acceptodds
Under review as a conference paper at ICLR 2027

Absorbability-Guided CoT Compression for Efficient Reasoning

Abstract

Reasoning models often generate longer chains of thought (CoTs) than a task requires, increasing inference cost. Accuracy-preserving compression requires the current model to absorb the compressed supervision into its reasoning capability. Existing methods choose compression levels from trajectory or task properties without evaluating this model-dependent condition. We introduce **Absorbability-Guided CoT Compression** (), which uses local adaptation to select model-absorbable supervision. For each batch, measures reference-trajectory quality and evaluates candidate token budgets through short SFT updates on a fresh copy of the current model. When reference trajectories are reliable, it selects the shortest budget whose adapted accuracy remains within a tolerance of the reference accuracy. When they are unreliable, it increases the budget until the adapted model reaches an absolute accuracy threshold. Across DeepSeek-R1-Distill-Qwen models at 7B, 14B, and 32B scales on GSM8K, MATH-500, and AIME2024, reduces response length by 19.2%–67.3% while matching or exceeding original-model accuracy in seven of nine model–benchmark settings. These results establish model absorbability as the governing criterion for accuracy-preserving CoT compression: the retained budget matches the amount of supervision the current model absorbs into its reasoning capability.

Primary area:
generative models
Keywords:
efficient reasoning, large language models, chain-of-thought compression

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yuf3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Absorbability-Guided CoT Compression (AGC²), which selects compressed reasoning supervision according to whether the current model can learn from it while preserving accuracy. For each training batch, the method evaluates different token budgets through short local SFT updates and retains the shortest acceptable candidate. Experiments on three model scales and three mathematical benchmarks show substantial reductions in response length while maintaining or improving accuracy in most settings.

Critical strengths
  1. The paper introduces a meaningful model-dependent view of CoT compression. Instead of selecting compression levels only from trajectory length or token importance, AGC² tests whether the current model can actually learn from each compressed candidate through short local adaptation.

  2. AGC² achieves a strong accuracy–efficiency trade-off across 7B, 14B, and 32B models. It reduces response length by 19.2%–67.3% while matching or exceeding the original model’s accuracy in seven of nine model–benchmark settings.

Critical weaknesses
  1. The absorbability score is evaluated on the same batch from which the answer-verified SFT examples are drawn. This partial train-evaluation overlap may favor candidates that fit the current batch rather than measuring generalizable absorbability.

  2. The computational overhead of AGC² is not sufficiently evaluated. For each training batch, the method may construct multiple compressed candidates, perform separate SFT updates from fresh model copies, and evaluate each adapted model on the full batch. However, the paper does not report the total training time, GPU cost, or comparison with the training cost of the baselines.

Questions for authors
  1. Since the answer-verified training subset overlaps with the batch used to evaluate each adapted candidate, how does AGC² distinguish generalizable absorbability from short-term fitting to the current batch? Could the authors evaluate candidate budgets on a disjoint held-out subset?

  2. What is the total training cost of AGC² compared with the baselines, given the repeated compression, adaptation, and evaluation required for each batch?

  3. When the reference batch is unreliable, incorrect trajectories are excluded from SFT and AGC² only lengthens the compressed trajectories from the correct subset. Why should increasing their token budget recover performance on the unanswered examples, and how often does this branch actually find an acceptable candidate?

  4. The empirical evaluation is limited to three datasets and different scales of the same DeepSeek-R1-Distill-Qwen family. GSM8K and MATH-500 may not fully represent challenging reasoning tasks, while AIME2024 contains only 30 problems, making its reported improvements sensitive to a small number of examples. Could the authors evaluate AGC² on larger and more challenging benchmarks, such as OlympiadBench or College Math, and on other model families, such as original Qwen or Llama-based models, to better demonstrate its robustness and generalizability?

3 comments
AuthorsOpenReview ↗

Response to Reviewer yuf3

We thank Reviewer yuf3 for the questions on absorbability, cost, the unreliable branch, and scope.

W1 & Q1: This test assesses the compressed supervision for the current batch, not held-out accuracy.

