stephen C. Levinson. 1987.Po- liteness : some universals in language usage. Num- ber 4 ...

    2022 arXiv

  63. [2023]

    is the pope catholic?

    "is the pope catholic?" applying chain-of- thought reasoning to understanding conversational implicatures.Preprint, arXiv:2305.13826. James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, ...

    arXiv

  64. [2024]

    InProceedings of the 23rd Chinese National Conference on Computational Linguistics (V olume 1: Main Conference), pages 1270–1285, Taiyuan, China

    Do large language models understand con- versational implicature- a case study with a Chinese sitcom. InProceedings of the 23rd Chinese National Conference on Computational Linguistics (V olume 1: Main Conference), pages 1270–1285, Taiyuan, China. Chinese Information Processin...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.