PaperlayerView exact revision on arXiv

Source-linked AI summary

Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literally

Zhizhuo Kou, Jian Yang, Pengfei Hao, Junyu Luo, Jingshu Peng, Sirui Han, Yike Guo

arXiv:2610.21204v1cs.AIcs.CL

TL;DR

The paper asks what extra thinking does to models’ everyday inferences, then measures chosen readings while varying thinking at fixed weights and probing reasoning traces. Thinking often cancels logical-word implicatures even as some context-driven inferences improve, and early trace answers can recover implicatures without reducing literal accuracy.

  • Problem

    It was unknown whether extra thinking changes the reading a model chooses when both readings are allowed, despite evaluations of intended-meaning recovery.

    Source: §1 · p. 1

  • Method

    The study holds model weights fixed while varying thinking and probes successive trace prefixes to track changes in chosen readings.

    Source: §1 · p. 2

  • Results

    Thinking reduces some and or implicatures in most tested open models and GPT-5, while presuppositions and indirect answers improve; traces show enriched readings early and literal readings at completion.

    Source: §1 · p. 2, §7 · p. 8, §6 · p. 8

  • Takeaways & Limitations

    Because single-answer accuracy cannot distinguish interpretation changes from competence loss, evaluations should report chosen readings alongside accuracy.

    Source: §6 · p. 8

    Headline results rely on forced-choice framings because free-form LLM ratings agreed with human consensus at best κ = 0.57.

    Source: Limitations · p. 9

Abstract

from arXiv · show

When a speaker says ``some students passed'', listeners infer that not all of them did. Reasoning models now let users decide how long a model thinks before it answers, but we do not know what extra thinking does to such everyday inferences. We study this directly: we keep a model's weights fixed, change only how much it thinks, and record which reading it chooses. A second tool cuts the model's own reasoning at successive points and asks for an answer, showing where the reading changes. Thinking makes models read logical words as logic: it cancels the implicatures of \emph{some} (\emph{not all}) and \emph{or} (\emph{not both}) in most tested open reasoning models and in GPT-5, while context-driven inferences such as indirect answers improve. The trace shows why: models first entertain the enriched reading and cancel it later. Answering from that early reading, an exploratory inference-time remedy, recovers these implicatures without losing literal accuracy. Evaluations of reasoning models should measure how they interpret, not only how accurate they are.

1 Introduction

The introduction asks whether user-controlled thinking changes the readings models choose, a question prior evaluations had not measured directly. It proposes fixed-weight dose comparisons and trace-prefix probes, and reports that thinking can cancel logical-word implicatures while context-driven inferences improve.

1 Introduction

  • The introduction frames some and or as logical words whose pragmatic inferences are cancelled when models favor their literal readings.Figure 1 contrasts some and or as ∃x and ∨ with adequate, which retains “not good.”Source: Figure 1 · p. 1, §1 · p. 1
  • Earlier evaluations scored intended-meaning recovery, but no study had measured which reading a model chooses when both readings are allowed.Prior work reported reduced scalar-implicature recognition in Qwen3 thinking mode (Spinoso-Di Piano et al., 2026).Source: §1 · p. 1
  • The study holds model weights fixed while varying thinking and cuts reasoning traces at successive points to locate reading changes.It applies these tools to 40 open and seven closed reasoning models across logical words, lexical scales, presuppositions, and indirect answers.Source: §1 · pp. 1–2
  • Thinking shifts models toward literal readings of some and or, while context-dependent inferences improve; trace prefixes show enriched readings appearing early before later cancellation.The reported pattern spans most tested open models and GPT-5.Source: §1 · p. 2
  • The contributions include a fixed-weight diagnostic protocol, an early-trace remedy that preserves literal performance, and released test materials and traces.The remedy improves scalar-felicity accuracy across two 8B models; released resources include expert-checked scenarios, expanded test sets, trace logs, and translated slices.Source: §1 · p. 2

2 Background and Related Work

The background contrasts accounts predicting more enrichment with thinking against an account in which default implicatures can be cancelled. Prior language-model studies examined accuracy at fixed settings, whereas this paper measures chosen readings across thinking levels and traces.