  • Training/test separation. Compression, SFT, and budget selection use the training split; downstream GSM8K, MATH-500, and AIME2024 test examples are never used.
  • The object being tested. Teacher-forced SFT uses compressed targets for Btref\mathcal{B}_t^{\text{ref}}; evaluation asks the adapted model to answer the full batch Btlong\mathcal{B}_t^{\text{long}} by free generation from the questions alone. The score asks whether those targets are usable after local adaptation.
  • Why not substitute unrelated samples? The goal is to test whether the candidate targets provide absorbable supervision for this batch. Samples not being compressed are not part of this batch’s training data, so their accuracy does not measure whether the model absorbs the candidate supervision; they are unsuitable for selecting a candidate at this step.

W2 & Q2: We measured 7B training costs for AGC2AGC^2 and TokenSkip on the same GPU. We use TokenSkip because its code and configurations are public. The costs are below.

7B methodWall time (h)Allocated GPU-hoursPeak worker allocated memory (GB)
TokenSkip0.8252.13949.49
AGC2AGC^221.38385.53145.48

Peak GPU memory is slightly lower than TokenSkip; the additional training time comes from candidate compression, local adaptation, and in-batch evaluation over up to K=3K=3 candidate budgets per batch. This is a one-time offline cost: at inference, AGC2AGC^2 saves 711 tokens on GSM8K (-67.3%), 1,419 on MATH-500 (-49.7%), and 2,453 on AIME2024 (-23.2%). The extra 83.4 GPU-hours are amortized after roughly 2×1052\times 10^5–4×1054\times 10^5 queries. We acknowledge the higher overhead than single-pass SFT and will explore faster proxy evaluations.

Q3: We thank you for identifying the ambiguity in the unreliable-reference branch. We will correct Sections 4.2–4.3 to distinguish reference accuracy, the supervision subset, and post-adaptation accuracy, and fix the inequality in Figure 2: acceptance requires at,b≥τa_{t,b}\geq\tau; at,b<τa_{t,b}<\tau means the search must continue if trials remain.

The initial at,0a_{t,0} is the correctness rate of the stored references, not model accuracy under a short cap. When at,0<τa_{t,0}<\tau, incorrect references stay excluded from SFT; a larger budget compresses only the correct subset Btref\mathcal{B}_t^{\text{ref}}, not the incorrect examples. Evaluation is still on the full batch Btlong\mathcal{B}_t^{\text{long}}.

Longer targets preserve more reasoning from the correct subset but do not guarantee that the model solves the remaining questions. Our earlier description implied assured recovery and was too strong. The branch accepts a candidate only if full-batch accuracy reaches τ\tau; otherwise, it keeps the pre-batch parameters after at most K=3K=3 trials. Batches are grouped by reference-CoT length, not by a verified difficulty label; we will correct this too.

In the 7B experiment, 78.9% of scored batches enter the reliable branch and 21.1% the unreliable branch; three scheduled batches are skipped without a reference score. Of the unreliable batches, 16.1% find an accepted candidate in 3 trials, while 83.9% do not.

Q4: We appreciate the suggestion to test harder benchmarks and another model family. On OlympiadBench with DeepSeek-R1-Distill-Qwen-7B, baselines are from Appendix Table 6 of Liu et al., ACL 2026.

MethodAccuracy (%)Mean response tokens
Original-7B50.705,213
D-Prompt49.404,676
NoThinking43.401,491
TH2T52.003,122
AGC2AGC^252.974,427

AGC2AGC^2 reduces mean length by 15.08% while increasing accuracy by 2.27% over Original. These results demonstrate the effectiveness of our method on a larger and more challenging benchmark.

We also evaluate DeepSeek-R1-Distill-Llama-8B on GSM8K, MATH-500, and AIME2024:

MethodGSM8K Acc (%) / LenMATH-500 Acc (%) / LenAIME2024 Acc (%) / Len
Original90.30 / 1,63389.40 / 3,94143.33 / 12,137
CoT-Valve87.30 / 1,315—20.00 / 7,411
Ours81.88 / 1,30187.80 / 4,34340.00 / 12,211

Because the TH2T appendix results appear problematic and the appendix lacks generation hyperparameters for DeepSeek-R1-Distill-Llama-8B, we compare with CoT-Valve (Table 3 of Ma et al., ACL 2025). On GSM8K, AGC2AGC^2 shortens responses by 20.34% while accuracy changes from 90.30% to 81.88%. On MATH-500 and AIME2024, CoT-Valve's aggressive compression drops AIME2024 accuracy from 43.33% to 20.00%. In contrast, AGC2AGC^2 detects that this violates local absorbability on these trajectories, so it preserves or slightly expands the token budget and maintains much stronger accuracy. We will include these results in the revision.

