acceptodds
Under review as a conference paper at ICLR 2027

ThinkMid: Thinking and Correcting in the Middle of Response Generation

Abstract

Large Reasoning Models (LRMs) scale at test time by spending more computation on a problem, yet existing methods largely decide how much to think before answering, rather than when to think again while answering. We introduce , a training-free decoding framework that dynamically allocates additional reasoning during response generation. ThinkMid preserves the model’s original reasoning trace, generates the response incrementally, and monitors sentence-level uncertainty using token entropy. When a sentence is sufficiently uncertain, generation pauses and the same model performs an auxiliary self-review in the existing conversational context. If an error is identified, a further reasoning turn produces a local correction; the flagged segment is replaced, and generation resumes from the revised prefix. ThinkMid requires no external verifier or additional training and is complementary to existing test-time scaling strategies. Across mathematical, scientific, and expert-level reasoning benchmarks using two LRMs with different architectures and reasoning formats, ThinkMid improves over the single-solve baseline and compute-matched test-time scaling methods while invoking additional reasoning for only a small fraction of generated sentences. These results indicate that where additional computation is inserted during generation can be as important as how much computation is used.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Test-Time Scaling, Self-Reflection, Large Reasoning Models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QgtqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ThinkMid, a training-free decoding method that monitors mean token entropy of each response sentence. When a sentence is uncertain, the same model reviews its response in a temporary turn; if an error is confirmed, a correction replaces the sentence and decoding resumes. On AIME 2025, GPQA Diamond, and an HLE subset with two LRMs, ThinkMid improves over the single-solve baseline and is reported to outperform compute-matched sequential, parallel, and iterative scaling baselines.

Critical strengths
  1. The idea of spending extra compute inside the response, where errors arise and propagate, is well motivated and complementary to existing test-time scaling.
  2. The method is simple, training-free, and works across two architectures and thinking formats.
  3. Separating the trigger (entropy) from the decision to edit (self-review) is a sensible guard against harmful self-correction, and the per-problem compute matching is careful.
Critical weaknesses
  1. Source of gains is unclear. ThinkMid discards high-entropy segments, adds thinking budget, and regenerates them, but these effects are never disentangled. A simple baseline that rolls back and resamples flagged segments without review is missing. The random-placement ablation still uses review, and the Section 5 "plain continuation" keeps the flagged sentence. Fixed-interval placement also matches ThinkMid's accuracy, suggesting extra regeneration opportunities, not self-verification, may drive the gains.
  2. Weak statistical evidence limited by the small evaluation sets. All results are single-run without confidence intervals. Margins over the strongest baseline amount to 0–4 attempts on AIME and 1–2 questions on GPQA and HLE, well within noise.
  3. Narrow analysis and evaluation. Ablations and τ selection use only AIME with one model (30 problems), and τ appears tuned on the test set. Benchmarks are small and limited to math/science QA.
  4. Compute accounting and baselines. It is unclear whether reused context is counted for auxiliary turns. Under the matched budget, BoN has only N≈1–2, making it a weak baseline. Latency is not reported.
Questions for authors
  1. How does ThinkMid compare with entropy-triggered resampling at matched compute?
  2. Can you report multi-seed results with confidence intervals and paired significance tests?
  3. How are prompt tokens counted for review and correction turns relative to BoN candidates?
  4. What are the trigger rates, review verdict distributions, and correction acceptance rates?
5 comments

We sincerely appreciate your careful review and constructive comments. Please see our responses to your comments one by one below.

Weakness #1 and Question #1: Source of gains and comparison with entropy-triggered resampling.

Thank you very much for your valuable suggestions. We would like to clarify that, as highlighted in the Critical Strengths identified by Reviewer 4Aag, ThinkMid separates uncertainty from correctness: high entropy triggers review rather than automatically discarding or rewriting a segment. The Review stage determines whether to request correction, while the Correction stage determines whether and how to revise the segment.