2 Background and Related Work

  • Generalized implicatures include some → ‘not all’ and or → ‘not both’, while particularized implicatures depend on specific context; lexical-scale enrichment varies across scales.Generalized implicatures can be blocked by context, unlike context-specific particularized inferences.Source: §2 · p. 2
  • Figure 2 summarizes that thinking makes some more literal in most open reasoning models, with colors distinguishing literal shifts from more pragmatic shifts.The caption identifies vermilion as a counted literal shift and blue as more pragmatic.Source: Figure 2 · p. 2
  • The effort account predicts more thinking should increase enrichment, whereas the default-and-cancel account predicts cancellation when literal truth conditions speak against a default implicature.The paper uses “default” for the no-thinking reading and “cancel” for a shift toward the literal reading, without committing to a cognitive mechanism.Source: §2 · p. 2
  • A developmental view predicts that thinking may amplify a model’s existing enrichment preference, with direction changing by capability.The account draws on the finding that children accept underinformative some more often than adults.Source: §2 · p. 2
  • Prior work found chain-of-thought can improve implicature resolution (Kim et al., 2023), encourage literal interpretations (Park et al., 2024), or reduce Qwen3 scalar-implicature recognition (Spinoso-Di Piano et al., 2026).Those studies measured accuracy at fixed settings; this paper varies thinking on fixed weights and examines readings, traces, and training stages.Source: §2 · p. 2

3 Method

The method varies thinking while holding weights fixed, records chosen readings, and probes successive prefixes of each model’s reasoning trace. Items span logical words and other inference types, with multiple framings, model conditions, and statistical checks.

3 Method

  • The study varies thinking conditions with the same weights and uses a trace-prefix diagnostic to locate where readings change.The figure schematic includes thinking off, capped, or natural conditions and readouts after successive trace prefixes.Source: §3 · p. 2, Figure 3 · p. 3
  • Items pair a trigger sentence with a question about speaker meaning; “yes” denotes the pragmatic reading and “no” the literal reading in the running example.The study draws logical-word and presupposition items from IMPPRES, lexical scales from scalar-diversity sets, and particularized implicatures from several datasets.Source: §3 · p. 3
  • Primary framings ask what the speaker meant or whether a sentence is an adequate description, while true/false questions provide a logic-mode comparison.Options are shuffled each trial; headline numbers use held-out items, with full-set replications marked separately.Source: §3 · p. 3
  • Conditions include thinking off, capped token budgets, or natural stopping up to an 8,192-token safety cap; gpt-oss uses low and high reasoning effort.Qwen3’s off condition uses its own switch; its thinking template differs by an empty thinking block.Source: §3 · p. 3
  • The primary readout is the share of committed answers choosing the pragmatic reading, while a second readout compares answer options after the trace.For particularized implicatures and presuppositions, the reading measure is accuracy; category sentences also use d′ and criterion c from hits and false alarms.Source: §3 · p. 3
  • Trace prefixes are cut at successive token lengths, and forced answers reveal whether the model endorses a pragmatic reading early before ending literally.The prefix curve describes average endorsed readings and is an intervention on the trace, not a readout of a hidden interpretation.Source: §3 · pp. 3–4
  • The panel includes Qwen3, OLMo-3-Think, gpt-oss, ten further open models, seven closed models, and checkpoints along base-to-RL lineages.Conditions use three decoding seeds and 95% item-cluster-bootstrap intervals; preregistered tests include logistic models and equivalence tests.Source: §3 · p. 4

4 Thinking Makes Some and Or More Literal

Thinking reduces pragmatic readings of some and or across many reasoning models, while context-driven inferences improve. Trace probes show enriched readings often appear early and are later cancelled, and post-training shifts the default reading at recipe-dependent stages.

