acceptodds
Under review as a conference paper at ICLR 2027

First Generate, Then Delete: Test-Time Thinking-Token Deletion for Multimodal Reasoning

Abstract

Large reasoning models typically solve complex problems by generating long chain-of-thought traces before producing final answers. In this work, we observe a phenomenon that not every generated thinking token is equally beneficial for subsequent answer decoding. Motivated by this observation, we propose Test-time Token Deletion (TTD), a first-generate-then-delete framework for test-time reasoning intervention. Given a question, the model first produces an initial reasoning trace, and TTD then selectively deletes a subset of thinking tokens and re-decodes the final solution conditioned on the remaining trace. To select tokens for deletion, we instantiate TTD with a training-free entropy-guided rule and a test-time mask-learning variant optimized with forward-only evaluations, while keeping the base model frozen. Extensive experiments on various multimodal reasoning benchmarks show that TTD improves reasoning accuracy while reducing response length. Our findings suggest that selective thinking-token pruning provides a complementary axis of test-time scaling beyond generating more tokens.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Multimodal Large Language Models, Chain-of-Thought Reasoning, Test-Time Scaling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1SL9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Test-time Token Deletion (TTD), a novel "first-generate-then-delete" framework for test-time reasoning intervention in Multimodal Large Language Models (MLLMs). The authors observe that not all generated intermediate thinking tokens (Chain-of-Thought traces) are equally beneficial when subsequently used as a conditioning context for final-answer decoding. To address this, TTD first allows the model to produce a complete reasoning trace, then selectively prunes a subset of thinking tokens, and finally re-decodes the final answer based on the remaining trace. To implement token pruning without modifying the base model parameters, the authors introduce two instantiations:

  1. Training-free Entropy-Guided Deletion: Uses truncated top-kk token entropy during generation as an uncertainty signal to filter out high-entropy thinking tokens (often function words or discourse markers that act as distractors during final-answer decoding).
  2. Test-time Mask Learning: Optimizes an instance-adaptive mask via forward-only Simultaneous Perturbation Stochastic Approximation (SPSA), guided by a joint objective combining output entropy minimization and a local trace-perturbation robustness term.

Extensive experiments across eight multimodal mathematical and general reasoning benchmarks demonstrate that TTD consistently improves reasoning accuracy (by up to 4.7%) while simultaneously reducing response lengths by roughly 38%. Furthermore, the method exhibits strong transferability across model scales, pure-text reasoning tasks, and tool-augmented multimodal workflows.

Critical strengths
  1. A Counter-Intuitive and Insightful Perspective: Traditional test-time scaling methods focus on generating more tokens (e.g., via longer CoTs, tree search, or self-consistency). This paper challenges the dogma by demonstrating that deleting redundant or distracting thinking tokens from a completed trace can improve accuracy while shortening response lengths.

  2. Effective Training-Free and Forward-Only Design: By keeping the base MLLM frozen and relying either on a lightweight entropy threshold or a forward-only SPSA mask-learning strategy, TTD avoids the heavy computational or memory burdens typically associated with gradient-based model fine-tuning or full parameter adaptation at test time.

  3. Thorough and Broad Empirical Validation: The evaluation is exceptionally comprehensive, spanning multiple model families and parameter sizes (Qwen2.5-VL, ThinkLite, MM-Eureka, LLaMA3.2-Vision), diverse benchmark categories (math, general reasoning, science), pure-text tasks, and tool-augmented workflows.

Critical weaknesses
  1. Inference Latency Overhead of Mask Learning: While the training-free entropy-guided variant is efficient (adding only ~0.125s per example), the test-time mask-learning variant requires T=80T=80 SPSA optimization iterations—taking around 20 seconds per example. This high latency overhead diminishes its practical appeal for real-time interactive applications despite reducing token length.

  2. Sensitivity to Hyperparameters and Thresholds: The performance of the training-free variant relies heavily on the deletion ratio/threshold τ\tau, while the mask-learning variant involves several sensitive hyperparameters (such as perturbation scale ϵ\epsilon, step size η\eta, and robustness weight λ\lambda). The paper lacks a deeper theoretical analysis or automatic tuning mechanism to ensure stability across out-of-domain inputs.