Thank you for the detailed response and additional experiments. The reported training cost, unreliable-branch statistics, and expanded evaluations clarify several of my questions. However, I remain concerned about same-batch adaptation and evaluation, the substantial computational overhead, and the mixed cross-family results. I therefore maintain my current rating.

We thank the reviewer for the follow-up and for acknowledging the clarifications on training cost, the unreliable branch, and the expanded evaluations. We address the three remaining concerns directly.

Same-batch adaptation and evaluation. This score is a per-batch absorbability test, not a held-out generalization estimate, and it is used only to select a budget. After the model is updated on the compressed targets for a batch, we evaluate it by free generation from the questions alone, not by teacher-forcing on those targets. The evaluation batch Btlong\mathcal{B}_t^{\text{long}} is also a superset of the SFT subset Btref\mathcal{B}_t^{\text{ref}}: it includes problems whose reference CoTs were excluded. A candidate passes only if the compressed reasoning is learnable within that limited update. The overlap is intentional: the candidate targets are the training data for that batch, so the relevant test is whether the model can absorb them. A disjoint held-out set would instead measure transfer to other problems, which is a different question and is noisy after two local steps. Generalization is checked separately on the held-out test sets; on the Qwen family, the downstream MATH-500 and AIME2024 gains indicate that the per-batch signal is not merely fitting the batch.

Computational overhead. We agree that the measured overhead is substantial relative to static SFT. For 7B model, the end-to-end allocated cost is 21.383 h and 85.531 GPU-hours versus 0.825 h and 2.139 GPU-hours for TokenSkip, an additional 83.4 GPU-hours. These allocated hours include compressor waiting; the measured student SFT and evaluation execution is 13.748 GPU-hours, versus 1.552 for TokenSkip, and most of it (13.256) is in-batch generation for evaluation. The search is bounded: 1.57 candidates and 3.13 local steps per scheduled batch on average. This is a one-time offline cost, not a per-query cost, and peak GPU memory is slightly lower (45.48 vs 49.49 GB). At inference, AGC2AGC^2 reduces response length by 67.3%, 49.7%, and 23.2% on GSM8K, MATH-500, and AIME2024, so the cost is amortized after roughly 2×1052\times10^5–4×1054\times10^5 queries. Static SFT is cheaper but weaker on harder tasks; we will state this trade-off explicitly and explore cheaper proxy evaluation in future works.

Cross-family results. On DeepSeek-R1-Distill-Llama-8B, AGC2AGC^2 is 20.34% shorter on GSM8K but 8.42 points lower; on MATH-500 and AIME2024, it does not shorten responses and remains 1.6 and 3.3 points below Original. The method avoids CoT-Valve's catastrophic AIME collapse (40.00 vs 20.00), but it does not improve over Original on this family. The Qwen results show the intended regime: MATH-500 gains at 7B/14B/32B and AIME2024 gains at 14B/32B with shorter responses. We will supplement the results and discuss the performance degradation in the revise version.

We thank the reviewer again for the continued engagement. These comments have helped us sharpen the absorbability test, the cost analysis, and the cross-family evaluation, and we would be grateful for any further suggestions that can improve the paper.

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

This paper proposes Absorbability-Guided CoT Compression (AGC²), a method for selecting compressed chain-of-thought supervision based on whether the current model can absorb it through local supervised fine-tuning. For each training batch, the method evaluates multiple candidate token budgets using short SFT updates from a common model state. If the reference trajectories are sufficiently accurate, it searches for the shortest budget whose adapted model preserves the reference accuracy. Otherwise, it increases the budget until the adapted model reaches an absolute accuracy threshold. The method is evaluated on DeepSeek-R1-Distill-Qwen models at 7B, 14B, and 32B scales using GSM8K, MATH-500, and AIME2024. The reported results show response-length reductions of 19.2%–67.3%, while matching or exceeding the original model accuracy in seven of nine model–benchmark settings. The paper also provides sensitivity analyses for the reference-quality threshold, accuracy tolerance, search patience, update strategy, and compressor choice.

