acceptodds
Under review as a conference paper at ICLR 2027

SCRATCH: Efficient Math Inference for Large Reasoning Models

Abstract

Extended reasoning traces can contain repeated calculations, explanations, and abandoned branches that increase output length without contributing to the final answer. Inspired by human scratch work, we propose SCRATCH, a curriculum fine-tuning framework for concise mathematical reasoning. Scratch Trace Construction (STC) selects intact units containing computations, conditions, and justifications from native traces with verified final answers. Progressive Compression Curriculum (PCC) progressively shifts training exposure from less compressed to more compressed targets. Across three models and three mathematical benchmarks, SCRATCH reduces generated tokens by 24.87% on average, with a mean accuracy change of percentage points relative to the unadapted models. Peak savings reach 37.71% on MATH-500 with DeepSeek-1.5B. Against five efficient reasoning methods, it achieves the highest accuracy in eight of nine settings and the shortest outputs in eight. Controlled ablations show that progressive ordering improves accuracy and shortens outputs across all three mathematical benchmarks when training targets, effective batches, cumulative sample exposure, and the final consolidation stage are held fixed. Without additional task-specific training, SCRATCH reduces generated tokens on the transfer benchmarks by 34.52% on average, with a mean accuracy change of percentage points.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Efficient Reasoning, Overthinking, PEFT

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

This paper studies redundancy in long CoT traces generated by reasoning models and proposes SCRATCH, a two-stage SFT framework for learning to produce shorter reasoning traces. In the first stage, Scratch Trace Construction (STC), the target model generates solutions to training problems, and only responses with verified correct final answers are retained. Each accepted CoT is divided into units at paragraph and numbered-step boundaries. GPT-4o-mini receives the question, the verified answer, and all numbered units, and selects those intended to preserve the key calculations, conditions, and logical connections while removing redundant explanations and unnecessary branches. In the second stage, Progressive Compression Curriculum (PCC), the model is fine-tuned on the compressed traces, starting with less-compressed examples, progressively introducing more-compressed ones, and ending with a consolidation stage focused on highly compressed examples. Across three models and three mathematical benchmarks, the authors report an average reduction of 24.87% in generated tokens and an average accuracy change of −0.36 percentage points relative to the original models.

Critical strengths

S1. SCRATCH reduces output length across multiple models and math benchmarks with small average accuracy changes. The shorter-generation behavior also transfers to non-math tasks without additional fine-tuning.

S2. Configurations E and F hold training targets, batch contents, per-example exposure, and the final consolidation stage fixed, changing only the order of earlier batches. This provides a relatively clean test of progressive ordering.

S3. SCRATCH builds compressed supervision offline and uses standard LoRA SFT. Inference requires no architectural changes or additional compression model, simplifying integration into existing pipelines.

Critical weaknesses

W1. The paper reports mean results from five decoding samples per question but provides no standard errors or confidence intervals. Given the modest gains in the PCC ordering ablation, paired confidence intervals or bootstrap estimates based on the existing evaluation outputs would help assess the reliability of these differences without additional model training.

W2. A straightforward alternative is to sample multiple solutions per problem, select the shortest correct response, and use it for SFT, as explored by Munkhbat et al. (2025). The paper does not compare STC with this baseline, leaving its incremental benefit unclear.

W3. SPIRIT (Cui et al., 2025) already prunes reasoning steps to construct concise supervision, while Train Long, Think Short (Hammoud et al., 2025) uses a curriculum to encourage progressively shorter reasoning. Although their selection criteria and training algorithms differ from SCRATCH, the paper should clarify what is novel beyond combining these established ideas and why its specific design matters.

Questions for authors

Q1. The original MATH dataset (2021) contains historical AIME problems, whereas evaluation covers AIME 2020–2025. Have the authors checked the 1,000 selected MATH training problems for exact or near-duplicate overlaps with evaluation questions, particularly AIME 2020? Please report the results.

Q2. Steps unnecessary for a particular correct solution may still be useful for learning, such as attempts that illustrate errors or revisions. How do the authors view this potential difference between what is needed for the final answer and what may benefit the model during training?

Q3. Could removing unsuccessful attempts reduce strategy diversity or backtracking ability? Pass@5 measures solution coverage rather than path diversity. Could the authors analyze alternative solution strategies and error recovery before and after training?