Following your suggestion, we added four ablations, including entropy-triggered resampling, to examine the contributions of mid-response intervention, the verification gate, and targeted local correction. ThinkMid achieves 83.3% accuracy, compared with 73.3% for entropy-triggered resampling. Please refer to our response to Reviewer 4Aag Weakness #3 for the experimental details and complete results.

Weakness #2 and Question #2: Statistical evidence and multi-seed robustness.

We appreciate your valuable suggestions.

(1) Multi-seed evaluation and statistical uncertainty.

We would like to clarify that our AIME 2025 results aggregate runs with four distinct random seeds, rather than a single run. GPQA and HLE were evaluated with one attempt per question because their evaluation sets contain more questions, following common practice in related work.

Following your suggestion, we additionally report multi-seed results with confidence intervals and paired significance tests. Please refer to our response to Reviewer UnNa Weakness #2 for the detailed results. ThinkMid improves over Vanilla under every tested seed, supporting the consistency of its gains in this setting.

(2) Overall performance across settings.

We acknowledge that gains over the strongest baseline in individual settings can be modest. However, the strongest baseline varies across settings, and ThinkMid shows larger overall gains when compared against any fixed baseline across the six settings. We believe ThinkMid performs better overall than the compute-matched baselines: it ranks first or second in all six settings and consistently improves over Vanilla, whereas some baselines reduce accuracy in certain settings.

Weakness #3: Scope of ablations, threshold selection, and benchmark coverage.

Thank you very much for your valuable suggestions.

(1) Additional ablations and sensitivity analysis.

Following your suggestion, we added ablation experiments and entropy-threshold sensitivity analysis on HLE for Qwen3.6-35B-A3B, with results reported in Table A1. At our default threshold, entropy-guided review achieves higher accuracy than both fixed-interval and random placement, while consuming fewer tokens than fixed-interval review. This supports more effective allocation of review computation. Moreover, all three tested thresholds yield the same accuracy, demonstrating stable accuracy across the tested threshold range.

Table A1. Ablation and entropy-threshold sensitivity results on HLE with Qwen3.6-35B-A3B.

VariantAccuracy (%)Tokens
Vanilla9.015,121
Fixed interval14.029,757
Random16.023,911
Entropy-triggered, τ = 0.318.039,911
Entropy-triggered, τ = 0.518.027,401
Entropy-triggered, τ = 0.718.021,226

(2) Threshold selection.

Regarding threshold selection, we clarify that we did not tune τ separately for individual models or benchmarks. We selected a generally effective threshold during implementation and used the same value across all settings. We expect that tailoring τ to individual settings could yield further improvements.

(3) Benchmark coverage and further evaluation.

Regarding evaluation scope, our experiments cover three commonly used reasoning benchmarks and multiple subject categories. GPQA and HLE, together with repeated evaluations on AIME 2025, are widely used in related reasoning-model research. These benchmarks were deliberately selected to assess tasks requiring reasoning: our claim concerns improvements on reasoning tasks, rather than universal improvements across all task types.

Given the limited rebuttal period, we could not complete further experiments. If the paper is accepted, we commit to including ablation studies and hyperparameter sensitivity analyses across all model–benchmark settings, together with evaluations on larger benchmarks covering additional task categories, in the camera-ready version.

Weakness #4 and Question #3: Compute accounting, BoN baselines, and latency.

Thank you very much for your comments.

(1) Compute accounting.

In our token accounting, cached prefix tokens are not counted again when reused by auxiliary turns, since the prefix is reused without repeated prefilling. All newly processed tokens are counted, including the review and correction prompts and their generated thinking and responses. Cache reuse is described in Section 3.3, while Section 4.1 and Appendix C.3 describe token accounting; we will make the treatment of cached prefixes explicit in the revision.

(2) BoN under the matched budget.

