acceptodds
Under review as a conference paper at ICLR 2027

MetaThink: Learning When to Think via Multi-Token Metacognitive Reasoning

Abstract

Large reasoning models achieve strong performance through chain-of-thought (CoT) reasoning, but uniformly applying lengthy deliberation to all problems incurs substantial computational overhead. Existing adaptive reasoning methods train models to selectively engage CoT, yet some rely on expensive and inconsistent external annotations, and they typically use a single-token decision mechanism that creates an information bottleneck and leads to decision boundary collapse. Inspired by metacognitive gating in human dual-process theory, we propose MetaThink, which introduces a multi-token metacognitive reasoning stage where the model generates a brief natural language self-assessment before deciding whether to engage CoT. Combined with an asymmetric reward that assigns higher rewards to correct non-CoT responses, MetaThink enables the model to autonomously discover which problems require reasoning through reinforcement learning, without external difficulty annotations. Experiments on the mathematical reasoning domain show that MetaThink achieves a competitive accuracy-efficiency tradeoff compared to existing adaptive reasoning methods across multiple model scales.

Primary area:
reinforcement learning
Keywords:
Large Reasoning Models, Reinforcement Learning with Verifiable Rewards, Adaptive Reasoning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_heZkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MetaThink, an adaptive reasoning framework based on multi-token metacognitive reasoning. It aims to reduce the computational overhead caused by large reasoning models using long CoT reasoning for problems of different difficulty levels. MetaThink introduces a multi-token natural language metacognitive stage before the actual reasoning process, allowing the model to assess whether a problem requires CoT. The authors also design a reward function that gives higher rewards to correct non-CoT responses, enabling the model to learn an adaptive reasoning strategy through GRPO. Experiments are conducted on DeepSeek-R1-Distill-Qwen-1.5B and 7B across three mathematical reasoning benchmarks.

Critical strengths
  1. The proposed method is relatively simple and can be easily integrated into the standard GRPO training framework.

  2. MetaThink achieves better experimental results than several existing adaptive reasoning methods.

Critical weaknesses
  1. Compared with single-token decisions, MetaThink introduces additional autoregressive computation and natural language generation. Therefore, it is unclear whether the performance improvements actually come from the multi-token metacognitive stage. The current evidence is not sufficient to show how much benefit the metacognitive content itself provides compared with other decision mechanisms under the same computational budget.

  2. The experiments mainly focus on mathematical reasoning tasks with verifiable answers. It remains unclear how well the proposed method generalizes to other types of tasks.

Questions for authors
  1. Could the authors provide more details or quantitative results comparing MetaThink with control methods that use a similar number of decision tokens but do not generate problem-specific metacognitive content? For example, a control method could use a fixed or generic multi-token prefix before making the same CoT/Non-CoT decision.

  2. For problems where MetaThink chooses Non-CoT, how many incorrect answers become correct when CoT is forced? Conversely, for problems where MetaThink chooses CoT, how many can still be answered correctly when Non-CoT is forced? In addition, under the same CoT usage rate or average token budget, how much accuracy improvement does MetaThink achieve compared with random routing or other simple routing strategies?

4 comments

We sincerely thank you for your insightful and constructive comments. In response, we have thoroughly revised the manuscript to enhance its clarity and rigor. Below, we address each of your comments in turn, and revisions made specifically in response to your comments are highlighted in blue\color{blue}{\text{blue}} in the revised manuscript for your convenience. Two of the new experiments were also requested by other reviewers and therefore carry their colors: the length-matched training control (Q1) is highlighted in red\color{red}{\text{red}} (shared with Reviewer 1), and the coding experiment (Q2) is highlighted in green\color{green}{\text{green}} (shared with Reviewer 2).

Q1: The benefit of metacognitive content is not shown against controls with a similar number of decision tokens but no problem-specific content.

A1: Thank you for your suggestion. Following it, we trained a Fixed variant on DeepSeek-R1-Distill-Qwen-1.5B (λ = 1.1), replacing the learned meta block with a fixed, length-matched generic template followed by the same decision statement, under identical SFT and RL settings.