Q4. The reported mean token length includes both correct and incorrect responses. Could the authors report token lengths and reductions separately for these two groups, to clarify whether the overall savings come mainly from correct or incorrect responses?

3 comments

Response 1 of 2

We appreciate your recognition of SCRATCH's efficiency, controlled ordering comparison, and simple deployment. Your questions about its contribution beyond prior work and the learning value of intermediate attempts helped us focus the additional experiments. We have added the suggested comparisons and looked more closely at which reasoning units are retained, as well as strategy diversity and error recovery after training.

W1. Stability across five evaluation runs

Each cell gives Baseline / SCRATCH over the five evaluation runs in Table 1. Accuracy standard deviation (SD) is expressed in percentage points. For tokens, relative SD (RSD) is the SD of the five run-level means divided by that method's own mean, expressed as a percentage.

Accuracy SD (percentage points).

ModelGSM8KMATH-500AIME 2020–2025
DeepSeek-1.5B0.42 / 1.280.65 / 1.722.12 / 2.29
DeepScaleR-1.5B0.62 / 0.711.13 / 1.002.81 / 1.71
Qwen3-14B0.34 / 0.280.71 / 0.951.65 / 2.55

Token RSD (%).

ModelGSM8KMATH-500AIME 2020–2025
DeepSeek-1.5B2.28 / 1.872.03 / 4.220.55 / 0.74
DeepScaleR-1.5B1.29 / 1.641.82 / 1.061.32 / 2.93
Qwen3-14B0.71 / 1.091.34 / 1.760.67 / 1.56

Across all nine settings, every SCRATCH run uses fewer tokens on average than every Baseline run.

We also assessed training variation with three independent trainings per configuration on DeepSeek-1.5B. Averaging the trainings and benchmarks equally, Full SCRATCH has 0.84 points higher accuracy and 4.43% fewer tokens than STC + prefix shuffle + consolidation, and 18.71% fewer tokens than STC + Standard SFT with a −0.08-point accuracy change. On DeepScaleR-1.5B and Qwen3-14B, it is more accurate and shorter than the prefix-shuffle control in all six settings. These results agree with PCC's intended role: stronger supervision for compact reasoning, with accuracy largely preserved.

W2/W3. Other methods

We compared SCRATCH with shortest-correct SFT (Munkhbat et al., 2025) and adaptations of SPIRIT-FT (Cui et al., 2025) and TLTS (Hammoud et al., 2025) on DeepSeek-1.5B, given the compute needed for GRPO. All methods use the same 1,800-question pool, evaluation settings, and output caps. Optimizer updates are matched, while TLTS also generates responses during online reinforcement learning. Each row is one trained model evaluated with five answers per question. Cells give accuracy (%) / mean tokens.

MethodGSM8KMATH-500AIME 2020–2025Key distinction
Shortest-correct SFT (best of 16)74.24 / 457.279.32 / 3507.623.67 / 11955.3Shortest whole correct sample
SPIRIT-FT adaptation72.84 / 396.778.20 / 2988.923.44 / 11318.4Perplexity-guided removal; LLM merging
TLTS adaptation73.24 / 396.679.32 / 2867.323.78 / 11257.5Online reinforcement learning; absolute budget
STC + Standard SFT74.83 / 459.680.28 / 3712.424.22 / 12102.5Intact reasoning units, uniform SFT
Full SCRATCH74.04 / 377.581.00 / 2688.826.67 / 10986.6Intact reasoning units, curriculum by relative compression

Full SCRATCH is shortest on all three benchmarks, more accurate than both adaptations throughout, and most accurate on MATH-500 and AIME. Our shortest-correct control is the Naive BoN variant of Munkhbat et al. (2025): it keeps the shortest verified-correct response among 16 samples, covering 1,769 of 1,800 questions. Their full recipe also adds few-shot conditioning and sample augmentation. Whole-response selection leaves any redundancy inside the chosen answer; STC selects within a response. Under the same Standard SFT, STC targets are more accurate on all three benchmarks, with somewhat longer outputs. With PCC, Full SCRATCH averages 1.49 points higher accuracy and 16.29% fewer tokens than the shortest-correct SFT.