For BoN, each candidate’s prompt, thinking, and response tokens count toward the same per-problem budget, as specified in Appendix C.4. Its limited number of candidates therefore reflects the modest additional token budget required by ThinkMid: under the evaluated candidate-generation settings, generating more candidates would require a larger budget. This highlights ThinkMid’s advantage in allocating limited additional computation to targeted review and correction, whereas BoN spends its budget generating additional complete solutions.

(3) Latency.

Following your suggestion, we also report latency results in Table A2.

Table A2. Latency comparison across models and benchmarks.

ModelBenchmarkVanillaThinking ForcingBoN-JudgeSelf-RefineThinkMid
Qwen3.6-35B-A3BAIME 20258451,4791,3711,3321,484
Qwen3.6-35B-A3BGPQA667834780770864
Qwen3.6-35B-A3BHLE8111,2691,1821,2501,148
GPT-OSS-20BAIME 20257561,2841,1761,2641,279
GPT-OSS-20BGPQA597961936860774
GPT-OSS-20BHLE7891,4101,3711,4041,161

Question #4: Trigger rates, review verdict distributions, and correction acceptance rates.

Thank you very much for your valuable suggestion. Following your suggestion, we report the trigger rates, review verdict distributions, and correction acceptance rates in Table A3.

Table A3. Trigger rates, review verdict distributions, and correction acceptance rates.

ModelBenchmarkSentence trigger rate (%)Review verdicts: correct / error / uncertain (%)Correction acceptance rate (%)
Qwen3.6-35B-A3BAIME 20253.172.0 / 19.5 / 8.593.8
GPT-OSS-20BAIME 202516.459.8 / 26.2 / 14.060.7
Qwen3.6-35B-A3BGPQA6.184.5 / 10.3 / 5.270.0
GPT-OSS-20BGPQA16.356.1 / 32.5 / 11.537.5
Qwen3.6-35B-A3BHLE8.170.1 / 18.9 / 11.096.2
GPT-OSS-20BHLE25.652.6 / 32.8 / 14.657.0

Thank you for the additional experiments. The rebuttal addresses several concerns but does not fully resolve the limitations. I therefore maintain my original rating.

Reviewer_UnNaOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ThinkMid, a training-free test-time scaling framework that introduces additional reasoning during response generation. Unlike conventional approaches that extend initial reasoning, sample multiple solutions, or refine completed responses, ThinkMid monitors the mean token entropy of generated response segments and selectively triggers self-review when uncertainty exceeds a predefined threshold. If an error is identified, the same model generates a local correction, replaces the corresponding segment, and resumes generation from the revised prefix.

The authors evaluate ThinkMid on AIME 2025, GPQA Diamond, and Humanity's Last Exam (HLE), using Qwen3.6-35B-A3B and GPT-OSS-20B. The reported results show improvements over vanilla inference and competitive performance against compute-matched test-time scaling baselines. Additional ablations compare entropy-guided intervention with random and periodic review placement, while a limited case analysis investigates the effectiveness of individual corrections.

Critical strengths
  1. Interesting Perspective on Test-Time Scaling. The paper explores allocating additional reasoning during response generation rather than exclusively before producing an answer. The proposed selective review and local correction mechanism provides an interesting perspective on inference-time computation allocation.

  2. Promising Empirical Improvements. ThinkMid reports accuracy improvements over vanilla inference across the evaluated benchmarks, with particularly notable gains on AIME 2025 and HLE. The comparison against compute-matched baselines provides preliminary evidence of the potential effectiveness of mid-response intervention.

Critical weaknesses
  1. Insufficient Theoretical and Empirical Justification for Entropy-Guided Intervention. The central design of ThinkMid relies on mean token entropy to determine when additional reasoning should be invoked. However, the paper provides neither a rigorous theoretical explanation nor direct empirical observations establishing why token entropy should correlate with reasoning errors or intervention effectiveness. Token-level predictive uncertainty does not necessarily reflect semantic correctness. While prior work on high-entropy reasoning tokens is cited, the connection to error-prone response segments remains insufficiently established. The reported final-answer improvements alone cannot validate this central methodological assumption. Consequently, the motivation and underlying mechanism of entropy-guided intervention remain inadequately supported.