Critical strengths
  1. The paper distinguishes between a compressed trajectory that is merely answer-correct and one that can be effectively learned by the current model. This is a useful conceptual refinement over methods that select compression levels only from trajectory statistics or fixed heuristics.
  2. The branch-specific search rules, independent candidate updates, and sensitivity analyses make the method relatively easy to understand and reproduce. The reported results also show that AGC² can achieve a favorable accuracy–length trade-off on several settings, especially MATH-500 and AIME2024.
Critical weaknesses
  1. The paper reports substantial CoT length reductions, but does not analyze what reasoning content is removed during compression. Can the authors provide a quantitative and qualitative analysis of the original versus compressed trajectories—for example, whether the method removes redundant explanations, intermediate calculations, verification steps, or answer-critical reasoning—and explain how the compressed trajectories preserve the information necessary for solving the problem?
  2. Can the authors compare AGC² against a simpler baseline that uses GPT-compressed CoTs with a fixed budget and standard SFT? This comparison is necessary to determine whether the additional in-training budget search provides benefits beyond directly distilling shorter trajectories from GPT.
  3. What are the total training costs of AGC² compared with the baselines, including the number of local SFT updates, additional forward passes, wall-clock time, and GPU memory? Under what number of downstream queries does the reduced inference cost offset the additional training cost?
  4. Could training with AGC² induce a globally shorter reasoning policy that harms OOD or harder tasks requiring longer chains of thought? Please evaluate length distributions and accuracy on out-of-domain, difficulty-controlled, and long-horizon reasoning benchmarks.
Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer dzVj

We thank Reviewer dzVj for the questions about retained reasoning content, fixed-budget distillation, cost, and generalization.

W1: AGC2AGC^2 uses post-adaptation answer accuracy to test compressed supervision, rather than a fixed linguistic rule for deciding which reasoning step is removable. Figure 1 shows that, when the answer-derivation step is removed from heuristically compressed CoTs, the original models' completion accuracy decreases by 6.7–36.7 percentage points. This motivates testing compressed targets with the model that will learn from them.

We cannot determine the exact minimal information needed for a correct solution without a rigorous proof. We therefore test how learning from each candidate affects the current model's answer accuracy. In the reliable branch, a candidate is accepted when post-adaptation accuracy meets at,b≥at,0−Δa_{t,b}\geq a_{t,0}-\Delta; otherwise the search does not retain that candidate. In the unreliable branch, acceptance instead requires at,b≥τa_{t,b}\geq\tau.

Case study. In a GSM8K training example, Ali and Ernie have 80 boxes; Ali makes five circles using eight boxes each, and Ernie needs ten boxes per circle. Both the reference and compressed trajectory answer 4.

The original trajectory calculates the answer and then repeats the complete calculation:

Starting from the beginning: 80 boxes. Ali uses 5 circles * 8 boxes = 40 boxes. 80 - 40 = 40 boxes left. Ernie uses 10 boxes per circle, so 40 / 10 = 4 circles. Yeah, that seems right.

The compressed trajectory removes this repeated pass and a later hypothetical discussion about floor division, while retaining the three calculations needed for the solution (verbatim excerpts):

Ali's total boxes = 5 circles * 8 boxes/circle = 40 boxes.

Remaining boxes = Total boxes - Ali's boxes = 80 - 40 = 40 boxes.

Ernie's circles = Remaining boxes / Boxes per circle = 40 / 10 = 4.

It also retains a short 40+40=8040+40=80 verification. The stored texts contain 650 and 337 tokens under the same student tokenizer (48.15% shorter). The example illustrates removal of duplicated explanation while preserving the calculation chain.

W2: We appreciate the suggestion to use fixed-budget GPT compression followed by standard SFT to assess the contribution of budget search. Figure 6 contains our fixed-round ablation: it uses the same compressor but applies three shortening rounds without the absorbability-based stopping rule. On MATH-500 and AIME2024, it produces shorter responses with lower accuracy than the adaptive search.

W3: We measured the training costs of AGC2AGC^2 and TokenSkip for the 7B model on the same GPU hardware. We selected TokenSkip as a comparison baseline because its code and training configurations are publicly available. The measured costs are shown below.