4.1 Thinking Cancels the Implicatures of Some and Or

  • Across 40 open models, some/or shifts survive the pre-declared format and polarity checks in 30; GPT-5 also shows lower pragmatic readings with high effort.The reported changes for GPT-5 are −20.6 pp for some and −18.8 pp for or against minimal effort.Source: §4.1 · p. 4, Table 1 · p. 5
  • Thinking’s effects differ across inference types: it reduces some/or enrichment, but improves context-driven inference and, in larger Qwen3 models, lexical-scale enrichment.Fig-QA accuracy rises 10.7 pp and accuracy on six PUB tasks rises 10.1 pp; both gains are significant at every Qwen3 size from 4B to 32B.Source: §4.1 · p. 5

4.2 Sensitivity Analysis: The Shift Survives Stopping, Baseline, Format and Item Choices

  • Full thinking lowers pragmatic readings under every tested analysis choice, while a 128-token budget raises some readings before longer thinking lowers them.For Qwen3-8B, some rises by 37.2 pp at 128 tokens and falls 21.8 pp below its no-thinking level at 1024 tokens.Source: §4.2 · p. 5, Table 2 · p. 6

4.3 Early in the Trace the Enriched Reading Is Endorsed; at the End It Is Rejected

  • Trace-prefix answers often move from pragmatic early readings to literal final readings, with the pattern appearing across models and logical words.In Qwen3-8B, some rises from 35.8% with an empty trace to 52.0% after 128 tokens, then falls to 1.1% by the trace end.Source: Figure 4 · p. 6, §4.3 · pp. 6–7

4.4 Post-Training Moves the Default Reading

  • Post-training shifts the default reading at recipe-dependent stages, and the additional thinking effect can shrink as the default becomes literal.OLMo-3 changes after thinking SFT, while Tülu-3 changes with DPO; OLMo-3-Think’s some cost shrinks from −32.1 pp to −2.6 pp after the final RL step.Source: §4.4 · p. 7, Table 3 · p. 7

5 Selective Thinking for Scalar Items (Exploratory)

The exploratory remedy answers selected some/or questions from an early trace prefix, where the enriched reading is already present. On held-out items, it improves the pooled score while leaving literal-gold accuracy unchanged.

5 Selective Thinking for Scalar Items (Exploratory)

  • For questions about a speaker’s belief in a some/or sentence, the remedy uses the answer after the first 128 trace tokens; other items retain natural-thinking answers.Routing reads only the question template and speaker sentence, and routed answers come entirely from the prefix distribution.Source: §5 · p. 7
  • The remedy is exploratory, with its weight selected on development items; four pre-registered interventions in or training the thinking did not pass their development gates.The weight was frozen and evaluated once on held-out template families.Source: §5 · pp. 7–8
  • The equal-family score rises 1.4 pp over natural thinking, while every literal-gold safety stratum stays within its pre-declared bound.Literal accuracy is unchanged because routing leaves literal-gold strata untouched; the smallest one-sided lower bound is 0.0 pp.Source: Table 4 · p. 8, §5 · p. 8
  • On IMPPRES some/or felicity items, accuracy rises from 1.2% to 20.5% in Qwen3-8B and from 2.2% to 8.1% in DeepSeek-R1-0528-Qwen3-8B.A task-aware instruction raises the overall score more, by 2.3 pp, but barely moves these items: 3.5% and 2.6%; the approaches act on different items.Source: §5 · p. 8

6 Discussion

More thinking does not uniformly increase pragmatic enrichment: it reduces implicatures for logical words, while some context-dependent inferences improve. The pattern varies by inference type and model, and accuracy alone cannot reveal which reading a model chose.

6 Discussion

  • Thinking reduces scalar enrichment for some and or, while presuppositions and indirect answers improve; lexical-scale results rise in larger Qwen3 models and are mixed elsewhere.The pattern fits default-and-cancel better than a uniform increase in enrichment: traces show enriched readings early, followed by literal answers.Source: §6 · p. 8
  • A model with thinking on more often treats “some of the tests failed” as allowing that all failed, yet answers indirect questions more accurately.Because both readings are licensed, single-answer accuracy benchmarks cannot distinguish interpretation change from loss of competence.Source: §6 · p. 8
  • The pattern is not universal across formats: Granite-3.3 and Llama-3.1-Nemotron-Nano-8B read some more pragmatically in Yes/No format, while the smallest Qwen3 models and Phi-4-mini-reasoning also lose lexical scales.Together with the stage results, these exceptions show that longer deliberation need not produce a literal shift.Source: §6 · p. 8

7 Conclusion

The conclusion is that controlled deliberation changes which readings reasoning models endorse, and accuracy alone cannot distinguish these interpretation shifts. The paper’s trace diagnostic locates changes within reasoning and suggests an exploratory remedy.