  2. Limited Architectural Generalizability and Experimental Robustness. Both Qwen3.6-35B-A3B and GPT-OSS-20B are Mixture-of-Experts (MoE) models that rely on sparse expert routing. Despite differences in their architectures and reasoning formats, both share this fundamental computational mechanism. The paper does not investigate whether the effectiveness of entropy-guided intervention depends on properties specific to sparse expert routing, nor does it provide evaluations on dense models. Therefore, the claimed architectural generality remains insufficiently supported. Furthermore, the multi-seed evaluation is restricted to AIME 2025, containing only 30 distinct problems. Repeated sampling over such a small dataset offers limited robustness evidence and remains susceptible to seed-selection effects, especially without per-seed results. GPQA and HLE are evaluated with only one attempt per question, further limiting confidence in the stability and generalizability of the reported improvements.

Questions for authors
  1. Theoretical Motivation: What theoretical justification supports the use of mean token entropy as an indicator of reasoning errors or intervention utility? Can the authors provide mathematical analysis or direct empirical observations establishing this relationship?

  2. Sparse Expert Routing and Generalizability: Both evaluated models rely on sparse expert routing. How can the authors establish that the observed effectiveness of entropy-guided intervention is not specific to this shared architectural property? Can the authors provide results on dense reasoning models to substantiate the claimed generality?

  3. Entropy–Correctness Relationship: Can the authors quantify the relationship between segment-level token entropy and actual reasoning errors, for example through correlation analysis, AUROC/AUPRC, or error rates across entropy quantiles? Such evidence is essential for validating the proposed intervention criterion.

  4. Seed Robustness: What exact seeds were used in the four AIME 2025 attempts, and how were they selected? Can the authors report per-seed performance and statistical uncertainty? Given the limited diversity of the 30-problem benchmark, can the improvements be validated through additional seeds or larger reasoning benchmarks?

5 comments

We sincerely appreciate your careful review and constructive comments. Please see our responses to your comments one by one below.

Weakness #1 and Question #1: Theoretical and empirical justification for entropy-guided intervention.

Thank you very much for your comments. We would like to clarify that ThinkMid’s core design does not depend on entropy specifically, but on dynamically allocating additional thinking at uncertain points during response generation for review and correction. Entropy serves only as a signal for selecting segments to review; their correctness is assessed by the same LLM through additional reasoning in the Review stage. As highlighted in the Critical Strengths identified by Reviewers 4Aag and Qgtq, whether to request correction and how to intervene are determined by the Review and Correction stages, respectively. We adopt mean token entropy because it is simple and inexpensive, requiring no additional forward passes. Section 3.2 explicitly allows alternative uncertainty estimators, and we believe stronger signals could further improve intervention targeting, as discussed in Section 6.

Our existing experiments also support its practical utility: the placement ablation in Section 4.3 shows that entropy-guided review matches fixed-interval review’s accuracy while consuming fewer tokens, and outperforms random placement with the same average number of reviews. This controlled comparison supports its effectiveness in allocating review computation, complemented by the threshold sensitivity analysis in Section 4.4 and the qualitative intervention-point analysis in Section 5. We additionally report direct entropy–error observations in our response to Question #3 below.

Weakness #2 and Questions #2 and #4: Architectural generalizability and experimental robustness.

We appreciate your valuable suggestions.

(1) Architectural generalizability.

Following your suggestion, we additionally evaluated the dense model Qwen3.6-27B on AIME 2025. As shown in Table A1, ThinkMid outperforms all evaluated baselines. These results demonstrate that its benefits extend to a dense backbone and do not require sparse expert routing, providing additional evidence of architectural generalizability.