Questions for authors
  1. Error Propagation in Mask Learning: Since the test-time mask learning relies on forward-only SPSA optimization guided by entropy and local perturbation robustness, how does the framework prevent cases where the model is confident in a wrong pseudo-answer, thereby locking in an incorrect deletion mask?

  2. Granularity of Deletion: TTD evaluates deletion primarily at the individual token level. Have the authors explored chunk-level or sentence-level token pruning (e.g., removing entire reasoning sub-steps or redundant clauses)? Would coarser granularity better preserve logical continuity while reducing optimization search space?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1SL9 (1/3)

We appreciate very much your constructive comments on our paper. Please kindly find our response to your comments below, and all revisions made to the paper are highlighted in blue for your ease of reference. We hope that our response satisfactorily addresses the issues you raised. Please feel free to let us know if you have any additional concerns or questions.

W1. Further Discussion on the Complementary Roles and Practical Applicability of TTD.

  • We appreciate the reviewer’s valuable feedback.
    • We would first like to humbly highlight the complementary roles of entropy-guided and mask-learning TTD, as discussed in Appendix F.2. We agree with the reviewer that entropy-guided TTD is better suited to latency-sensitive interactive applications. It requires no mask optimization and introduces only one additional answer re-decoding stage. Mask-learning TTD devotes additional computation to instance-adaptive context-mask optimization. We have added new Figure 4 to make this accuracy–runtime trade-off explicit.
    • The motivation for mask-learning TTD is to provide an additional axis of test-time adaptation when more computation is available, by optimizing the model’s self-generated reasoning context while keeping its parameters frozen. Kindly refer to our comparison with MM-UPT in Section 4.2 and Table 16 of Appendix F.4, which performs unsupervised RL-based model adaptation and incurs an estimated adaptation cost of 60.48s per sample. TTD achieves higher average reasoning accuracy than MM-UPT while takes approximately 20s per example as well as avoids backpropagation through the base model and the storage of its gradients and optimizer states. The results demonstrate the potential value of reasoning-context optimization for improving test-time performance without updating the underlying model.
    • Furthermore, we would like to humbly highlight the results in Section 4.1 and Table 11 of Appendix E.3. ARES uses difficulty-aware post-training to improve reasoning efficiency, including reducing unnecessary reasoning on easier problems. The efficiency-oriented post-training is appealing for real-time interactive applications. Applying TTD to ARES further improves average accuracy while reducing reasoning context length on both mathematical benchmarks and general multimodal benchmarks. The results suggest that our selective reasoning context deletion can provide benefits even for a model already post-trained for efficient reasoning.
      • Motivated by this result, we hope our framework can pave the way for future work that incorporates feedback from TTD into iterative self-improvement post-training, with the aim of teaching models to generate more effective reasoning traces directly. Such integration could potentially benefit real-time interactive applications.
AuthorsOpenReview ↗

Response to Reviewer 1SL9 (2/3)

W2. Clarification of hyperparameter sensitivity and automatic selection.