7B methodWall time (h)Allocated GPU-hoursPeak worker allocated memory (GB)
TokenSkip0.8252.13949.49
AGC2AGC^221.38385.53145.48

Peak GPU memory is slightly lower for AGC2AGC^2, while the additional training time comes from candidate compression, local adaptation, and in-batch autoregressive evaluation across up to K=3K=3 candidate budgets per batch. However, this is a one-time offline training cost: at inference time, AGC2AGC^2 reduces average response length by 711 tokens on GSM8K (-67.3%), 1,419 tokens on MATH-500 (-49.7%), and 2,453 tokens on AIME2024 (-23.2%). Under standard GPU decoding throughput, the additional 83.4 GPU-hours of training are amortized after roughly 2×1052\times 10^5–4×1054\times 10^5 downstream queries. We acknowledge that our current training overhead is substantially higher than single-pass static SFT, and we will discuss this trade-off and explore faster proxy evaluations in future work.

W4: The training corpus contains GSM8K and MATH training problems; AIME2024 is held out from this adaptation corpus.

1. AIME2024 in the main results. The trained models still generate 8,117, 6,292, and 7,760 tokens on average at 7B, 14B, and 32B, rather than reducing all tasks to the few hundred tokens observed on GSM8K. Reported accuracies compared with Original show that the models can still produce long responses on a harder held-out benchmark, although AIME2024 has only 30 questions.

2. Additional 7B OlympiadBench evaluation. We report the results of baselines and AGC2AGC^2 on OlympiadBench with DeepSeek-R1-Distill-Qwen-7B. Baseline results are from the Appendix Table 6 of Liu et al., ACL 2026.

MethodAccuracy (%)Mean response tokens
Original-7B50.705,213
D-Prompt49.404,676
NoThinking43.401,491
TH2T52.003,122
AGC2AGC^252.974,427

For AGC2AGC^2 vs. the original model, mean length decreases by 15.08%, while accuracy increases by 2.27%. These results demonstrate the effectiveness of our method on a larger and more challenging benchmark.

Dear Reviewer dzVj,

We hope this message finds you well.

Once again, we sincerely appreciate your valuable time and insightful feedback.

We are reaching out to kindly remind you that, based on your suggestions, we have included additional experiments and discussions in our response. We would sincerely appreciate it if you could take valuable time out of your busy schedule to review our response at your earliest convenience. Your timely feedback will help us confirm whether we have appropriately addressed the issues and suggestions you raised.

Furthermore, your insights and expertise are invaluable to us. Please let us know if there is anything that can further improve the quality of the paper. We are pleased to address any remaining issues.

Best regards, Authors of Submission 42317

Dear Reviewer dzVj,

We hope this message finds you well. We are writing to kindly remind you that the discussion phase is drawing to a close on Nov. 18 (AoE).

We would sincerely appreciate it if you could take valuable time out of your busy schedule to review our response. Your timely feedback will help us confirm whether we have appropriately addressed the issues and suggestions you raised.

We would greatly appreciate any updates or further feedback from you. Your insights and expertise are invaluable to us, as we aim to address any concerns and improve our submission.

Thank you for your time and attention. We look forward to hearing from you soon.

Best regards, Authors of Submission 42317

Thank you for the detailed response and additional experiments. The case study and OlympiadBench results partially address my concerns about compression behavior and generalization. However, the fixed-round ablation does not fully establish the advantage over fixed-budget GPT-compressed CoTs with standard SFT. The substantial training overhead and the inconsistent baselines used in the break-even estimate also leave the practical efficiency benefits unclear. Overall, I appreciate the additional evidence, but my assessment remains borderline, leaning toward rejection. I therefore maintain my current rating.

Reviewer_BTy8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes a method of CoT compression while trying to preserve better accuracy. In essence, they check if a current model absorbs compression and adaptively increases the budget for length of the CoT in cases where it does not. They paper claims they are better in 6 of 9 test cases. However, I could only see them being better in 5 and worse in 4. The model does not take the least tokens in essence by design trying to achieve higher accuracy by allowing for longer CoTs in persumably the harder cases.

Critical strengths
  1. The method seems to be better in some cases on real benchmarks and LLMs.