Table A1. Results on AIME 2025 with Qwen3.6-27B.

MethodAccuracyTokens per problem
Vanilla80.0%16,970
Thinking Forcing86.7%26,053
BoN-Judge83.3%26,270
BoN-Vote80.0%25,177
Self-Refine80.0%30,384
ThinkMid90.0%25,997

(2) Seed robustness.

The four original AIME 2025 runs used seeds 42, 43, 0, and 1. Following your suggestion, we added two further seeds (44 and 45) for Qwen3.6-35B-A3B and report per-seed results, confidence intervals, and paired significance tests in Table A2. ThinkMid improves over Vanilla under every tested seed, and the seed-level confidence interval for the paired gain remains above zero. These results support consistent improvements across seeds in this setting, rather than gains confined to a particular seed.

We evaluated GPQA and HLE with one attempt per question because these evaluation sets contain more questions than AIME, following common evaluation practice in related work. The limited rebuttal period prevented us from completing additional runs across all settings. To extend the robustness assessment beyond AIME, we commit to completing multi-seed evaluations for the remaining model–benchmark settings, and reporting the full results in the camera-ready version if the paper is accepted.

Table A2. Multi-seed results on AIME 2025 with Qwen3.6-35B-A3B. Confidence intervals are 95%; accuracy intervals are in %, and Δ is in percentage points.

SeedVanilla (%)ThinkMid (%)Δ pt [95% CI]Gained / LostMcNemar p
4270.083.3+13.34 / 00.125
4370.086.7+16.75 / 00.062
066.780.0+13.34 / 00.125
173.383.3+10.04 / 10.375
4470.080.0+10.03 / 00.25
4570.083.3+13.34 / 00.125
All seeds70.0% [62.9, 76.2]82.8% [76.6, 87.6]+12.8 [+7.8, +17.8]24 / 11.6 × 10⁻⁶
Seed-level summaryVanilla accuracy (%)ThinkMid accuracy (%)Paired gain (pt)
Mean ± SD70.0 ± 2.182.8 ± 2.5+12.8 ± 2.5
95% t-interval[67.8, 72.2][80.1, 85.4][+10.1, +15.4]

Question #3: The relationship between segment-level entropy and reasoning errors.

Thank you very much for your valuable suggestion. Following your suggestion, we annotated sentences that introduce reasoning errors in the Vanilla generation trajectories from one run of Qwen3.6-35B-A3B on AIME 2025 and calculated error rates across mean token entropy bins. The results are reported in Table A3. Sentences with mean entropy ≥ 0.5 have an error rate of 26.7%, compared with 3.6% below this threshold. These observations provide direct empirical evidence that entropy can help prioritize segments with higher error rates for review in this setting.

Table A3. Sentence-level reasoning error rates across mean token entropy bins for a Vanilla run of Qwen3.6-35B-A3B on AIME 2025.

Mean token entropySentencesSentences introducing errorsError rate
< 0.120731.4%
0.1–0.219952.5%
0.2–0.31041110.6%
0.3–0.59533.2%
0.5–0.730723.3%
≥ 0.715533.3%

Dear Reviewer UnNa,

We hope this message finds you well. We would like to express our sincere appreciation for the time you have spent reviewing our work and for your insightful comments. Your feedback has been invaluable in helping us strengthen and refine our manuscript.

In response to your review, we have made substantial revisions and additions during the discussion phase, accompanied by detailed clarifications and new experimental results. We sincerely hope that these efforts have addressed your primary concerns regarding the methodology of this work.

As the discussion phase is drawing to a close, if any aspects of our responses remain unclear or insufficient, we would be deeply grateful if you could kindly point them out. We will do our utmost to further improve and supplement the manuscript within the remaining time.

Once again, thank you very much for your time, effort, and constructive evaluation of our submission. Your guidance is greatly appreciated.