  • We appreciate the reviewer’s valuable feedback.
    • We would first like to humbly highlight that our evaluation covers 16 benchmarks with 23463 evaluation examples, spanning multimodal mathematical reasoning, general multimodal reasoning, tool-augmented multimodal reasoning, and text-only mathematical and scientific reasoning. We also evaluate TTD across eight base models. The experiments provide empirical support for the applicability of TTD across the evaluated tasks, model families, and scales.
    • We would also like to humbly highlight that we have examined hyperparameter sensitivity in Figure 5 and Table 5. The results show TTD's relatively stable performance over moderate deletion ratios and within a neighborhood of the selected optimization settings. For baselines, we use the same threshold sweep protocol, as described in Appendix C.5. We also compare deletion-based methods under matched deletion budgets to control for the number of thinking tokens removed.
    • Following the reviewer’s suggestion and previous work [1], we conducted additional experiments using 30 randomly sampled, labeled validation examples from each benchmark to select hyperparameters based on validation accuracy. For entropy-guided TTD, we select the deletion ratio from {0.1,0.2,0.3,0.4}. For mask-learning TTD, the candidate values are {0.1,0.3,0.5,0.7} for the perturbation scale, {0.1,0.2,0.3,0.4} for the step size, and {1,2} for the robustness weight. The selected hyperparameters are then fixed for all remaining evaluation examples.
      • The results, shown in the table below, indicate that the automatically selected configurations achieve comparable or even better accuracy than our defaults.
MethodMathVistaMMMU
Entropy-Guided TTD69.5052.78
Entropy-Guided TTD + automatic selection69.8052.55
Mask Learning TTD71.5353.41
Mask Learning TTD + automatic selection71.9053.67

[1] Pushing the Limits of Simple Pipelines for Few-Shot Learning: External Data and Fine-Tuning Make a Difference. In CVPR, 2022

AuthorsOpenReview ↗

Response to Reviewer 1SL9 (3/3)

Q1. Error propagation in mask learning.

  • We appreciate the reviewer’s valuable feedback. Entropy minimization alone can be unreliable at test time, motivating us to introduce the additional robustness term. This term penalizes changes in predictive confidence under perturbations of the retained reasoning trace, discouraging solutions whose confidence depends strongly on a particular arrangement of the generated context. Predictions are also recomputed for each candidate mask, so the optimization does not enforce agreement with a fixed initial pseudo-answer. As shown in Table 1, adding this robustness constraint improves accuracy over entropy-only mask learning and narrows the performance gap to the oracle learner.

Q2. Further analysis of deletion granularity

We appreciate the reviewer’s insightful suggestion and have conducted additional experiments to examine the effect of deletion granularity.

  • Specifically, we compare the original token-level mask learning with two sentence-based variants: (1) sentence mask learning, which optimize one mask variable shared by all tokens within each sentence; and (2) token mask learning + sentence voting, which first learn individual token masks and then apply majority voting within each sentence to retain or remove it as a whole. For the latter, a sentence is retained if at least half of its tokens are selected by the learned token masks. We regenerate the final answer after applying the resulting masks to the same reasoning context.
  • The results, shown in the below table, indicate that token-level masking achieves the highest accuracy on all four benchmarks, outperforming direct sentence masking and post-hoc sentence voting. Sentence voting retains improvements over the base model, but reduces accuracy relative to the original token masks on every benchmark. The result suggests that useful and redundant or distracting thinking tokens can coexist within the same sentence, and thus assigning a single decision to the entire sentence may retain unnecessary tokens or discard useful conditioning context.
  • We agree that direct sentence masking reduces the number of learnable mask variables, from one per token to one per sentence. Post-hoc sentence voting, by comparison, changes only the final deletion granularity and leaves the token-level optimization unchanged. The results suggest a trade-off between reduced mask dimensionality and finer control over context deletion, supporting our choice of token-level deletion.
  • We have incorporated these comparisons and discussion into the revised manuscript.
MethodMathVisionMathVerseMathVistaWeMathAverage
Base Model25.0643.6466.8363.5449.77
Mask Learning TTD (token mask)27.1646.3071.5367.6453.16
Mask Learning TTD (sentence mask)25.3642.7269.0064.7750.46
Mask Learning TTD (token mask + sentence voting)25.8245.2370.9066.1552.03

Dear Reviewer 1SL9,

Thank you again for your time and constructive feedback. We would greatly appreciate your feedback, and we would be happy to provide additional clarification.

Thank you again for your time and consideration.

Best regards, The authors

Reviewer_JYH3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Based on the observation that not every thinking token is equally beneficial for subsequent answer decoding, this paper proposes Test-time Token Deletion (TTD): the model first generates initial reasoning traces, selected tokens are then deleted by TTD, and the final answer is re-decoded based on the remaining trace. TTD is instantiated with a training-free entropy-guided deletion rule and a forward-only simultaneous perturbation stochastic approximation (SPSA) mask-learning variant. Based on Qwen2.5-VL-7B, the proposed method achieves up to a 4.7% Pass@1 improvement while reduces approximately 38% responses length. Additional experiments cover tool-augmented reasoning, pure-text reasoning.