MATH500 AccMATH500 CoT RateAvg TokensAvg Meta Tokens
Fixed76.80.865231.0
Ours85.479.0255731.0

The Fixed variant still collapses (CoT rate 0.8%, accuracy 76.8), while ours does not. Since the two variants differ only in whether the meta block carries meaningful content, this shows that it is the metacognitive content, not a longer decision sequence, that prevents decision collapse.

We have added this experiment to Section 4.3 (L426-L430) and Table 5 (L454-L459), highlighted in red\color{red}{\text{red}} because it is shared with Reviewer 1. The sentence relating this training-time control to the Appendix E interventions (Section 4.3, L431) is highlighted in blue\color{blue}{\text{blue}}.

Q2: The experiments focus on mathematical reasoning; generalization to other task types is unclear.

A2: Thank you for the question. We conducted an experiment in the coding domain. We trained DeepSeek-R1-Distill-Qwen-1.5B on the DeepCoder-Preview-Dataset and evaluated on MBPP+ and HumanEval+:

MBPP+ Pass@1CoT RateHumanEval+ Pass@1CoT Rate
Original Thinking48.710062.0100
Vanilla CoT RL55.710073.5100
AdaptThink54.163.871.840.6
Ours (λ=1.1)55.356.373.142.6

MetaThink matches the Pass@1 of Vanilla CoT RL while using CoT on only 56.3% (MBPP+) and 42.6% (HumanEval+) of problems, and outperforms AdaptThink in Pass@1 on both benchmarks. This shows that our method generalizes beyond mathematics. Extending it to interactive tool-use agents is a promising direction for future work.

We have added this experiment as the new Section 4.5 (L481-L485), with the results in Table 8 (L494-L503), highlighted in green\color{green}{\text{green}} because it is shared with Reviewer 2.

Q3: How do routing decisions hold up when the opposite mode is forced, and how does MetaThink compare with random or other simple routing at the same CoT rate / token budget?

A3: Thank you for the questions. We analyze these questions using MetaThink-1.5B (λ=1.2) on MATH500 because its 56.4% CoT rate gives balanced CoT and Non-CoT samples.

(1) Forced-mode analysis. We forced the opposite mode on each problem and regenerated:

Routed Mode# ProblemsSubsetCorrect after Forcing Opposite ModeRate
Non-CoT21821 incorrect1885.7% (recovery)
CoT282227 correct14865.2% (retention)

Non-CoT decisions are mostly correct. 90.4% of these problems are already solved, and forcing CoT gains only 18 problems (3.6% of MATH500), the small cost λ=1.2 accepts in exchange for efficiency. CoT decisions are necessary. Forcing Non-CoT would lose 79 correct answers (34.8%).

(2) Simple routing strategies. Under matched cost (question-length routing assigns CoT to the longest questions at the same CoT rate):

Routing StrategyMATH500 Acc
Random (same CoT rate)79.9
Random (same token budget)83.1
Question length (same CoT rate)81.8
MetaThink84.8

MetaThink outperforms these strategies by +4.9, +1.7, and +3.0, respectively. This confirms that the gains come from problem-aware routing, not just from using CoT less often.

Both analyses have been added as the new Section 4.4 (L461-L479), highlighted in blue\color{blue}{\text{blue}}: the simple routing strategies are reported in Table 6 (L468-L474) and the forced-mode analysis in Table 7 (L487-L492).

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with only four days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you again for your time and effort in reviewing our submission.

Dear Reviewer heZk,

Thank you for your careful review and constructive suggestions. We have addressed each of your concerns:

  1. We add the requested fixed, length-matched training control, supporting the contribution of metacognitive content beyond additional decision tokens.
  2. We add coding experiments demonstrating generalization beyond mathematical reasoning.
  3. We provide the requested forced-mode analysis, quantifying when switching the selected reasoning mode helps or hurts.
  4. We compare against random and question-length routing under matched costs, demonstrating the benefit of learned routing.

As the discussion phase draws to a close, we believe these experiments address your concerns about the mechanism, generalization, and routing quality. We would appreciate your confirmation and would be happy to clarify any remaining points.

Thank you again for your time and thoughtful feedback.

Best regards,

Authors

Reviewer_3yVVOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MetaThink is a framework that generates a brief self-assessment before deciding whether to produce a full chain of thought. Trained with GRPO using rewards that favor correct answers without reasoning steps, enforce format consistency, and penalize length, it requires no manual difficulty labels. Evaluations on DeepSeek-R1-Distill-Qwen models (1.5B and 7B) demonstrate improved accuracy per token on GSM8K, MATH500, and AIME 2024. Ablations confirm that the multi-token assessment enables adaptive reasoning.