What distinguishes SCRATCH. SPIRIT-FT uses the target model's perplexity to select removals and merges steps with an LLM when needed. STC instead selects computations, conditions, and dependencies while preserving the retained units verbatim. A matched control uses GPT-4o-mini to rewrite the same native traces concisely. STC + Standard SFT is 2.39 points more accurate on average than Direct Rewrite + Standard SFT. This suggests a benefit from keeping the model's own expressions, rather than constructing new ones.

TLTS applies a decreasing absolute token-budget target during online GRPO, with rewards for correctness, length, and format. PCC places the curriculum in offline supervised training: targets stay fixed, while their frequency and ordering change according to relative compression, followed by consolidation. SCRATCH pairs verbatim native-unit selection with this offline curriculum. The comparisons and ablations support both choices. In the final version, we will clarify these distinctions in Related Work.

Response 2 of 2

Q1. Overlap with AIME

We verified that the entire selected pool of 2,000 questions used for training and development (1,000 GSM8K and 1,000 MATH) has no overlap with the 180 AIME 2020–2025 evaluation problems, including all 30 AIME 2020 problems.

Q2. What the answer needs versus what helps learning

Thank you for highlighting the learning value of errors and revisions. We distinguish a complete response that ends in a wrong answer from useful unsuccessful attempts within a verified-correct response. SCRATCH excludes the former from positive SFT targets. Using such responses as rejected examples in direct preference optimization (DPO) is a direction for future work.

Within a verified-correct response, STC asks the annotator to retain unsuccessful attempts and corrections that supply useful assumptions, conditions or information needed later. In an expert audit of 50 native traces across the three backbones, the default annotator retained 60 of 62 useful error-correction units and 157 of 160 critical-condition units. Q3 examines the related question directly: whether the trained model still finds different correct strategies and recovers from errors.

Q3. Strategy diversity and error recovery

We tested two abilities on 50 questions (25 from MATH and 25 from AIME). Question selection did not depend on either model's outputs, and evaluators were not told which model produced each answer.

  1. Diversity of correct strategies. For each question, we examined five answers from each model and grouped the correct ones by mathematical approach. Rewordings of the same approach count as one strategy. The table reports the number of questions with at least two distinct correct strategies.
  2. Recovery from errors. For each question, we prepared two partial solutions that contain a recoverable error, for 100 in total. Baseline and SCRATCH received the same 100 starts. A recovery counts as successful only if the model corrects the error, completes a valid derivation, and reaches the correct answer.

Table entries show Baseline → SCRATCH.

ModelQuestions with ≥2 correct strategies (of 50)Successful recoveries (of 100)
DeepSeek-1.5B11 → 1061 → 60
DeepScaleR-1.5B17 → 1669 → 66
Qwen3-14B30 → 2983 → 82

On this sample, the counts remain close to Baseline, differing by one multi-strategy question and one to three successful recoveries per model.

Q4. Correct and incorrect response lengths

Both correct and incorrect responses become shorter in all nine model–benchmark settings, with reductions for correct responses ranging from 14.50% to 36.35%. DeepSeek-1.5B on MATH-500 has the largest reduction for correct responses:

Response groupBaseline tokensSCRATCH tokensReduction
Correct3582.52280.136.35%
Incorrect7493.94431.140.87%

Averaged over the nine settings, shorter correct responses account for about 65% of the overall token saving. Correct responses contribute more in seven of the nine settings; the exceptions are the two 1.5B models on AIME.

The new comparisons help explain how intact reasoning units and PCC contribute to SCRATCH. Your questions also led us to examine useful corrections and the model's behavior after compression, beyond average accuracy and length. Please let me know if any additional experiments or analyses might help you reconsider our manuscript.

I appreciate the authors’ response, which addresses some of my concerns, and I am maintaining my original rating.

Reviewer_htY3OpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes SCRATCH, a curriculum fine‑tuning framework for concise mathematical reasoning to mitigate LLM over‑thinking. It includes two components: STC selects complete reasoning units from verified long reasoning traces to build compact supervision targets; PCC trains the model progressively from low‑ to high‑compression samples with a final consolidation phase. Evaluations on three models across math benchmarks show SCRATCH reduces generated tokens by 24.87% on average with negligible accuracy loss. It also achieves token compression on out‑of‑domain reasoning tasks without extra task‑specific training. Controlled ablations verify the efficacy of both STC and PCC.