  2. The idea has some small novelty in that it has not been used before and adaptively considering the length/budget of CoT, when to compress more and when less during SFT is interesting. They show that is works in some cases.

Critical weaknesses
  1. The experiments are run using only one LLM of one family. So, we do not know if this works only for this family or beyond.

  2. The LLMs used are small in size relatively speaking. The SOTA for open-weight LLMs is much larger. Does the same thing hold there? Or for that matter in any other LLM? We do not know.

  3. The experiments are also not that conclusive. If you are given Table 1, the comparison between THT2 and this method is very close. I am not sure one is superior to the other or this is a result of some randomness and small sample size.

  4. A whole lot of graphs have 3 data points, etc. and if you look at them some of them go up and down. It is not clear if we can draw any conclusion from them. Is this random variation or statistically siginifcant? There are no variances shown - for Figures 3 and 4, these are perhaps point data and drawing a line between them is weird. We do not know what the results can be in between what has been tested. It would be better to show these as box plots with the mean and the variance to show how varied the results are. That would give me a better idea of what to trust. Anyway, a lot more data points and statistical significant tests would be needed to have anything meaningful.

  5. The main algorithm and some interesting results are in the Appendix. The conference requires the main paper to be complete standalone.

  6. The paper is badly written. There is unnecessary complexity using mathematical terms for very simple equations. I like formal explanations and statements. But, things like accuracy can be said in words and would actually be more understandable if it was said in word briefly perhaps with a brief example.

  7. The paper would also be more understandable if a small example would be worked out.

  8. The paper is not reproducible in its current form because things are gone into the Appendix - again which we are not required to read.

Questions for authors

Address the weaknesses please.

3 comments
AuthorsOpenReview ↗

Response to Reviewer BTy8

We thank Reviewer BTy8 for the feedback.

W1 & W2: We appreciate your request to evaluate generalization beyond a single model family. For a fair and consistent comparison across baselines, our original main experiments used DeepSeek-R1-Distill-Qwen-7B, -14B, and -32B. We have now added an evaluation of DeepSeek-R1-Distill-Llama-8B on GSM8K, MATH-500, and AIME2024:

MethodGSM8K Acc (%) / LenMATH-500 Acc (%) / LenAIME2024 Acc (%) / Len
Original90.30 / 1,63389.40 / 3,94143.33 / 12,137
CoT-Valve87.30 / 1,315—20.00 / 7,411
Ours81.88 / 1,30187.80 / 4,34340.00 / 12,211

Because the TH2T appendix results appear problematic and lack generation hyperparameters for DeepSeek-R1-Distill-Llama-8B, we compare with CoT-Valve's reported results (Table 3 of Ma et al., ACL 2025). On GSM8K, AGC2AGC^2 shortens responses by 20.34%, with accuracy changing from 90.30% to 81.88%. On MATH-500 and AIME2024, CoT-Valve's aggressive compression drops AIME2024 accuracy from 43.33% to 20.00%. In contrast, AGC2AGC^2 detects that aggressive shortening violates the local absorbability criterion on these trajectories, so it preserves or slightly expands the token budget and maintains much stronger accuracy. We will include these results and discussions in the revised manuscript.

W3 & W4: We address both concerns together.

Main-table comparison: Our objective is to select shorter supervision that satisfies the current model's local absorbability criterion, rather than maximize compression alone. On MATH-500, the reported AGC2AGC^2 accuracies are 88.4%, 87.6%, and 89.0%, versus 86.8%, 86.4%, and 88.0% for TH2T at 7B, 14B, and 32B.

Multi-seed sensitivity evaluation: We evaluated all five distinct configurations in Figures 3 and 4: τ∈{.90,.92,.95}\tau\in\{.90,.92,.95\} at Δ=0,h=1\Delta=0,h=1, and Δ∈{0,.02,.05}\Delta\in\{0,.02,.05\} at τ=.90,h=1\tau=.90,h=1, using five seeds. Both are reported as mean ± sample standard deviation; reduction is against Original. The results across different seeds show a task-dependent accuracy–length trade-off across the tested settings.