7 Conclusion

  • On fixed weights, thinking reduces the implicatures of some and or in most open models and GPT-5, while presuppositions and indirect answers improve.Larger Qwen3 models read lexical scales more pragmatically.Source: §7 · p. 8
  • Early trace prefixes give the enriched reading, while completed traces give the literal one; checkpoint comparisons locate the drop in specific post-training stages.The training-stage comparisons are described as locating the drop, not identifying a mechanism.Source: §7 · p. 8
  • Answering from the reading entertained early recovers scalar implicatures on held-out items without changing literal accuracy.The dose protocol and trace diagnostic measure interpretation changes that accuracy alone cannot distinguish.Source: §7 · p. 8

Limitations

The evidence has scope and measurement boundaries: graded dose-response results rely mainly on Qwen3 and gpt-oss, headline items are English, and outcomes use forced choices. The trace probe and checkpoint comparisons do not identify a mechanism.

Limitations

  • The graded dose-response rests mainly on Qwen3 and gpt-oss; other open models use a forced-empty thinking block they were not trained on, and closed models are accessed only through APIs.Training-stage checkpoint comparisons are correlational.Source: Limitations · p. 8
  • All headline items are English; a machine-translated German and Chinese slice shows the same drop.The non-English evidence is limited to that translated slice.Source: Limitations · p. 9
  • Headline results use forced choices because free-form answers could not be coded reliably by LLM raters; the best LLM-rater agreement with human consensus was κ = 0.57.The remedy was tested on two 8B models sharing a base model, so generality across lineages and scales is untested; human labels were consensus labels without independent double annotation.Source: Limitations · p. 9

Ethical considerations

The paper documents its data and research practices and warns that literal interpretations can miss user intent when reasoning is enabled. It recommends evaluating interpretation alongside accuracy for conversational use.

Ethical considerations

  • The study uses publicly released models and datasets under their licences, and releases new items, code, prompts, and traces under open licences.IMPPRES is non-commercial; two stimulus repositories lack licence files, and NOPE is not redistributed.Source: Ethical considerations · p. 9
  • No participants were recruited, and no personal data were collected or used.The human labels are discussion-based consensus judgments from a panel of five or more trained annotators.Source: Ethical considerations · p. 9
  • Thinking may make assistants answer underinformative or indirect utterances literally, missing what users mean despite technical correctness.The authors recommend that deployers who enable thinking by default for conversational use evaluate interpretation, not only accuracy.Source: Ethical considerations · p. 9
  • AI assistants helped with code writing and polish; researchers designed and checked the research.Source: Ethical considerations · p. 9

A Data and Resources

The appendix inventories the models, checkpoints, artefacts, and licences used. Its experimental setup describes the model comparisons, item sources, readouts, reasoning manipulations, and statistical procedures.

A.2 Artefacts and Licences

  • The appendix lists model checkpoints and artefact licences; openPangu use excludes the European Union, and NOPE is used for research evaluation but not redistributed.New items, code, and traces are released under an open licence.Source: §A.2 · p. 13, §A.1 · p. 13

B Experimental Setup

A.1 Models and Checkpoints

  • The primary fixed-weight manipulation uses Qwen3 models whose thinking can be switched without changing the template beyond an empty thinking block.Because Qwen3 model size and training recipe are confounded, the study also compares same-recipe size pairs in OLMo-3 and gpt-oss.Source: §B.1 · pp. 13–14

B.1 Method Details

  • Items cover generalized implicatures, particularized implicatures, and presuppositions; headline generalized-implicature results use fresh items, with released sets treated as replications.Items are split into development and test halves, and findings are reported on the test half.Source: §B.1 · p. 14
  • Scalar items use felicity framings as primary and truth judgments as a logic-mode comparison; particularized implicatures use meaning questions, with some also tested in true/false form.Open answers are reported descriptively because LLM coding did not reach the predeclared agreement with a human audit.Source: §B.1 · p. 14
  • Thinking is varied through natural or capped budgets and length-stratified samples; dose–response claims rely on budget contrasts because switches and effort levels can change prompts.Intent-Seeded Thinking places a fixed sentence about speaker intent at the start or near the end of the thinking block, with token-matched controls.Source: §B.1 · p. 15
  • Each condition uses three decoding seeds, with confidence intervals bootstrapped over items; primary tests use item-clustered logistic regressions and Holm-corrected p-values.Findings use held-out items; prompts and parsing were fixed on the other half, and hypotheses and thresholds were set before the main runs.Source: §B.1 · p. 15

B.2 Experimental Details

This section documents the experimental materials and supporting analyses, including item and answer validation, prior-work baselines, prompt examples, per-model results, judging limitations, and robustness checks.