Critical strengths

S1. A simple training modification improves per-token accuracy. The 7B R1-distilled model achieves 91.0% on GSM8K with fewer tokens than Vanilla CoT RL's 90.6%, and 94.0% on MATH500 using 1,978 tokens versus 92.6% with 2,291 tokens.

S2. Controlled experiments on decision collapse demonstrate that the multi-token <meta> block resolves the information bottleneck inherent in binary single-token routing, establishing a contribution beyond aggregate benchmark gains.

Critical weaknesses

W1. Sensitivity to hyperparameter tuning and lack of uncertainty esitimates: The accuracy-efficiency tradeoff depends heavily on choosing the right asymmetry factor λ ∈ {1.1, 1.2, 1.3, 1.4, 1.5}. Finding the optimal λ may be difficult when applying the framework to new domains with different token budgets and reasoning difficulty. Additionally, some experiments lack uncertainty estimates (Table 1 and 4), making it hard to interpret small accuracy differences like 91.0% versus 90.6% on GSM8K.

W2. The paper lacks baseline comparisons with relevant RL-based fast/slow and hybrid-reasoning methods, such as [1-3]. Please include these for comprehensive evaluation. Thinkless is particularly relevant because it learns short/long response selection and uses DeGRPO to mitigate single-token mode collapse.

[1] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Thinkless: LLM Learns When to Think. NeurIPS 2025.

[2] Stephen Chung, Wenyu Du, and Jie Fu. Thinker: Learning to Think Fast and Slow. NeurIPS 2025.

[3] Lingjie Jiang et al. Think Only When You Need with Large Hybrid-Reasoning Models. NeurIPS 2025.

Questions for authors

Q1. Have you explored applying the method to non-mathematical domains (such as coding or interactive tool-use agents)? How would you adapt the asymmetric reward structure when verifiable binary rewards are absent or replaced by execution-based/process-supervised rewards?

4 comments

We sincerely thank you for your insightful and constructive comments. In response, we have thoroughly revised the manuscript to enhance its clarity and rigor. Below, we address each of your comments in turn, and all associated revisions are highlighted in green\color{green}{\text{green}} in the revised manuscript for your convenience.

Q1: The tradeoff depends heavily on λ, and finding the right λ may be difficult in new domains.

A1: Thank you for raising this and we apologize for the lack of clarity. We clarify three points:

  • λ is a control knob, not a parameter with a single optimum. Under Eq. 1 (ignoring the format reward, which is the same for both modes), the expected task reward is λ·p_n for non-CoT and p_c for CoT, where p_n and p_c are the model’s success rates in each mode. The optimal policy therefore skips CoT iff λ·p_n > p_c, so λ directly sets the tolerated relative accuracy loss when skipping CoT (≈9% for λ=1.1, ≈33% for λ=1.5). Because this criterion is defined on the model’s own success rates rather than on token budgets or problem difficulty, λ has the same meaning across domains, and practitioners can set it according to their accuracy–efficiency requirement.

  • MetaThink remains competitive across the whole λ range. It generally achieves higher accuracy than the adaptive baselines (Table 1), and CoT usage consistently decreases as λ increases on every benchmark, so the tradeoff behaves predictably.

  • λ=1.1 is a safe default. It gives the highest accuracy among all λ settings in every experiment (Tables 1 and 9) while still reducing CoT usage, and the same λ grid transfers without modification across model scales (1.5B/7B) and initializations (R1-distilled and Qwen2.5 base).

We have added this interpretation of λ, the default choice λ=1.1, and the shared λ grid to Section 3.3 (L216-L220).

Q2: Some experiments lack uncertainty estimates (Tables 1 and 4), e.g., 91.0% vs. 90.6% on GSM8K.