Critical strengths
  1. This paper addresses the real problem of reasoning overthinking and reduces inference cost by retaining essential logical steps instead of naive token truncation.
  2. The authors conduct extensive comparisons against five efficient reasoning baselines on both mathematical and cross‑domain tasks, and report diverse metrics including accuracy, token reduction and pass@5.
Critical weaknesses
  1. The STC module heavily depends on verified ground‑truth answers and an external annotation LLM, and the paper does not evaluate its performance under noisy verification or weak annotators.
  2. The PCC module adopts fixed compression grouping ratio and exposure weights, and the paper does not study the hyperparameter sensitivity of these settings.
  3. SCRATCH causes accuracy degradation on cross‑domain tasks, especially for small‑size models on BBH, and the paper does not diagnose the root cause of this phenomenon.
  4. All experiments are only conducted on models no larger than 14B, and the paper’s claim of compatibility for bigger models lacks empirical evidence.
Questions for authors
  1. How would a weaker annotation model affect STC scratch trace quality and final performance?
  2. Could you provide small‑scale noise injection experiments to test STC under imperfect answer verification?
  3. Why does compression hurt non‑mathematical reasoning more on transfer benchmarks?
1 comment

Thank you for your encouraging assessment. Your suggestions guided the additional robustness experiments and transfer analyses below. Unless stated otherwise, we ran them on DeepSeek-1.5B, and each trained model is evaluated with five answers per question.

Weaker annotators (W1/Q1)

To test a weaker annotator, we replaced GPT-4o-mini with Qwen3-4B for unit selection on all 1,800 training traces, keeping the native traces, unit segmentation, and curriculum rule unchanged. Accuracy changed by −0.53, −0.68, and −1.11 percentage points on GSM8K, MATH-500, and AIME, relative to default SCRATCH.

To see how the selected units differ, we compared four annotators with an expert reference on 50 native traces (25 GSM8K, 25 MATH) spanning the three backbones. Two experts marked the essential units independently, and a third resolved disagreements. F1 balances an annotator's missed necessary units against its unnecessarily selected ones.

AnnotatorF1 against expert selection (%)
GPT-4o-mini (default)96.19
Qwen3-32B96.44
DeepSeek-V4.1-Flash-101796.66
Qwen3-4B92.96

GPT-4o-mini is close to the two higher-scoring annotators, and the 4B model's lower match accompanies the accuracy loss above.

Imperfect verification and curriculum settings (W1/Q2, W2)

To simulate a verifier that accepts wrong solutions, we replaced 5% or 10% of training examples with an erroneous solution and its wrong answer. That is 90 or 180 of the 1,800 examples, split equally between GSM8K and MATH.

The Progressive Compression Curriculum (PCC) ranks targets by relative compression, the fraction of native reasoning tokens removed, into Low, Medium, and High groups. It gives the more compressed groups more weight across the Early, Middle and Later stages, before a final consolidation stage. Compression group proportions are the shares of training pairs in each group; Later-stage exposure weights set how often each group is sampled in the Later stage. We changed each separately, with everything else fixed. Values are accuracy changes in percentage points relative to default SCRATCH; proportions and weights are listed as Low/Medium/High.

SettingGSM8KMATH-500AIME
5% erroneous examples−0.67−0.96−0.89
10% erroneous examples−1.45−2.08−2.00
Compression group proportions 60/30/10% → 50/35/15%−0.15−0.24−0.34
Later-stage exposure weights 40/40/20% → 50/35/15%+0.06+0.16+0.11

Accuracy falls by at most 0.96 points with 5% erroneous examples and 2.08 points with 10%. The curriculum changes move accuracy by −0.34 to +0.16 points and mainly affect length. Larger Medium and High groups shorten outputs by 0.46–1.92%. Sampling the Low group more often in the Later stage lengthens them by 1.92–4.23%.