Sincerely, The Authors of Submission46126

Reviewer_4AagOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

ThinkMid pauses response generation when a sentence has high average token entropy, asks the same model to review it, and replaces the sentence if the model identifies a repair. It is a sensible, training-free decoding heuristic, with promising gains on AIME and HLE. The main issue is that the paper’s claims about novelty, superiority, and reliable correction are stronger than the evidence supports.

Critical strengths
  • The design sensibly separates uncertainty from correctness: high entropy triggers inspection, rather than automatically causing a rewrite. Allowing the reviewer and corrector to decline is also a reasonable safeguard.
  • The implementation handles both attention-only and recurrent hybrid architectures. Restoring the generation state and inserting a local correction makes the proposal fairly concrete and practical.
  • The placement and threshold ablations provide useful initial evidence that selective review can achieve similar accuracy to frequent periodic review at lower token cost.
Critical weaknesses
  • The results do not establish the broad superiority claimed. GPQA gains over vanilla amount to just two additional correct answers for Qwen and one for GPT-OSS. HLE uses only 100 questions, and no confidence intervals or paired significance analysis are reported. Table 1 also contradicts “outperforms all compute-matched baselines”: ThinkMid loses to Self-Refine on Qwen/GPQA and ties thinking forcing on GPT-OSS/AIME. The results are encouraging, but the wording needs tightening.
  • The novelty framing overlooks relevant precedents. FLARE [1] already uses low-confidence tokens to trigger sentence regeneration during an answer, while HN-decode [2] allocates additional search to high-entropy positions. These differ from ThinkMid, but they weaken the broad claim that earlier methods do not revisit uncertain content during generation. The more specific contribution is combining sentence-level triggering, same-model review, and local replacement while retaining the original reasoning trace. The claim that iterative refinement leaves responses unchecked is also plainly inconsistent with Self-Refine.
  • The experiments do not isolate which design choices matter. Compared with Self-Refine, ThinkMid changes review timing, prompts, and revision scope together. A completed-answer review using the same review/correction prompts would help establish the benefit of intervening midway. Likewise, there is no direct ablation of the verification gate or local replacement. The counterfactual analysis is useful, but its 11 intervention points include only four correct plain continuations and sometimes multiple points from one problem; that is thin evidence for corrections generally avoiding harm.
  • The baseline budget allocation leaves an important comparison missing. Vanilla is forced out of thinking at 16,000 tokens, whereas thinking forcing prevents early termination until its larger allowance is consumed. A natural-completion baseline with a larger thinking cap would sit between these extremes. Without that comparison and the fraction of runs hitting the initial cap, it is unclear how much ThinkMid repairs errors arising during response writing versus compensates for prematurely interrupted reasoning.

Minor points

  • The original thinking and earlier response remain in context, including potentially incorrect premises. Claims that continuation uses the correction “alone” should be qualified.
  • The settings are described as “officially recommended,” but Qwen recommends temperature 1.0 for general thinking tasks, rather than 0.6.

[1] Jiang, Z., Xu, F.F., Gao, L., Sun, Z., Liu, Q., Dwivedi-Yu, J., Yang, Y., Callan, J. and Neubig, G., 2023, December. Active retrieval augmented generation. In Proceedings of the 2023 conference on empirical methods in natural language processing (pp. 7969-7992).

[2] He, J., Liu, M., Olaleye, O.P., Agarwal, A., Avendi, M., Abbasi, Y., Rowe, M., Patel, H.L., Li, P., Sheng, T. and Ravi, S., 2026. Think Twice Before You Write--an Entropy-based Decoding Strategy to Enhance LLM Reasoning. arXiv preprint arXiv:2604.00018.

Questions for authors
  • What prompt is used for HLE grading, and have the solver model’s decisions to overturn exact-match failures been independently validated?
9 comments

We sincerely appreciate your careful review and constructive comments. Please see our responses to your comments one by one below.

Weakness #1:: The results do not establish the broad superiority claimed.