A2: Thank you for the suggestion. All results in the paper are averaged over independent runs (Section 4.1, L304-L305). For the comparison you mention, we report mean ± standard deviation over 3 random seeds on DeepSeek-R1-Distill-Qwen-7B:

GSM8kMATH500AIME
Vanilla CoT RL90.6 ± 0.392.6 ± 0.660.4 ± 3.5
Ours (lambda=1.1)91.0 ± 0.494.0 ± 0.861.9 ± 3.3
Avg Tokens (Vanilla -> Ours)944 -> 5472291 -> 197810236 -> 8666

These results have been added as the new Table 4 (L379-L386) and are discussed in Section 4.2 (L371-L373). Please note that table numbers have shifted in the revision: the Qwen2.5 results (Table 4 in the original submission) are now Table 9.

Q3: Comparisons with RL-based fast/slow and hybrid-reasoning methods [1-3] are missing, especially Thinkless.

A3: Thank you for the suggestion. Due to limited computational resources, we first add Thinkless as a baseline on DeepSeek-R1-Distill-Qwen-1.5B under the same setting. Each cell reports Acc / CoT Rate / Avg Tokens:

GSM8KMATH500AIME
Vanilla CoT RL84.3 / 100.0 / 164585.4 / 100.0 / 375237.9 / 100.0 / 10507
Thinkless82.3 / 15.6 / 85682.9 / 68.4 / 261035.6 / 100.0 / 9285
Ours (λ=1.1)83.9 / 44.9 / 88485.4 / 79.0 / 255742.9 / 100.0 / 10722
Ours (λ=1.2)83.5 / 11.2 / 53684.8 / 56.4 / 269339.4 / 100.0 / 10202

Overall, MetaThink achieves a better accuracy–CoT tradeoff than Thinkless. Notably, Thinkless requires a modified objective (DeGRPO) to mitigate single-token mode collapse, whereas MetaThink avoids collapse with standard GRPO through its multi-token metacognitive stage.

Thinkless has been added to the compared methods in Section 4.1 (L314-L315) and to Table 1 (L338), and the comparison is discussed in Section 4.2 (L388-L390). In the revised manuscript, we also discuss all three methods [1-3] in Related Work (Section 2.2, L111-L117), and we will add experimental comparisons with [2] and [3] in a later version.

Q4-(1): Has the method been applied to non-mathematical domains such as coding or tool-use agents?

A4: Thank you for the question. We conducted an experiment in the coding domain. We trained DeepSeek-R1-Distill-Qwen-1.5B on the DeepCoder-Preview-Dataset and evaluated on MBPP+ and HumanEval+:

MBPP+ Pass@1CoT RateHumanEval+ Pass@1CoT Rate
Original Thinking48.710062.0100
Vanilla CoT RL55.710073.5100
AdaptThink54.163.871.840.6
Ours (λ=1.1)55.356.373.142.6

MetaThink matches the Pass@1 of Vanilla CoT RL while using CoT on only 56.3% (MBPP+) and 42.6% (HumanEval+) of problems, and outperforms AdaptThink in Pass@1 on both benchmarks. This shows that our method generalizes beyond mathematics. Extending it to interactive tool-use agents is a promising direction for future work.

We have added this experiment as the new Section 4.5 (L481-L485), with the results in Table 8 (L494-L503).

Q4-(2): How would the asymmetric reward be adapted when verifiable binary rewards are absent or replaced by execution-based / process-supervised rewards?

Thank you for the question and we are sorry for the lack of clarity. The asymmetric design does not rely on binary rewards. It can be written generally as

R_task(x, y) = λ · R_base(x, y) if r = ∅, and R_base(x, y) otherwise,

where R_base is any scalar reward normalized to [0, 1]; Eq. 1 is the special case with binary correctness. The optimal policy then skips CoT iff λ · E[R_base | non-CoT] > E[R_base | CoT], so λ keeps the same meaning as the tolerated relative quality loss. Note that λ is applied only to non-CoT responses rather than to the whole group, so it is not removed by GRPO’s group-wise advantage normalization.

  • Execution-based rewards: The coding experiment above (Section 4.5, L482-L483) already uses execution-based rewards (unit-test outcomes), showing that the design transfers directly. Graded signals such as the fraction of passed tests also fit naturally.
  • Process-supervised rewards: Since non-CoT responses contain no reasoning steps, we can aggregate step scores into a trajectory-level score (treating the direct answer as a single step) before applying λ. This compares the two modes on the same scale, so CoT is not favored merely for having more steps to reward.
  • Non-verifiable domains: R_base can be a reward model or LLM-as-judge score normalized to [0, 1]. The normalization matters: with negative rewards, multiplying by λ would amplify penalties instead of rewarding efficiency.