BBH accuracy for small models (W3/Q3)

BBH is a collection of reasoning tasks beyond mathematics, and our score gives equal weight to its 27 task types. To locate the losses, we examined six task types with 100 questions each. For both 1.5B models, the largest drops are in logical deduction and state tracking (3.4–4.4 percentage points). These tasks require carrying intermediate facts forward, such as who holds an object after several exchanges. Qwen3-14B loses 0.6–0.8 points on these tasks. On the two 1.5B models, Boolean expressions, dates, and sorting decline less.

We then read 50 outputs per small model in which the original model was correct, and SCRATCH was not. An omitted intermediate state or condition was involved in 33 DeepSeek-1.5B cases (66%) and 36 DeepScaleR-1.5B cases (72%). This suggests that the brevity learned on mathematics carries over to these tasks, and that small models sometimes drop a state or condition their later reasoning still needs. The smaller losses at 14B are consistent with smaller models relying more on writing out intermediate states.

Larger models (W4)

Our mathematics experiments on three backbones at 1.5B and 14B support the same model-agnostic STC/PCC procedure: select intact units from each model's own verified traces, then schedule those fixed targets by relative compression. Neither step depends on a particular architecture or parameter count. We therefore expect this pipeline to extend beyond 14B, including closed-weight models that provide the required reasoning traces and allow post-training with the curriculum's sampling and ordering controls. Due to limitations in computational resources, we were unable to carry out verification, but we consider this an important direction for future work.

We hope these additional experiments and analyses answer your questions, and would be glad to discuss any points you would like us to explore further.

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

This paper proposes SCRATCH, a framework for improving the inference efficiency of mathematical reasoning models while largely preserving answer accuracy. It consists of two components: 1.Scratch Trace Construction (STC): Uses an annotation model to select essential reasoning units from verified model-generated solutions, preserving computations, conditions, and necessary justifications while removing redundant reasoning. 2.Progressive Compression Curriculum (PCC): Organizes training examples by their relative compression ratios and progressively increases exposure to highly compressed reasoning traces, followed by a consolidation stage. Experiments on three models (DeepSeek-1.5B, DeepScaleR-1.5B, and Qwen3-14B) across GSM8K, MATH-500, and AIME show an average 24.87% reduction in generated tokens, with only a 0.36 percentage-point decrease in accuracy. SCRATCH also achieves 34.52% average token reduction on three transfer benchmarks. Controlled ablations demonstrate the benefits of reasoning-unit selection and curriculum ordering.

Critical strengths
  1. Strong empirical improvements in the accuracy–efficiency trade-off. SCRATCH consistently reduces reasoning length across different model families and mathematical benchmarks, outperforming five competing efficient-reasoning methods in accuracy and token efficiency in most settings. The additional evaluation on CommonsenseQA, ARC-Challenge, and BBH demonstrates that the learned compression behavior transfers beyond mathematical reasoning. These results support the practical relevance and generality of the proposed approach.

  2. Well-controlled ablations supporting the curriculum design. The paper carefully distinguishes the effects of target construction, sample exposure, consolidation, and training order. In particular, the comparison between shuffled and progressively ordered training preserves effective batches, cumulative sample exposure, and the final consolidation stage. This provides meaningful evidence that curriculum ordering contributes improvements beyond simply increasing exposure to compressed examples.

Critical weaknesses
  1. Limited evidence for the robustness and generality of the curriculum's benefits. Although PCC is a central methodological contribution, the controlled ablations are conducted only on DeepSeek-1.5B. Moreover, the experiments do not report results across independently trained checkpoints or confidence intervals. The observed accuracy improvements from progressive ordering are relatively modest (0.73–1.67 percentage points), making it difficult to determine whether these benefits are robust to training randomness. The five decoding seeds used for evaluation do not address this concern.

  2. Insufficiently controlled comparisons with existing compression methods. The competing training-based methods use their original datasets and optimization procedures rather than matched training budgets. Consequently, differences in training data, supervision quality, and optimization may contribute to the reported performance gaps. Furthermore, the STC ablation compares against regex-based extraction, which is a relatively weak alternative to LLM-guided unit selection. A stronger compression baseline using LLM-based rewriting or summarization under a comparable budget would help isolate the specific benefit of preserving intact reasoning units.

Questions for authors
  1. Robustness of PCC: Could the authors report results from multiple independent training runs for the controlled curriculum ablations, ideally extending the comparison to DeepScaleR-1.5B or Qwen3-14B? This would help establish whether the observed ordering improvements are statistically reliable and generalize across model scales.