B.2 Experimental Details

  • Open-answer results are descriptive because gpt-oss-20b agreement with human consensus was κ = 0.33 [0.12, 0.54], below the ≥0.70 acceptance rule.Accordingly, every headline result uses forced-choice formats.Source: §B.2 · p. 16
  • Human consensus matched the intended reading in all 50 audited speaker-knowledge scenarios, meeting the pre-declared validity bar.The Wilson 95% lower bound was 86.2%; the audit judged “some” misleading when the speaker checked every object and fine when they had not.Source: §B.2 · p. 16

B.3 Prior-Work Baselines

  • On ImplicatureX, thinking reduced Qwen3-8B scalar-implicature recognition from 100.0% to 65.2% and raised cancellation from 46.4% to 94.2%.This reproduces the original report; LUDWIG accuracy rose with thinking at every number of shots.Source: §B.3 · p. 16, Table 6 · p. 16
  • GPT-5 became more literal with effort in both answer formats, while Claude models’ pragmatic rates varied by format and polarity probes exposed possible Yes-response shifts.The passage also reports model-specific exceptions: Claude Opus 4.8 showed a small pragmatic shift under both polarities, and Granite-3.3 and Llama-3.1-Nemotron-Nano-8B showed evidence of genuine shifts; Gemini 2.5 Flash results were uninterpretable because of answer failures.Source: §B.3 · pp. 16–17, Table 7 · p. 18, Table 8 · p. 18

B.4 Verbatim Prompts

  • The prompts include Yes/No, polarity-free three-way, and dataset-option formats, plus budget-stop and forced-close variants for eliciting answers from thinking traces.The forced-close variant uses the first N tokens of the same trace before the native close tag.Source: §B.4 · pp. 17–19
  • The verbatim prompts ask whether a speaker’s statement supports a second statement, request a Yes-or-No judgment, and require only the answer letter.Examples include “some” paired with “not all” and “all” paired with “some.”Source: §B.4 · p. 19
  • A shown response concludes that if all teenagers criticize Guy, then some do, so the speaker does not believe the “some” statement is false.It answers No to the question of whether the speaker believes that statement is false.Source: §B.4 · pp. 19–20

C Additional Results

C.1 Additional Results

Additional analyses show that literal shifts generalize across formats, languages, and model families, while trace probes and seed experiments identify where readings change and how some enrichment can be restored.

C.1 Additional Results

  • In three-way choices, 21 of 24 models shift toward literal readings, while several exceptions depend on answer format or begin near a floor.Granite-3.3 and Llama-3.1-Nemotron-Nano-8B shift pragmatic only in Yes/No format; Magistral-Small and openPangu shifts track answering Yes.Source: §C.1 · p. 20
  • Trace probes show early-prefix effects vary: sampled lexical-scale readings jump in some models, while Skywork-OR1’s or reading drops immediately.For Qwen3.5 and Phi-4-mini-reasoning, truncation often produces invalid answers, so those curves describe surviving traces only.Source: §C.1 · pp. 20–21, Figure 6 · p. 22, Figure 7 · p. 22
  • Thinking lowers pragmatic readings in German and Chinese as in English, although the Chinese translations were not human-verified.For Qwen3-8B, changes are −37.9 pp in German, −17.1 pp in Chinese, and −39.2 pp on English source items.Source: §C.1 · p. 21
  • Qwen3.8-27B’s pooled some shift conceals opposing relation-specific changes, including −59.7 pp for both-negated pairs and +56.2 pp for a mixed-negation relation.The section also points to results for further context-driven sets and per-family analyses, without giving their values here.Source: §C.1 · pp. 21–22, Table 10 · p. 21
  • Intent-Seeded Thinking restores more enrichment than length-matched controls, but its gains differ by model and indirect-answer accuracy can fall.In Qwen3.5-9B, the seed raises pragmatic readings of some to 44.9%, versus 1.3% under natural thinking; accuracy on indirect answers moves from 83.5% to 78.2% in Qwen3-8B.Source: §C.1 · pp. 22–23
  • 30 of 40 open reasoning models show a literal shift on some and or, and the full-set results cover 300 held-out items per trigger.Table 12 summarizes the held-out split; Table 9 describes the full held-out sets and formats.Source: Table 9 · p. 21, Table 12 · p. 24

