acceptodds
Under review as a conference paper at ICLR 2027

From Self-Correction to Self-Evolution: Agentic Test-Time Learning for Mathematical Reasoning

Abstract

Recent advances in self-correction (SC) and recursive self-improvement (RSI) offer new ways to improve large language models' reasoning. A natural approach is to accumulate correction experience as reusable skills to guide subsequent problem solving. To this end, we propose SC2SE (Self-Correction to Self-Evolution), an agentic test-time learning framework for mathematical reasoning. With model parameters frozen and without requiring ground-truth answers as learning feedback, SC2SE learns from its own verification and revision experience through four collaborating agents for problem solving, stepwise correction, skill evolution, and role calibration. These agents distill correction experience into skills; role calibration checks each skill's suitability for the solver role before and after skill aggregation, turning diagnoses of existing solutions into actionable guidance for subsequent problem solving. Updated skills guide the next round of problem solving, and new answers undergo further verification and correction, forming a test-time self-evolution loop: correction → skill accumulation → skill-assisted answering → further correction. We conduct same-pool evaluation on four mainstream large language models and five public mathematics datasets. The results show that skill-assisted initial answers achieve higher mean accuracy (ACC) than the compared RSI baselines; with further correction, SC2SE improves mean ACC by 14.3 percentage points over the highest mean ACC among the compared baselines. Together with component comparisons and process case studies, these results support the complementary roles of self-correction and skill accumulation: correction provides experience for skill learning, while accumulated skills support subsequent problem solving and revision.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, mathematical reasoning, self-correction, test-time learning, skill evolution, agentic systems

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_P3DtOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SC2SE, which is a training-free loop where agents turn the self-reflection on math problem solving into skill library. Four specialized agents, solver, stepwise corrector, skill evolver, role calibrator, composes a skill creation pipeline.

Critical strengths
  • Extensive evaluation results: 4 LLM models, 5 competitive benchmarks
Critical weaknesses

Limited novelty, little add-on to existing literature. Need clear contribution upon existing works such as SkillOpt, SkillAudit, EvoLib, ProgCo. Role calibration seems to be the only difference, which is also not very interesting.

Questions for authors

What is the evaluation result without role calibration?

1 comment
AuthorsOpenReview ↗

Response to Reviewer P3Dt

Thank you for your suggestions. We have added a role calibration ablation and a comparison of SC2SE with related methods. We address each question below.

Q1. How does accuracy change when role calibration is turned off?

Turning off role calibration reduced initial accuracy by 7.1 percentage points and accuracy after correction by 3.5 points. By the time we ran this test, DeepSeek had updated its API and the Flash version used in the submission could no longer be selected. Both groups used the updated DeepSeek Flash. We compared the groups on the same five datasets with the same shared initial answers and random seed. The table gives the ablation results.

Answer stageCalibration offCalibration onGain with calibration on
Initial answer58.2%65.3%+7.1 percentage points
After correction72.9%76.5%+3.5 percentage points

With role calibration on, initial accuracy increased by 7.1 percentage points and accuracy after correction increased by 3.5 points. Calibration makes the skills easier for the solver to use on new problems and thus directly affects the initial answer. The corrector does not read the skill library. The gain after correction may reflect some additional correct initial answers remaining correct after verification.

How SC2SE differs from existing methods

SC2SE studies how to turn an error found on one problem into a step the solver can use in its initial answer to the next problem. For example, dividing z2=zz^2=z by zz causes the solver to miss z=0z=0. SC2SE extracts the following guidance from this correction: “Before dividing by an expression that contains an unknown, first consider the case where the expression is zero.” Before and after merging skills, the calibration agent checks whether the solver can use this guidance from the new problem's conditions alone.

We drew on SkillOpt's structured skill edits. SkillOpt proposes edits from scored trajectories and selects updates using a separate validation set. SkillAudit compares how the same task is performed with and without skills and uses this comparison to revise the skills. SC2SE extracts guidance from solving and correcting math problems at test time, without feedback from reference answers. SkillOpt · SkillAudit

ProgCo uses feedback from programmatic verification to revise the current answer. Its mean accuracy across four models was 45.6%. ProgCo paper

We compared SC2SE with Dynamic Cheatsheet (DC) and ACE. Your suggestion to include another related method helps test whether learning skills from solving problems is effective. We therefore also evaluated EvoLib, which extracts knowledge from problem solving for use on later problems. The table reports the four-model averages and the additional results on the updated DeepSeek Flash separately. In the additional comparison, we scored the final answers produced by each method's own learning and answering procedure. EvoLib paper

ExperimentMethodAccuracy
Four-model meanDC42.8%
ACE44.1%
SC2SE64.6%
Additional test on updated DeepSeek FlashEvoLib65.9%
SC2SE76.5%

Across four models, SC2SE averaged 64.6% accuracy, compared with 42.8% for DC and 44.1% for ACE. On the updated DeepSeek Flash, SC2SE achieved 76.5% accuracy, 10.6 percentage points above EvoLib's 65.9%.