Thanks for the additional experiments. I'll raise the score.

We would like to express our gratitude for your insightful comments and constructive suggestions. We are encouraged that our responses have addressed your concerns, and we appreciate your recognition of the key contributions of our work. Your feedback has been invaluable in improving the quality and clarity of the paper, and we thank you for the time and effort you have devoted to reviewing our submission.

Reviewer_4HU7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes MetaThink, a method that trains reasoning models to decide whether to generate Chain-of-Thought (CoT) before answering. It uses a short metacognitive text block followed by a binary decision, trained through SFT and GRPO with an asymmetric reward favoring correct non-CoT answers. The authors argue that multi-token metacognitive reasoning prevents the decision collapse observed with single-token routing. Experiments on two model backbones and three mathematical reasoning benchmarks compare against existing adaptive reasoning methods, with ablations on routing and metacognitive content.

Critical strengths
  1. The idea is clean and the method is simple. Letting the model write a short self-assessment and decide for itself whether to think, with no difficulty labels and no architectural change, is an attractive formulation of adaptive reasoning. If the claim that learned metacognitive content drives the decision held, it would be a useful result for the field.

  2. The experimental effort is substantial. All adaptive baselines are RL-trained under a shared setup on two backbones with three seeds, token costs are reported alongside CoT rates, non-CoT rollouts are audited for hidden reasoning, and the single-token collapse is reproduced on both backbones with training curves. The authors also addressed concrete points from the previous round.

Critical weaknesses
  1. The central claim is not supported by the paper's own evidence. The single-token ablation (Section 4.3) shows only that a longer buffer under the same reward avoids collapse, not that the content of the buffer matters. The intervention study in Appendix E points the other way: replacing the meta block with another problem's text while keeping the decision sentence barely changes routing AUROC in Table 10a, whereas flipping the decision sentence destroys it. Routing therefore appears to be carried by the final sentence, which is functionally a single decision token. The missing control is a training-time run with a length-matched fixed template in place of the learned content.

  2. The comparison to existing adaptive methods is not reliable. The reproduced AdaptThink-1.5B numbers on AIME in Table 1 are internally inconsistent (a low CoT rate with an average length above the always-CoT baseline), all RL baselines are given an SFT cold start that their original recipes do not use, and the reproduced AdaptThink-7B is several points below its published GSM8K result at far fewer tokens. On AIME the method runs CoT on nearly every problem, so the accuracy there is not an adaptive result. Two numbers also changed between versions without explanation (the λ=1.5 CoT rate on MATH500 for Qwen2.5-7B, and AdaptThink's efficiency in Table 9).

Questions for authors
  1. Does metacognitive content actually prevent decision collapse? Please evaluate a fixed, length-matched generic meta block followed by the same decision statement, using identical SFT and RL settings. This would help distinguish the contribution of meaningful metacognitive content from that of simply using a longer decision sequence.

  2. Baseline reproducibility and fairness. Why does the reproduced AdaptThink-1.5B underperform published results on AIME, and what is its average non-CoT response length? Does adding a 4K SFT cold start to AdaptThink and AutoThink follow their original training protocols? Please also clarify the two previously identified numerical discrepancies.

4 comments

We sincerely thank you for your insightful and constructive comments. In response, we have thoroughly revised the manuscript to enhance its clarity and rigor. Below, we address each of your comments in turn, and all associated revisions are highlighted in red\color{red}{\text{red}} in the revised manuscript for your convenience.

Q1: A training-time control with a fixed, length-matched generic meta block is missing.

A1: Thank you for your suggestion. Following it, we trained a Fixed variant on DeepSeek-R1-Distill-Qwen-1.5B (λ = 1.1), replacing the learned meta block with a fixed, length-matched generic template followed by the same decision statement, under identical SFT and RL settings.

MATH500 AccMATH500 CoT RateToken LenMeta Token Len
Fixed76.80.865231.0
Ours85.479.0255731.0

The Fixed variant still collapses (CoT rate 0.8%, accuracy 76.8), while ours does not. Since the two variants differ only in whether the meta block carries meaningful content, this shows that it is the metacognitive content, not a longer decision sequence, that prevents decision collapse.