Critical strengths
  1. The paper conducts extensive experiments across diverse reasoning settings, including multimodal mathematical reasoning, tool-augmented reasoning, and text-only reasoning, as well as across multiple base models and model scales, including Qwen2.5-VL-3B/7B, ThinkLite-VL-7B, and MM-Eureka-7B. These results provide strong empirical support for the effectiveness of the proposed method.
  2. The paper is generally well written and easy to follow, with a clear presentation of the proposed TTD method. In addition, Fig.2 and Fig.3 provides an intuitive visualization of the frequency of successfully deleted tokens and high-entropy tokens, offering a concrete illustration of how TTD operates in practice.
Critical weaknesses

The end-to-end efficiency evaluation could be more comprehensive. Although TTD reduces the length of the final reasoning trace, it still incurs the computational cost of generating the initial CoT. The reported approximately 0.125-second entropy-guided overhead in Appx. F.2 does not include this initial generation cost, while the mask-learning variant requires approximately 20 seconds per example (Tab.12). Reporting end-to-end wall-clock time and FLOPs, including initial CoT generation, token deletion, mask optimization, and final answer decoding, under matched compute and output-length budgets would provide a clearer assessment of the practical efficiency gains.

Questions for authors

See Weaknesses part.

1 comment
AuthorsOpenReview ↗

Response to Reviewer JYH3

Thank you sincerely for your thoughtful feedback on our work. Below, we have provided a detailed explanation for your concerns as follows. Please do not hesitate to let us know if you have any further questions.

Clarification of computational costs and the accuracy–runtime trade-off

We sincerely thank the reviewer for the constructive comments and have incorporated the following discussion into our revised manuscript.

  • We have revised Table 12 (Table 13 in the revised manuscript) to explicitly include initial reasoning-trace generation. Initial generation takes approximately 8 s/example. The training-free entropy-guided TTD requires one additional forward for final-answer re-decoding after token deletion, taking approximately 0.125 s/example and yielding an estimated total runtime of 8.125 s/example.
  • We have also added a new Figure 4 in the revised manuscript, combining the accuracy results from Table 2 with the end-to-end runtime analysis from Table 13 to directly examine the accuracy–runtime trade-off between competitive methods. The results show that entropy-guided TTD achieves higher average accuracy than MTI, TTI, and s1 with a lower estimated end-to-end inference time. Mask-learning TTD further improves accuracy, with additional computation devoted to instance-adaptive mask optimization.
Reviewer_zD3SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Test-time Token Deletion (TTD) for multimodal reasoning. The model first generates a complete reasoning trace, removes selected tokens, and answers again using the remaining trace. The model weights stay frozen. Token selection uses either an entropy threshold or test-time mask optimization. Experiments show accuracy gains across several benchmarks and model backbones, alongside shorter retained traces.