  2. Effectiveness of STC: How does STC compare with LLM-based reasoning-trace rewriting or summarization when using the same annotation model, training examples, and optimization budget? This comparison would help determine whether selecting intact reasoning units offers a specific advantage over alternative approaches to constructing concise supervision.

  3. Practical efficiency: Could the authors provide wall-clock inference latency measurements and an estimate of the offline annotation and training costs? Since the primary motivation is inference efficiency, understanding how token reductions translate into actual computational savings would strengthen the practical significance of the results.

2 comments

Response 1 of 2

Thank you for your thoughtful review and for recognizing SCRATCH's accuracy–efficiency gains and the controlled curriculum ablations. We share your view that these gains should be examined across independent training runs and other backbones, alongside a comparison with LLM rewriting that tests the value of retaining intact reasoning units. Following your suggestions, we added these experiments, direct latency measurements and estimates of the one-time preparation cost.

W1/Q1. Independent training and curriculum ablations

STC keeps intact reasoning units from the model's own solutions after final-answer verification. PCC trains on these fixed targets, progressively emphasizing higher relative compression and ending with compact example consolidation. Its purpose is to teach shorter reasoning while retaining accuracy.

The controls use the same STC targets and LoRA setup. STC + Standard SFT samples uniformly with the same updates and example presentations. STC + all-step shuffle reorders intact PCC batches across all steps. STC + prefix shuffle + consolidation shuffles only the Early, Middle and Later stages, keeping the final consolidation batches and their order.

Three independent trainings on DeepSeek-1.5B. Within the discussion phase, we trained each of the three configurations below three times, matching starting models, questions and targets within each repetition. Each trained model generated one answer per question. Cells show accuracy (%) / mean tokens, averaged over the trainings. The last column compares Full SCRATCH with each row, averaging the three trainings and benchmarks equally.

ConfigurationGSM8KMATH-500AIME 2020–2025Full SCRATCH vs. row (mean)
STC + Standard SFT74.12 / 460.781.13 / 3760.826.67 / 12144.8−0.08 points / 18.71% fewer
STC + prefix shuffle + consolidation73.59 / 403.380.20 / 2842.725.37 / 11162.5+0.84 points / 4.43% fewer
Full SCRATCH73.82 / 377.480.47 / 2684.627.41 / 10999.4—

Full SCRATCH shortens outputs while retaining mean accuracy close to Standard SFT, and improves both means over the prefix-shuffle control. The mean percentage token reduction across benchmarks is positive in each independent training repetition against both controls.

Ablations on the two additional backbones. Each configuration is one trained model evaluated with five answers per question. Cells show accuracy (%) / mean tokens.

BackboneConfigurationGSM8KMATH-500AIME 2020–2025
DeepScaleR-1.5BSTC + Standard SFT83.21 / 1040.288.12 / 2799.432.22 / 7908.0
DeepScaleR-1.5BSTC + all-step shuffle80.79 / 795.585.24 / 2373.829.11 / 6254.0
DeepScaleR-1.5BSTC + prefix shuffle + consolidation81.91 / 841.286.60 / 2513.130.67 / 6134.7
DeepScaleR-1.5BFull SCRATCH82.79 / 815.587.40 / 2423.131.67 / 5904.9
Qwen3-14BSTC + Standard SFT95.63 / 1540.282.96 / 4315.766.67 / 11713.7
Qwen3-14BSTC + all-step shuffle94.91 / 1250.181.68 / 3659.763.89 / 11348.4
Qwen3-14BSTC + prefix shuffle + consolidation95.62 / 1315.783.08 / 3792.765.00 / 11141.6
Qwen3-14BFull SCRATCH96.21 / 1272.683.88 / 3693.766.11 / 10924.0

Relative to Standard SFT, Full SCRATCH reduces tokens by 20.12% on DeepScaleR-1.5B and 12.84% on Qwen3-14B, with mean accuracy changes of −0.56 and +0.31 points. It is more accurate and shorter than the prefix-shuffle control in all six settings. Keeping consolidation at the end also improves accuracy over all-step shuffle in all six. These ablations match PCC's intended role: stronger supervision for shorter reasoning, with accuracy largely preserved.