We sincerely appreciate your valuable suggestions and recognition of our encouraging results. We agree that “outperforms all compute-matched baselines” is imprecise and will carefully revise this statement to more accurately reflect the experimental results. Nevertheless, we believe ThinkMid performs better overall than all compute-matched baselines, achieving the best or tied-best accuracy in five of six settings and second-best in the remaining setting. On Qwen/GPQA, ThinkMid remains close to Self-Refine, the strongest baseline in this setting, achieving 84.3% accuracy compared with 85.4%. On GPT-OSS/GPQA, ThinkMid improves over Vanilla, whereas Self-Refine falls below it, demonstrating more consistent gains over Vanilla across the two models on this benchmark.

Following your suggestion, we report 95% confidence intervals and paired significance tests for HLE in Table A1. ThinkMid improves accuracy over Vanilla by 9 percentage points, with a paired 95% confidence interval of [+4, +15] and McNemar p = 0.004. Although the individual accuracy intervals overlap, the paired improvement interval excludes zero, providing statistical evidence of an improvement on the evaluated HLE subset. Furthermore, ThinkMid correctly answers nine questions that Vanilla misses while preserving all of Vanilla’s correct answers, showing that the observed gains are not offset by regressions on previously correct cases.

Table A1. HLE results with 95% confidence intervals and paired comparisons against Vanilla.

MethodCorrect / 100 (95% CI)vs. Vanilla: gained / lost, Δ pt [95% CI], McNemar p
Vanilla9 [4.8, 16.2]—
Thinking Forcing17 [10.9, 25.5]8 / 0, +8.0 [+3, +14], p = 0.008
BoN-Judge12 [7.0, 19.8]5 / 2, +3.0 [−2, +8], p = 0.45
BoN-Vote8 [4.1, 15.0]0 / 1, −1.0 [−3, 0], p = 1.00
Self-Refine13 [7.8, 21.0]4 / 0, +4.0 [+1, +8], p = 0.125
ThinkMid18 [11.7, 26.7]9 / 0, +9.0 [+4, +15], p = 0.004

Weakness #2: The novelty framing overlooks relevant precedents.

Thank you very much for your valuable suggestions. We would like to clarify that our intended claim is not that earlier methods never revisit uncertain content during generation. For reasoning models, we intended to contrast the additional thinking allocated before each response or revision by the three mainstream scaling strategies with ThinkMid’s dynamic allocation of thinking at uncertain points within an ongoing response. FLARE retrieves external information and regenerates sentences, while HN-decode explores alternative continuations through branching; neither introduces explicit review-and-correction thinking turns within the current response as ThinkMid does. Our core contribution is the mechanism for adding thinking during response generation to review and locally revise uncertain content, rather than using entropy to select intervention points or merely checking uncertain segments.

Regarding Self-Refine, we acknowledge that our wording was insufficiently precise. We intended to explain that its feedback and deliberation precede generation of the next revision, without a separate review stage interrupting that generation. The new revision may itself introduce errors that remain unchecked during its generation; this does not contradict Self-Refine’s review of previously completed outputs. We will revise the wording and incorporate FLARE and HN-decode into the related work, explicitly clarifying these distinctions.

Weakness #3: The experiments do not isolate which design choices matter.

Thank you very much for your valuable suggestions and for recognizing the value of our intervention-point analysis. Following your suggestions, we added four ablations: Post-hoc Refine reviews and revises completed answers using the same review and correction prompts; Correct Only directly corrects flagged segments without review; Entropy-Triggered Resampling directly resamples flagged segments; Review-Triggered Resampling retains the review stage but replaces correction with resampling. The results in Table A2 support the contributions of mid-response intervention, the verification gate, and targeted local correction over resampling. Regarding the intervention-point analysis, we would like to clarify that it is not intended to establish a general claim that corrections rarely cause harm. Section 5 explicitly presents it as a qualitative study of one run and states its limitations. Multiple interventions on the same problem correspond to distinct generation states, rather than independent test problems. Across the 11 intervention points, ThinkMid preserves the correct outcome in all four cases where plain continuation succeeds and produces correct answers in five cases where plain continuation fails. These paired observations support the effectiveness of the intervention mechanism in the examined cases and provide concrete evidence of how interventions can redirect generation trajectories.