Reviewer_2hhxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SC2SE, a training-free test-time learning framework for competition mathematics built from four agents. A solver answers with a skill library in its prompt, a stepwise corrector checks and revises the answer with tools, a skill-evolution agent turns solving trajectories into edits to the skill library, and a role-calibration agent makes sure each skill is something the solver can actually use. The skill library is learned without ground-truth answers. Experiments on several math competition benchmarks and four LLMs show that SC2SE outperforms existing self-correction and experience-learning baselines, both for initial answers and for corrected answers.

Critical strengths
  1. The motivation is clear. A corrector can point out what went wrong, but its diagnosis is not always something the solver can act on. Turning the corrector's diagnoses into guidance the solver can use is a sensible starting point, and role calibration states this idea cleanly.
  2. The method is fully specified and does not rely on ground-truth answers. The pipeline and prompts are described in enough detail to reproduce, and learning without labels makes the setting closer to real test-time use.
  3. The experiments are broad and reported transparently. The paper covers many baselines, models and datasets, and it also reports the corrector without skills and the costs. This lets readers judge for themselves where the gains come from.
Critical weaknesses
  1. Most of the gain comes from the corrector, not from self-evolution. Table 2 shows that the stepwise corrector without any skills already beats every baseline, and adding skills improves accuracy by only a few points on top of that. The "self-evolution" framing in the title and abstract therefore overstates what the skills contribute. The no-skill control is also missing for GPT Luna.
  2. Skills are learned and evaluated on the same problems. The skill library is built on the evaluation problems themselves, and the case study (p. 9) re-solves the very problem that produced the skill. The main experiments include no held-out evaluation, so it is not clear whether the learned skills transfer to unseen problems.
  3. There is no uncertainty reporting. Every number comes from a single temperature-0 run on 30 to 50 problems per dataset, so one problem can move accuracy by about 3 points. With margins of a few points, it is hard to tell whether the improvements are stable.
  4. Compute is neither matched nor reported against baselines. SC2SE adds four learning rounds plus a full verify-and-revise pass, but the paper does not compare token usage or model calls with the baselines. Role calibration is listed as a contribution, yet it is never ablated, so its effect is unknown.
Questions for authors
  1. Could you learn skills on one part of each dataset (or on competitions from other years) and evaluate both initial and corrected answers on disjoint problems, for at least two models?
  2. Could you add the no-skill result for GPT Luna and an ablation without role calibration? How many skills did role calibration reject?
  3. Could you report bootstrap confidence intervals over problems, or results from several sampled runs, for Tables 1 and 2?
  4. How many tokens and model calls per problem do SC2SE (learning plus evaluation) and each baseline use? How do baselines such as Self-Refine compare when given more rounds to match compute?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 2hhx (1/3)

Thank you for your suggestions. We have added results on how well the skills transfer to new problems, accuracy before and after using skills with confidence intervals, and the computational costs of learning skills and answering problems. We address each question below.

Q1. Do the learned skills transfer to unseen problems?

(1) Transfer across datasets. We held out one entire dataset at a time, built a library using only skills learned from the other four datasets, and used DeepSeek Flash to answer the held-out problems. The held-out problems were never used to learn or select skills, and we fixed the library before answering them. With skills, overall initial accuracy rose from 52.9% to 60.0%. Accuracy increased on four of the five held-out datasets and stayed the same on one.

Held-out test setBaseline initial accuracyInitial accuracy with skills from other datasets
AIME 202473.3%83.3% (+10.0)
AIME 202573.3%73.3% (+0.0)
BRUMO 202586.7%90.0% (+3.3)
HMMT February 202553.3%66.7% (+13.3)
AMO Bench8.0%16.0% (+8.0)
Overall52.9%60.0% (+7.1)

Parentheses give the change from baseline in percentage points.

(2) Transfer across years. By the time we ran these tests, DeepSeek had updated its API and the Flash version used in the submission could no longer be selected. We tested the updated DeepSeek Flash and GPT Luna. Each model pooled its own skills learned from the five source datasets. We then fixed each library and tested it on new problems from AIME 2026 and HMMT February 2026. The empty and pooled libraries used the same correction procedure.