Decoding variation in Table 1. The three-training results above address training variation. Here, each cell gives Baseline / SCRATCH over the five evaluation runs in Table 1. Accuracy standard deviation (SD) is expressed in percentage points. For tokens, relative SD (RSD) is the SD of the five run-level means divided by that method's own mean, expressed as a percentage.

Accuracy SD (percentage points).

ModelGSM8KMATH-500AIME 2020–2025
DeepSeek-1.5B0.42 / 1.280.65 / 1.722.12 / 2.29
DeepScaleR-1.5B0.62 / 0.711.13 / 1.002.81 / 1.71
Qwen3-14B0.34 / 0.280.71 / 0.951.65 / 2.55

Token RSD (%).

ModelGSM8KMATH-500AIME 2020–2025
DeepSeek-1.5B2.28 / 1.872.03 / 4.220.55 / 0.74
DeepScaleR-1.5B1.29 / 1.641.82 / 1.061.32 / 2.93
Qwen3-14B0.71 / 1.091.34 / 1.760.67 / 1.56

Across all nine settings, every SCRATCH run uses fewer tokens on average than every Baseline run.


Response 2 of 2

W2/Q2. STC versus LLM rewriting

We added Direct Rewrite + Standard SFT using the same gpt-4o-mini-2024-07-18 annotator, 1,800 native traces, questions and verified final answers. It produces concise solutions and may paraphrase, merge or reorder, with no target length. The rewrites and STC targets use the same Standard SFT, including optimizer updates and example presentations.

We also include our SPIRIT-FT adaptation, which selects step removals by the target model's perplexity on the remaining text and retains its LLM merging branch. It uses the same native traces and Standard SFT. Each row is one trained DeepSeek-1.5B model evaluated with five runs. Cells give accuracy (%) / mean tokens.

ConfigurationGSM8KMATH-500AIME 2020–2025
Direct Rewrite + Standard SFT72.24 / 423.577.80 / 3234.822.11 / 11494.1
SPIRIT-FT adaptation72.84 / 396.778.20 / 2988.923.44 / 11318.4
STC + Standard SFT74.83 / 459.680.28 / 3712.424.22 / 12102.5
Full SCRATCH74.04 / 377.581.00 / 2688.826.67 / 10986.6

With the same Standard SFT, STC targets are more accurate than direct rewrites on all three benchmarks, by 2.39 percentage points on average. Rewriting changes both the retained content and its expression; STC keeps the model's original wording and order. These results suggest that retaining the model's own expressions helps preserve accuracy during compression. With PCC, Full SCRATCH is both more accurate and shorter than direct rewriting, by 3.19 points and 10.72% fewer tokens on average, and also improves both metrics over SPIRIT-FT.

Q3. Latency and offline cost

The annotator is used once, to prepare training data. The trained model answers without it, so the per-response saving accumulates with use. We timed the original model (BASE) and SCRATCH directly on one RTX PRO 6000 Blackwell Server Edition GPU, processing one question at a time, from tokenization to the local return of the answer and excluding model loading and warm-up. Averaged equally over the nine model–benchmark settings, mean latency per response falls by 24.92%. Mean MATH-500 times:

BackboneBASESCRATCH
DeepSeek-1.5B16.481 s10.247 s
DeepScaleR-1.5B11.743 s9.196 s
Qwen3-14B94.782 s75.525 s

The table below gives the one-time cost of preparing data and training one model. Generating solutions includes rejected and resampled ones. Annotation covers the 2,000-question preparation pool and retries at the published gpt-4o-mini-2024-07-18 rates of USD 0.15 / 0.60 per million input / output tokens; output charges stay small because the annotator returns step IDs rather than text. GPU hours are priced at an assumed USD 2.50 per hour.

BackboneGenerating solutions, GPU-hTraining, GPU-hAnnotation, USDTotal, USD
DeepSeek-1.5B0.720.951.225.40
DeepScaleR-1.5B1.151.281.717.79
Qwen3-14B2.803.602.0118.01

At that price and with the MATH-500 timings above, DeepSeek-1.5B saves 6.23 s per response. This covers the annotation cost after about 280 comparable requests and the full one-time cost after about 1,250. The corresponding full-cost figures are about 4,400 requests for DeepScaleR-1.5B and 1,350 for Qwen3-14B.