Table A2. Ablation results.

MethodAccuracy (%)
Post-Hoc Refine76.7
Correct Only76.7
Entropy-Triggered Resampling73.3
Review-Triggered Resampling73.3
ThinkMid83.3

Weakness #4: The baseline budget allocation leaves an important comparison missing.

We appreciate your valuable suggestion. Following your suggestion, we evaluated Qwen3.6-35B-A3B on AIME 2025 with initial-thinking budgets of 65,562 and 81,946 tokens (the latter is Qwen’s recommended maximum output length for challenging benchmarks). As shown in Table A3, ThinkMid continues to outperform Vanilla and improves upon its own performance under the original budget, demonstrating that its effectiveness persists at these larger budgets. We acknowledge that Vanilla also improves with a larger initial-thinking budget. One major reason is that Qwen3.6 is a highly advanced model trained to autonomously extend its reasoning during the thinking phase.

Although ThinkMid solves only one additional problem, II-13, at each budget, this problem does not reach either cap: Vanilla naturally completes its initial thinking after 55,609 tokens in both settings but consistently answers incorrectly, whereas ThinkMid answers correctly in both. Simply increasing Vanilla’s initial-thinking budget from 65k to 81k therefore does not solve this problem, whereas ThinkMid does. Notably, ThinkMid achieves this correction using only 9,155 additional tokens, substantially fewer than the 16,384-token increase in Vanilla’s initial-thinking allowance. These observations demonstrate that ThinkMid can still repair an incorrect response after extensive, naturally completed initial reasoning.

We also emphasize that limited initial-thinking budgets are a common and important practical setting. Under such budgets, some problems are expected to reach the cap before reasoning completes naturally. In this setting, ThinkMid’s ability to selectively supplement and extend reasoning at critical points during response generation is itself a core advantage, alongside correcting errors in the response.

Table A3. Accuracy and initial-thinking budget hit rates on AIME 2025 with Qwen3.6-35B-A3B.

Initial-thinking budget (tokens)Vanilla accuracy (%)ThinkMid accuracy (%)Vanilla budget hit rate (%)ThinkMid budget hit rate (%)
16,00070.083.393.393.3
65,56286.790.013.313.3
81,94686.790.06.76.7

Weakness #5: Context retained during continuation.

Thank you very much for pointing this out. We agree that the original thinking and earlier response remain in context and may contain incorrect premises. We will revise the corresponding wording in the revised version to describe this more precisely.

Weakness #6: Decoding settings.

Thank you very much for your comment. Our wording referred to a decoding configuration widely adopted in related reasoning-model research: temperature 0.6 and top-p 0.95 are used in both the DeepSeek-R1 and original Qwen3 technical reports, with top-k 20 additionally specified for Qwen3. We applied this configuration consistently across all models and benchmarks, rather than adopting different recommendations for individual models or benchmarks. We agree that “officially recommended” is imprecise for Qwen3.6 and will revise the wording to clarify the source and rationale of our settings.

Question #1: HLE prompt and validation

Thank you very much for your comment. Our grading prompt is based on the official HLE judge prompt [Ref1]. It provides the question, the model’s extracted answer, and the reference answer, and instructs the grader only to determine whether the answers match, without solving the problem or arguing for an alternative answer. We will include the full prompt in the revised version. We did not use an additional verifier model; all reported grading results, including cases where the solver overturned an exact-match failure, were manually checked and verified.

  • [Ref1] A benchmark of expert-level academic questions to assess AI capabilities. Nature, 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.