D Pre-registration and Analysis Details

  • Four pre-registered remedies failed their development gates, and the analysis section reports outcomes for the pre-registered predictions.The supplementary results also include run settings and per-phenomenon tables.Source: Table 13 · p. 25, Table 14 · p. 25

D.1 Accounts, Hypotheses and Operationalisation

The section contrasts effort, default-and-cancel, and capability accounts, then specifies how the study measures deliberation and tests predictions. Its reported outcomes include mixed training-stage and load patterns, alongside remedies that did not pass their gates.

D.1 Accounts, Hypotheses and Operationalisation

  • Training-stage outcomes differ: the literal shift appears at DPO in Tülu-3 but at thinking SFT in OLMo-3, while OLMo-3’s thinking cost shrinks over RL steps.These outcomes do not support one uniform stage boundary across the two models.Source: §D.1 · p. 26
  • With 2,048 words of unrelated preceding text and thinking off, logical-word pragmatic readings move in opposite directions for Qwen3-8B and Qwen3-14B.Rates change from 22.6% to 14.2% for Qwen3-8B and from 14.5% to 20.9% for Qwen3-14B.Source: §D.1 · p. 26
  • The default-and-cancel account predicts that deliberation cancels generalized implicatures while context-driven particularized implicatures may hold or improve; the effort account predicts more enrichment.The capability account instead predicts that the direction of generalized-implicature change depends on the model’s existing competence.Source: §D.1 · p. 27
  • The study defines R as the share of committed answers with a pragmatic reading and varies thinking budget from none to natural stopping under an 8,192-token cap.Felicity judgments are primary; true/false framing serves as a logic-mode comparison, and category items also use ternary and 1–7 scales.Source: §D.1 · pp. 27, 29
  • The operational predictions include a trace-prefix test for a shift toward literal readings and an intent-seed test requiring restored enrichment without reduced particularized-implicature accuracy.The section states that hypotheses, thresholds, falsifiers, and the decision rule were fixed before the main runs.Source: §D.1 · pp. 29–30

D.2 Pre-registered Predictions and Outcomes

  • All four pre-registered remedies failed their pre-declared development gates.The remedies included a sentence at the start of thinking, a check before it ends, fine-tuning, and preference optimization on correct traces.Source: §D.2 · p. 30

D.3 Statistical Procedures

The statistical procedures use item-clustered estimates and corrected comparisons, while the related-work discussion distinguishes this study’s controlled test-time manipulation from prior fixed-setting or accuracy-focused work.

D.3 Statistical Procedures

  • The planned by-item mixed-effects models were replaced by logistic regressions with item-clustered robust errors because a mixed-model solver was unavailable.Budget-by-class interaction estimates are −0.067, −0.063, −0.049, and −0.037 for 4B, 8B, 14B, and 32B, respectively.Source: §D.3 · p. 30
  • 47 of 48 some and or shifts survive Holm correction on full held-out sets, compared with fewer significant shifts on smaller core sets.Secondary analyses are uncorrected and labeled as such.Source: §D.3 · p. 30
  • Core results use 26–33 items per word, producing wide item-level intervals, so headline numbers are replicated on 300-item held-out sets.Confidence intervals use paired, stratified cluster bootstraps over items with 10,000 resamples.Source: §D.3 · pp. 30–31

E.1 Related Work

  • Prior work reports mixed reasoning effects on social understanding and implicature, whereas this study controls thinking budget on fixed weights and traces interpretation changes.The authors distinguish their outcome—choice between two licensed readings—from accuracy on tasks with one correct answer.Source: §E.1 · pp. 31–32
  • Earlier comparisons across training stages confound deliberation with post-training; this study varies inference budget while holding weights fixed for its primary contrast.Its stage analysis extends from supervised fine-tuning through preference tuning to RLVR.Source: §E.1 · p. 31

Related papers

Loading 2610.21204v1…