Since annotation cost depends on the annotator, we also checked how much that choice matters. We compared each annotator's step selections with an expert reference on 50 solutions using F1, which measures how well the selected steps match the reference by balancing missed necessary steps against unnecessary selected ones. GPT-4o-mini reaches 96.19%, close to Qwen3-32B (96.44%) and DeepSeek-V4.1-Flash-1017 (96.66%). The smaller Qwen3-4B reaches 92.96%, and relabeling all 1,800 training solutions with it changed DeepSeek-1.5B accuracy by −0.53 to −1.11 percentage points across the three benchmarks, relative to default SCRATCH.

Your suggestions helped us examine both how SCRATCH learns shorter reasoning and what the savings mean in practice. The added results support PCC's intended role of reducing tokens while largely preserving accuracy, and the cost estimates indicate when repeated use offsets the one-time preparation cost. We hope these results are helpful as you reconsider the work and would be happy to discuss any remaining questions or follow-up experiments you would find useful.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Responses

We thank the reviewers and Area Chair for their time and interest in SCRATCH and appreciate the reviewers' recognition of its accuracy-efficiency trade-off, controlled ablations, and simple deployment. Their questions guided the additional experiments below; references point to our detailed replies.

STC retains intact units from the model's own verified solutions; PCC gradually increases training exposure to targets with higher relative compression.

1. Curriculum benefits across trainings and backbones. We ran three independent trainings of Full SCRATCH and each of two controls on DeepSeek-1.5B. Averaged over trainings and three benchmarks, Full SCRATCH uses 18.71% fewer tokens than STC + Standard SFT with a −0.08-percentage-point accuracy change. Against STC + prefix shuffle + consolidation, which shuffles earlier batches while keeping the final stage fixed, it improves accuracy by 0.84 points and reduces tokens by 4.43%. In separate ablations on DeepScaleR-1.5B and Qwen3-14B, Full SCRATCH also improves both metrics over this control in all six settings. Across five evaluation runs in each of the nine settings, every SCRATCH run has lower mean token use than every Baseline run. (AFma, Response 1, W1/Q1; vXbd, Response 1, W1)

2. Comparisons supporting the design. On DeepSeek-1.5B, STC yields 2.39 percentage points higher mean accuracy than Direct Rewrite with the same annotator, native traces, and Standard SFT. Full SCRATCH improves mean accuracy by 1.49 points and uses 16.29% fewer tokens than the shortest-correct best-of-16 SFT (Naive BoN). It is also more accurate and shorter than our SPIRIT-FT and TLTS adaptations on all three benchmarks. These comparisons support intact-unit selection and an offline curriculum organized by relative compression. (AFma, Response 2, W2/Q2; vXbd, Response 1, W2/W3)

3. Practical savings. Direct timing on one RTX PRO 6000 GPU, one request at a time, shows a 24.92% mean latency reduction across nine settings. Estimated one-time costs for trace generation, annotation, and training are USD 5.40–18.01 per model, under reported API prices and an assumed USD 2.50/GPU-hour. The annotator is used only for data preparation; no extra model is needed at inference. (AFma, Response 2, Q3)

4. Useful reasoning and the source of savings. In an expert audit of 50 traces, the annotator retained 60 of 62 useful error-correction units. Tests of strategy diversity on 50 questions and recovery from 100 erroneous starts per model find counts close to Baseline. Correct responses become shorter in all nine settings and account for about 65% of token savings on average. We also verified no overlap between the selected training/development pool and AIME 2020–2025. (vXbd, Response 2, Q1–Q4)

5. Robustness and transfer. On DeepSeek-1.5B, we tested a weaker annotator, 5%/10% erroneous training examples and curriculum changes. The tested curriculum settings change accuracy by −0.34 to +0.16 percentage points; with 5% erroneous examples, the largest drop is 0.96 points. Our BBH analysis links smaller-model transfer errors to omitted intermediate states or conditions. (htY3, W1/Q1–Q2, W2, W3/Q3)

We hope this evidence helps the reviewers and Area Chair assess SCRATCH's contribution and practical value. We welcome your feedback on remaining questions during the discussion and would be glad to provide focused clarifications or follow-up analysis.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.