SettingGSM8K Acc (%)Reduction (%)MATH-500 Acc (%)Reduction (%)AIME2024 Acc (%)Reduction (%)
Original91.0 ± 0.37—85.2 ± 0.48—50.0 ± 5.27—
τ=.90,Δ=0\tau=.90,\Delta=090.3 ± 0.2567.03 ± 2.3290.0 ± 0.5449.24 ± 3.2152.0 ± 6.9110.96 ± 13.11
τ=.92,Δ=0\tau=.92,\Delta=087.9 ± 0.4362.99 ± 3.5185.2 ± 0.4228.27 ± 3.0946.7 ± 2.792.97 ± 5.88
τ=.95,Δ=0\tau=.95,\Delta=088.7 ± 0.1760.84 ± 4.4088.0 ± 0.7728.86 ± 2.9250.0 ± 4.085.84 ± 6.45
τ=.90,Δ=.02\tau=.90,\Delta=.0288.6 ± 0.4254.98 ± 3.8190.2 ± 0.4831.72 ± 2.1948.7 ± 3.801.22 ± 13.33
τ=.90,Δ=.05\tau=.90,\Delta=.0590.4 ± 0.3367.89 ± 0.6587.5 ± 1.2264.13 ± 1.7628.7 ± 6.0651.12 ± 9.81

Paired accuracy analysis: We ran exploratory two-sided exact McNemar tests on per-question outcomes with Holm correction. For default (τ=.90,Δ=0\tau=.90,\Delta=0) vs. Original, adjusted p-values are 2.80×10−92.80\times10^{-9} (GSM8K), 0.00269 (MATH-500), and 1.0 (AIME2024). The differences are statistically significant on GSM8K and MATH-500, but not on AIME2024.

W5 & W8: We appreciate your emphasis on making the main text self-contained and the method reproducible. We will improve the standalone presentation. The final version allows one extra page; if accepted, we will use it to move Algorithm 1 and more results from the appendix into the main text. We will also release the training and evaluation code to support reproducibility.

W6 & W7: We appreciate your suggestions to explain the notation in plain language and illustrate how correctness is determined. The notation distinguishes three quantities:

  • Reference trajectory accuracy (at,0a_{t,0}, Eq. 3): the fraction of problems in Btlong\mathcal{B}_t^{\text{long}} whose original reference CoT yields the correct final answer.
  • Answer-verified sub-batch (Btref\mathcal{B}_t^{\text{ref}}, Eq. 4): the reference trajectories with correct extracted answers, used to form the SFT targets.
  • Post-adaptation absorbability (at,ba_{t,b}, Eq. 7): the locally updated model's answer accuracy on the full batch Btlong\mathcal{B}_t^{\text{long}}.

In the revised manuscript, we will simplify the content of redundant equations and add formal explanations.

We will also explain how correctness is checked. For a reference CoT, we extract the final answer and compare it with the ground-truth answer, using numerical or symbolic equivalence where appropriate. A match contributes 1 to the sum in Eq. (3), and that reference enters the answer-verified subset in Eq. (4). After local adaptation, the model answers from the question alone; we apply the same procedure to compute Eq. (7). For example, an extracted 4\boxed{4} with ground truth 4 is correct, whereas an extracted 3 is not. If three of four problems are correct, the batch accuracy is 3/4=0.753/4=0.75.

Dear Reviewer BTy8,

We hope this message finds you well.

Once again, we sincerely appreciate your valuable time and insightful feedback.

We are reaching out to kindly remind you that, based on your suggestions, we have included additional experiments and discussions in our response. We would sincerely appreciate it if you could take valuable time out of your busy schedule to review our response at your earliest convenience. Your timely feedback will help us confirm whether we have appropriately addressed the issues and suggestions you raised.

Furthermore, your insights and expertise are invaluable to us. Please let us know if there is anything that can further improve the quality of the paper. We are pleased to address any remaining issues.

Best regards, Authors of Submission 42317

Dear Reviewer BTy8,

We hope this message finds you well. We are writing to kindly remind you that the discussion phase is drawing to a close on Nov. 18 (AoE).

We would sincerely appreciate it if you could take valuable time out of your busy schedule to review our response. Your timely feedback will help us confirm whether we have appropriately addressed the issues and suggestions you raised.

We would greatly appreciate any updates or further feedback from you. Your insights and expertise are invaluable to us, as we aim to address any concerns and improve our submission.

Thank you for your time and attention. We look forward to hearing from you soon.

Best regards, Authors of Submission 42317

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.