Critical strengths
  1. The paper makes a useful distinction between tokens needed to generate reasoning and tokens useful when that reasoning is later read as context. Editing completed traces offers an interesting way to improve final answers without updating model weights.
  2. The evaluation is broad. It covers multiple benchmarks and backbones, with additional text-only and tool-use experiments. Random deletion, repeated decoding, and self-consistency comparisons provide useful supporting evidence.
Critical weaknesses
  1. The reported entropy values do not clearly match the definition. Equations (2)–(3) imply that truncated entropy cannot exceed full vocabulary entropy. For a vocabulary around 150K, the upper bound is approximately 11.92 nats or 17.19 bits. Yet Figure 3 and Table 11 report values above these bounds, including 27.02. An omitted scaling or aggregation step might explain this, but the paper needs to clarify it. The actual top-k value is also missing.
  2. The main results do not fully separate the benefits of re-answering, deletion, and mask search. Figures 2 and 9 show that re-answering with the complete trace already improves accuracy, but this control is absent from Table 2. The two TTD variants also need clearer, separate reporting. Finally, shorter retained traces do not necessarily mean lower total generation cost. The original trace and search-time answer decoding must still be counted.
  3. The baseline comparison is incomplete for assessing novelty. Most competing methods extend or revise reasoning, while related context-pruning methods are discussed without direct comparisons. Attention-based token selection and simple position- or sentence-based pruning would be useful baselines under the same re-answering protocol. Without these comparisons, it is unclear whether the proposed selection strategies offer a meaningful advance over existing ways to prune context.
Questions for authors
  1. How exactly are the entropy values in Figure 3 and Table 11 computed? Please specify k, the log base, any normalization or scaling, and whether values are averaged or summed.
  2. Using the same initial traces and seeds, how do full-trace re-answering, entropy-threshold deletion, and mask-learning TTD compare? Which variant is reported in Table 2?
  3. What does “Length” include? Please report end-to-end costs for both variants, including initial trace generation and mask search. Does the advantage remain against alternatives given the same computation budget?
  4. How does TTD compare with attention-based token selection or simple position- and sentence-based pruning, using the same deletion budget and re-answering protocol? These comparisons would help establish the value of the proposed selection strategies.
7 comments
AuthorsOpenReview ↗

Response to Reviewer zD3S (1/2)

We sincerely thank the reviewer for providing valuable feedback. We detail our response below point by point. Some experimental results have been updated in the revised paper, and any modifications made to the paper are highlighted in blue for your convenience. Please kindly let us know whether you have any further concerns.

W1 & Q1. Clarification of entropy computation in Figure 3 and Table 11.

We appreciate the reviewer’s valuable suggestion and apologize for any potential confusion.

  • The ”average thinking-token entropy” reported in Figure 3(a) and Table 11 refers to the average reasoning context entropy. Specifically, we sum the thinking-token entropy scores within each reasoning context, without dividing by its length, and then average these sums across samples in the corresponding group. Figure 3(b) uses token-level entropy.
  • We use k=50 and the natural logarithm base (i.e., e) without additional normalization.
  • We have revised the captions of Figure 3(a) and Table 11 (Table 12 in the revised manuscript) to use “average reasoning context entropy” and clarified the computation procedure in the revised manuscript.

W2 & Q2. Clarification of the TTD variants and re-answering ablation.

We appreciate the reviewer’s valuable suggestion.

  • We use Table 2 to compare TTD with representative test-time intervention and scaling methods. Both TTD (Ours) and TTD (Oracle) in this table refer to the Mask-Learning variant, with the oracle using ground-truth feedback only for analysis. Figure 11 evaluates Entropy-Guided TTD.
  • In addition to Figures 2 and 9, Figure 5 includes Directly Answer w/ CoT as a full-trace re-answering ablation control, using the same initial reasoning traces, random seeds and evaluation protocol. The controlled comparison demonstrates the effectiveness of entropy-guided deletion in improving reasoning performance.
  • We follow the reviewer’s suggestion and report Entropy-Guided TTD and Mask-Learning TTD separately in the table below. Both variants improve over the base model and achieve higher average accuracy than existing intervention and scaling baselines in both benchmark groups. The results show that training-free entropy-guided variant already yields improvements, while instance-adaptive mask learning provides further gains.
  • We thank the reviewer again for this helpful suggestion. We have incorporated these clarifications and the separate reporting of both variants into the revised Table 3.
MethodMathVisionMathVerseMathVistaWeMathAverageMMMUScienceQAA-OKVQAMMBenchAverage
Base Model25.0643.6466.8363.5449.7751.4485.2182.3284.7975.94
Entropy-Guided TTD26.7145.4169.5066.4452.0252.7885.7284.5485.2677.08
Mask Learning TTD27.1646.3071.5367.6453.1653.4186.0084.8985.7877.52

W2 & Q3. Clarification of Length and end-to-end computational costs

We deeply appreciate the reviewer’s feedback.