Regarding Appendix E (Table 15), the intervention study probes what the trained model relies on at inference time, whereas our claim concerns the role of meta content during training. The two are consistent: the decision sentence carries the final routing decision, while meaningful meta content is what allows this decision to be learned without collapse, as the training-time control above directly confirms.

We have added this experiment to Section 4.3 (L426-L430) and Table 5 (L454-L459), with a pointer from Section 3.1 (L138-L139). The relation between this training-time control and the inference-time interventions is stated at the end of Appendix E (L1229-L1232).

Q2: Concerns about the reproduced baselines and several reported numbers.

A2: Thank you for the detailed comments. We address each point below.

(1) AdaptThink-1.5B on AIME shows a low CoT rate but an average length above the always-CoT baseline. What is its non-CoT response length?

The Non-CoT responses of the reproduced AdaptThink on AIME are indeed very long. The reason is that Original NoThinking already produces 5,919 tokens on AIME (Table 1), and AdaptThink's objective places no length constraint on the non-thinking branch. As a result, its Non-CoT responses still contain implicit reasoning, which explains the long average length:

AIME Non-CoT Token Len
AdaptThink10523

We have added this explanation to Section 4.2 (L404-L407); the corresponding AdaptThink numbers are in Table 1 (L336).

(2) Does the 4K SFT cold start added to AdaptThink and AutoThink follow their original training protocols?

No, it is our addition. Our method requires an SFT cold start, which allows it to apply to any model rather than only reasoning models. Comparing against baselines without this stage would be unfair, so to align settings, we added an SFT cold start to each baseline that strictly follows its own training format. AdaptThink and AutoThink share the same cold-start checkpoint as their RL starting point, so we report them together. As shown below, the cold start does not change model performance, so the reported results reflect the baselines' performance under the same setting:

AIME AccCoT Rate
DeepSeek-R1-Distill-Qwen-1.5B29.4100
AdaptThink / AutoThink SFT Cold Start29.2100
Ours Cold Start (forced CoT)28.8100

The added SFT cold start is now stated explicitly in the Training Details of Section 4.1 (L264-L269), and this comparison is reported in Table 2 (L359-L365).

(3) The reproduced AdaptThink underperforms its published results: AdaptThink-1.5B on AIME, and AdaptThink-7B on GSM8K (several points lower at far fewer tokens).

Both gaps, for AdaptThink-1.5B on AIME and for AdaptThink-7B on GSM8K, come from unifying the settings across all methods: in our experiments, every method is trained with a learning rate of 1e-6 and evaluated at temperature 0.7, which differs from the original AdaptThink setup. The SFT cold start is also our addition, but as shown in (2) it does not change model performance. These differences from the original AdaptThink recipe are now stated explicitly in Section 4.1 (L268-L269); the shared training settings are listed at L261-L264 and the shared evaluation settings at L320-L322.

(4) On AIME the method uses CoT on nearly every problem, so the accuracy there is not an adaptive result.

For 1.5B and 7B models, direct answers rarely succeed under the L_max limit on AIME-level problems in the training distribution, so the policy naturally converges to CoT. This reflects model capacity, not a limitation of our method. With a stronger model, easier AIME problems become solvable without CoT, and the CoT rate should drop. To verify this, we applied our full pipeline (SFT cold start + RL) to DeepSeek-R1-Distill-Qwen-14B:

AIME AccAIME CoT Rate
Original Thinking69.6100
Vanilla CoT RL74.0100
MetaThink (lambda=1.2)69.889.6

On 14B, the CoT rate drops to 89.6% while accuracy stays close to the thinking mode (69.8 vs. 69.6), confirming that our method routes adaptively when the model's capacity allows. These results are reported in Table 3 (L359-L365) and discussed in Section 4.2 (L396-L399).

(5) Two numbers changed between versions without explanation (the λ=1.5 CoT rate on MATH500 for Qwen2.5-7B, and AdaptThink's efficiency in Table 9).