ModelInitial, empty libraryInitial, pooled libraryAfter correction, empty libraryAfter correction, pooled library
DeepSeek Flash$^{*}$63.49%69.84% (+6.35)80.95%84.13% (+3.17)
GPT Luna23.81%30.16% (+6.35)50.79%57.14% (+6.35)

∗^{*} denotes the updated DeepSeek Flash.

For Luna, the pooled library improved both initial accuracy and accuracy after correction by 6.35 percentage points.

In this cross-year test on the updated DeepSeek Flash, the pooled library outperformed the empty library by 6.35 percentage points on initial answers. The gap narrowed to 3.17 points after correction. The skill library directly guides the solver's initial answer. Both groups then use the same corrector to check each step against the problem and revise the solution. The empty-library group can also fix errors in its initial answers at this stage, which may narrow the gap after correction.

Q2. What do the skills and role calibration each contribute?

The skill library gives the solver steps for solving problems. Role calibration checks whether those steps apply to new problems and rewrites skills when needed. The solver uses the skills to write an initial answer, which the corrector then checks and revises.

Across the four models, mean initial accuracy rose from 42.2% without skills to 47.1% with skills and reached 64.6% after correction (see Table 2 in Q3 for results by model). On GPT Luna, final accuracy with skills was lower than that of the independently implemented no-skill reference but higher than that of the empty-library control in the same implementation.

Across the four models, role calibration attempted two repairs. The Qwen 35B repair passed rechecking. The DeepSeek Flash repair did not succeed.

We compared role calibration on and off on the updated DeepSeek Flash across five datasets, with the random seed fixed at 42. With calibration off, the system skipped the role checks before and after skill aggregation. All other settings were the same. The table gives the ablation results.

Role calibrationInitial accuracyAccuracy after correction
Off58.2%72.9%
On65.3% (+7.1)76.5% (+3.5)

Parentheses give the gain in percentage points with calibration on relative to calibration off.

With role calibration on, initial accuracy increased by 7.1 percentage points and accuracy after correction increased by 3.5 points. Calibration makes the skills easier for the solver to use on new problems and thus directly affects the initial answer. The corrector does not read the skill library. The gain after correction may reflect some additional correct initial answers remaining correct after verification.

AuthorsOpenReview ↗

Response to Reviewer 2hhx (2/3)

Q3. 95% confidence intervals for Tables 1 and 2

For Tables 1 and 2, we added paired bootstrap 95% confidence intervals for each method's difference from the shared initial answers (Base). We resampled problems with replacement within each of the five datasets, keeping the number of problems per dataset fixed. In each of 50,000 replicates, we used the same sampled problem IDs for all methods and all four models. Both tables report accuracy as a percentage. The last column gives the change in four-model mean accuracy relative to Base, in percentage points.

Table 1. Accuracy on the five datasets by method

MethodFlashProQwen 35BLunaMeanChange from Base [95% CI]
Base52.938.853.523.542.2—
Reflex60.037.654.731.245.9+3.7 [2.1, 5.3]
Self-Refine61.851.257.630.650.3+8.1 [5.7, 10.6]
Self-Reflection55.340.057.132.946.3+4.1 [2.2, 6.2]
CheckList55.341.257.130.045.9+3.7 [1.6, 5.7]
CoT-Tool54.140.041.822.439.6−2.6 [−6.2, 0.9]
ProgCo58.241.257.125.945.6+3.4 [2.1, 4.9]
ProgCo-Py62.448.258.828.249.4+7.2 [5.3, 9.3]
Dynamic Cheatsheet58.244.752.415.942.8+0.6 [−2.9, 4.1]
ACE55.345.953.521.844.1+1.9 [−0.4, 4.4]
SC2SE initial answer60.044.157.626.547.1+4.9 [1.6, 8.2]
SC2SE after correction74.770.065.348.264.6+22.4 [18.1, 26.6]

Table 2. Initial accuracy and accuracy after correction with and without skills

SettingFlashProQwen 35BLunaMeanChange from Base [95% CI]
Base (shared initial answers)52.938.853.523.542.2—
SC2SE initial answers with skills60.044.157.626.547.1+4.9 [1.6, 8.2]
SC2SE no-skill reference, after correction72.466.564.152.463.8+21.6 [17.9, 25.3]
Full SC2SE, after correction74.770.065.348.264.6+22.4 [18.1, 26.6]

Table 1 shows that full SC2SE improved mean accuracy over Base by 22.4 percentage points, with a 95% confidence interval of [18.1, 26.6]. Table 2 shows the contribution of skills at each stage. Skills improved mean initial accuracy by 4.9 percentage points [1.6, 8.2]. After both groups underwent correction, the mean difference between the full method and the no-skill reference was +0.7 percentage points [−2.4, 3.8]. On these problems, the gain from skills was clearer at the initial answer stage.