  • Length denotes the average number of thinking tokens in the reasoning context used for final-answer decoding. For Random Deletion and TTD, this corresponds to the retained context after deletion. We use this metric to characterize how selective thinking-token deletion changes the reasoning context supplied to the final-answer decoding and highlight that our framework provides a complementary axis of test-time scaling through editing generated reasoning context via deletion, beyond generating more tokens. We have clarified this definition in the revised manuscript.
  • We agree with the reviewer that length alone does not capture total inference cost (e.g., traces of the same length can incur different costs depending on the number of forward computations). We would like to humbly clarify that our analysis treats length and computational cost as distinct analysis, with the latter evaluated separately using inference time in Section 4.2 and Appendix F.2 (Table 13).
  • Following the reviewer’s great suggestion, we have added a new Figure 4 in the revised manuscript, combining the accuracy results from Table 2 with the runtime analysis from Table 13 and including cost of initial trace generation.
    • As shown in the Figure 4, Entropy-guided TTD achieves higher average accuracy than MTI, TTI, and s1 while requiring less end-to-end time. Mask-learning TTD further improves accuracy, with additional computation devoted to adaptive mask optimization.
AuthorsOpenReview ↗

Response to Reviewer zD3S (2/2)

W3 & Q4. Comparisons with attention-, position-, and sentence-based pruning.

  • We appreciate the reviewer's insightful feedback.
    • We would first like to humbly clarify the relationship of TTD to existing context-pruning methods to further complement the discussion in Appendix B.
      • Attention-based KV-cache compression methods such as KVzip and H₂O primarily seek to reduce memory usage and inference cost while preserving task performance, and often operate on prefilled prompt or cached states during decoding.
      • Although TTD shares the broad idea of token deletion with previous context-pruning methods, our work specifically studies completed, self-generated reasoning traces as conditioning context for subsequent final-answer decoding. Separating reasoning generation from its subsequent use as context allows us to investigate the distinction between tokens useful for generating reasoning and tokens useful when that reasoning is later read. Our leave-one-out deletion analysis examines this distinction and motivates entropy-guided deletion and instance-adaptive mask learning with a frozen base model.
    • Following the reviewer’s great suggestion, we conducted additional experiments for direct comparisons with alternative context-pruning strategies on four mathematical and four general multimodal benchmarks.
      • All methods use the same initial reasoning traces, target deletion budgets, and re-answering protocol. Pruning is restricted to the generated reasoning context and the original question and image remain unchanged.
      • Sentence-based deletion. We include random deletion at sentence granularity and a sentence-level entropy-guided TTD variant that ranks sentences by their mean token entropy.
      • Position-based deletion. Head, Middle, and Tail Deletion remove contiguous token spans from the beginning, middle, and end of the generated reasoning context, respectively.
      • For the attention-based baselines, KVzip uses context reconstruction-attention to identify important KV pairs, while H₂O combines accumulated-attention heavy hitters with recent tokens. We treat the completed reasoning trace as prefill context and adapt the importance criteria of KVzip and H₂O to the same token deletion and re-answering protocol. This comparison evaluates their attention-based criteria within our intervention setting. Specifically, we follow their standard KV pair scoring rule and then average resulting scores across heads and layers to obtain one importance score per reasoning token. For H₂O, we reserve 50% of the retention budget for the most recent reasoning tokens.
        • The results, shown in the table below, yield three findings. (1) The deletion criterion matters at both token and sentence granularities. Entropy-guided TTD outperforms random deletion at both granularities. And finer token-level deletion providing additional gains; (2) Tail Deletion harms all four mathematical benchmarks. This pattern echoes the usefulness of preserving recent context in previous method (e.g, H₂O). However, position-based strategies do not provide uniform benefits across tasks; (3) The proposed TTD provides obvious gains beyond the attention-based alternatives on both mathematical benchmarks and general multimodal benchmarks.
      • We thank the reviewer again for this helpful suggestion. We have incorporated these additional comparisons, implementation details, and clarifications into the revised manuscript.
MethodMathVisionMathVerseMathVistaWeMathAverageMMMUScienceQAA-OKVQAMMBenchAverage
Base Model25.0643.6466.8363.5449.7751.4485.2182.3284.7975.94
Random Deletion (token)24.3443.6866.6765.0349.9349.5584.4683.1984.9075.52
Random Deletion (sentence)24.7644.0168.0062.9849.9450.2284.8782.9684.9175.74
Head-Position Deletion (token)26.0844.9868.9064.7751.1849.7884.4383.4984.5475.56
Middle-Position Deletion (token)25.6544.8968.5063.8550.7250.5685.2782.6284.8075.81
Tail-Position Deletion (token)23.4238.7562.3059.1945.9249.0083.0484.4584.9675.36
KVZip (token)25.7244.5465.4062.8749.6351.4485.4283.5885.2376.41
H$_2$O (token)26.1546.0168.6065.1151.4651.6785.6282.4485.0576.19
Entropy-guided TTD (sentence)26.2144.9468.5065.6351.3251.8985.4784.2785.0576.67
Entropy-guided TTD (token)26.7145.4169.5066.4452.0252.7885.7284.5485.2677.08
Mask Learning TTD (token)27.1646.3071.5367.6453.1653.4186.0084.8985.7877.52
Reviewer_zD3SOpenReview ↗