For the λ=1.5 CoT rate on MATH500 for Qwen2.5-7B, the evaluation script in the earlier version used an incorrect parsing method; the current number comes from the corrected script and appears in Table 9 (L838) of the revised manuscript. For AdaptThink's efficiency (Table 9 in the original submission, now Table 14, L996), its training settings in the earlier version were not aligned with the other methods; the current number is obtained under the aligned settings.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with only four days remaining, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we're eager to address any remaining issues to improve our work.

Thank you again for your time and effort in reviewing our submission.

Dear Reviewer 4HU7,

Thank you for your careful review and constructive comments. We have addressed each of your concerns:

  1. We add the requested fixed, length-matched training control, supporting the role of learned metacognitive content in preventing decision collapse.
  2. We clarify the distinction between training-time learning and inference-time routing, reconciling the new control with the intervention results.
  3. We clarify the baseline reproduction settings, provide a cold-start comparison, and explain AdaptThink’s long non-CoT responses.
  4. We add a larger-model experiment demonstrating adaptive CoT usage on AIME.
  5. We explain and correct both numerical discrepancies identified in your review.

As the discussion phase draws to a close, we believe these additions address your concerns. We would appreciate your confirmation and would be happy to clarify any remaining points.

Thank you again for your time and thoughtful feedback.

Best regards,

Authors

Comments on the paper

3 comments
Area_Chair_9E25OpenReview ↗

Please Review the Author Rebuttal and Participate in the Discussion

Dear Reviewers,

Thank you for your efforts in reviewing this submission. As the discussion period is approaching its end, I would like to kindly encourage everyone to actively participate in the discussion and carefully consider the authors' rebuttal.

In particular, if you have not yet responded, please share your thoughts on whether the authors have adequately addressed your concerns. Please also engage with fellow reviewers where appropriate, and provide your final assessment, including any score updates or a brief justification for maintaining your original rating.

Your timely input would be greatly appreciated and will help us reach a fair and well-informed decision.

Thank you again for your time and valuable contributions!

Dear Area Chair 9E25,

Thank you for your time and for the helpful reminder. We have posted a brief summary of our rebuttal under each review and look forward to the reviewers' replies. At the end of the discussion phase, we will also post an overall summary of the rebuttal for you and all reviewers, which we hope will assist your evaluation.

Thank you again for your time.

Best regards,

Authors

AuthorsOpenReview ↗

Summary of Discussion Status and Revisions

Many thanks to the PCs, SAC, AC, and reviewers for their time during the discussion. We have addressed all concerns raised by the three reviewers, as summarized below together with our main revisions.

Reviewer responses

  • For Reviewer 4HU7, we added the requested fixed meta block control and a 14B experiment, and clarified our baseline setup. We have not yet had the chance to hear back from the reviewer.
  • For Reviewer 3yVV, we added the Thinkless baseline, seed variance, a principled interpretation of λ, and a coding experiment, after which the reviewer raised the score.
  • For Reviewer heZk, we added the requested fixed meta block control, routing quality analyses, and a coding experiment. We have not yet had the chance to hear back from the reviewer.

We would welcome any further feedback.

Key additions

(highlighted in the revised PDF in red\color{red}{\text{red}} for Reviewer 4HU7, green\color{green}{\text{green}} for Reviewer 3yVV, and blue\color{blue}{\text{blue}} for Reviewer heZk)

  • Content, not length, prevents collapse (Sec. 4.3). With a fixed meta block of matched length, as Reviewers 4HU7 and heZk requested, the model still collapses to 0.8% CoT on MATH500, while MetaThink keeps 79.0% CoT and 8.6 points higher accuracy.
  • Routing quality (Sec. 4.4). At matched cost, MetaThink beats random and question length routing by 1.7 to 4.9 points, and forcing the opposite mode shows that its decisions reflect actual reasoning needs.
  • Baselines and reliability (Sec. 3.3, 4.1, 4.2). We added Thinkless, over which MetaThink achieves a better tradeoff, clarified our unified baseline setup, and added seed variance and a principled interpretation of λ.
  • Scale and domain (Sec. 4.2, 4.5). On a 14B model, AIME CoT usage drops to 89.6% while matching the original thinking accuracy. On coding benchmarks, MetaThink stays within 0.4 Pass@1 points of full CoT RL while using CoT on only about half of the problems.

We would be glad to clarify anything further. We are especially grateful to the AC for proactively reminding the reviewers to join the discussion, and to all reviewers for their constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.