Response to authors

Thanks for your response and the new comparisons. However, I still have several concerns.

First, sum token entropies without length average (Figure 3) explains the large values, but it also mixes entropy with CoT length. So it's still unclear whether wrong traces really have higher per-token uncertainty or are simply longer. Next, I notice that full-trace reanswering already get 51.32% acc in your Figure 10. Based on your current results, it seems that ''more than 50% of the gain'' may come from reanswering itself rather than token deletion. I also want to know how the reported 20 seconds (Figure 4) relates to the initial CoT generation time is confusing. I remain unsure whether mask learning retains its advantage over parallel sampling under a matched end-to-end inference budget.

I therefore maintain my original rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of changes

We extend our sincere thanks to the reviewers for their constructive feedback. We have summarized additional experiments and clarifications made during the rebuttal period as follows. Revisions to the manuscript are highlighted in blue.

Clarifications:

  1. Clarified entropy computation, the top-k setting, and the logarithm base, and revised the corresponding figure and table captions. (Reviewer zD3S W1 and Q1)
  2. Reported entropy-guided TTD and mask-learning TTD separately. (Reviewer zD3S W2 and Q2)
  3. Clarified our finding and ablation study on the benefits of full-trace re-answering and distinguished these benefits from the additional gains achieved by selective deletion. (Reviewer zD3S W2, Q2, and follow-up Q2)
  4. Expanded the discussion of TTD’s relationship to existing context-pruning methods. (Reviewer zD3S W3 and Q4)
  5. Clarified Length definition and details in end-to-end computational costs, and added an accuracy–runtime analysis in the new Figure 4 (Reviewer zD3S W2, Q3 follow-up Q3; Reviewer JYH3; Reviewer 1SL9 W1)
  6. Clarified the robustness objective during mask learning. (Reviewer 1SL9 Q1)
  7. Expanded the discussion of the complementary roles and practical applicability of entropy-guided TTD and mask-learning TTD. (Reviewer 1SL9 W1)

Additional Experiments:

  1. Added length-normalized mean token entropy and upper-percentile token-entropy statistics. (Reviewer zD3S Q1 and follow-up Q1)
  2. Added comparisons with attention-, position-, and sentence-based context pruning methods. (Reviewer zD3S W3 and Q4)
  3. Compared token-level mask learning with sentence-level mask learning and token-level mask learning followed by sentence voting. (Reviewer 1SL9 Q2)
  4. Added comparisons with parallel sampling, reporting accuracy and end-to-end runtime. (Reviewer zD3S follow-up Q4)
  5. Added automatic hyperparameter selection. (Reviewer 1SL9 W2)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.