AuthorsOpenReview ↗

Response to Reviewer 2hhx (3/3)

Q4. Computational cost and number of correction rounds

1. Computational cost per problem

We measured the mean cost per problem for DeepSeek Flash across the five datasets. Total tokens include both input and output. Some token counts were estimated from run logs.

MethodAccuracyModel calls per problemTotal tokens (k/problem)
Reflex60.0%319.4
Self-Refine61.8%636.0
Self-Reflection55.3%322.2
CheckList55.3%420.0
CoT-Tool54.1%310.9
Dynamic Cheatsheet58.2%2.2952.1
ACE55.3%5.3854.2
ProgCo58.2%7.8257.2
ProgCo-Py62.4%48.5517,004.5
SC2SE74.7%138.291,055.3

(1) Compared with standard feedback-based correction methods, SC2SE used more tokens but achieved higher accuracy. For example, SC2SE outperformed Self-Refine by 12.9 percentage points. Most of the additional cost came from learning skills and evaluating candidate libraries.

(2) Compared with ProgCo-Py, which also uses Python, SC2SE achieved 12.4 percentage points higher accuracy using just 6.2% as many total tokens per problem. When skill learning was included, SC2SE used more model calls.

SC2SE used fewer total tokens than ProgCo-Py in part because it moves directly to correction after the first failed step and skips the remaining checks. When a calculation is needed, the system selects a math tool for the task and reuses a successful result for identical requests on the same problem. At the soft limit on consecutive tool calls, it prompts the model to answer using the results so far. At the hard limit, it ends the current round of calls. SC2SE also compresses long inputs. The four Flash requests that triggered compression fell from 133,200 to 43,200 tokens on average.

2. Gains from additional correction rounds

We did not directly measure what accuracy the baselines would reach with the same total budget as SC2SE. We examined accuracy by round to see how much the baselines gained as the budget increased. All three baselines gained the most in round 1, and their gains per round generally declined thereafter. Accuracy gains ranged from 4.7 to 5.9 percentage points in round 1 and from 0 to 1.2 points in round 3. Parentheses in the table give the gain over the previous round in percentage points.

MethodInitial answerRound 1Round 2Round 3
Self-Refine52.9%58.2% (+5.3)60.6% (+2.4)61.8% (+1.2)
ProgCo52.9%57.6% (+4.7)57.6% (+0.0)58.2% (+0.6)
ProgCo-Py52.9%58.8% (+5.9)62.4% (+3.5)62.4% (+0.0)

We also tested Self-Refine on the updated DeepSeek Flash using the same problems and shared initial answers. In each round, it first generated feedback and then revised the solution. The table gives accuracy from rounds 0 through 8 and cumulative model calls and tokens since the shared initial answer.

Correction roundsAccuracyGain over previous round (pp)Cumulative calls per problemCumulative tokens (thousands per problem)
052.9%—00.0
158.8%+5.9212.6
261.2%+2.4425.3
364.7%+3.5637.4
465.3%+0.6848.7
565.3%+0.01059.9
665.9%+0.61271.4
765.3%-0.61482.5
865.9%+0.61693.9

Self-Refine gained 11.76 percentage points in the first three rounds and only 1.18 points in total over the next five. Accuracy was 65.9% in rounds 6 and 8 and 65.3% in round 7. From rounds 3 to 8, cumulative model calls per problem rose from 6 to 16 and cumulative tokens per problem rose from 37,400 to 93,900. The first three rounds accounted for most of the accuracy gain. Figure 11‾\color{#174EA6}{\underline{\textsf{Figure 11}}} in Appendix E.5 of the revised manuscript shows Self-Refine's accuracy from the shared initial answer (round 0) through round 8.

Reviewer_2hhxOpenReview ↗

Response to the authors' rebuttal

Thank you for the thorough response and the considerable effort behind the new experiments; the held-out transfer results in particular are a welcome addition. However, the new results also suggest, as I read them, that most of the final gain comes from the stepwise corrector rather than from skill evolution, which is presented as the main novelty. Together with the remaining difference in compute budget, this leads me to keep my current rating. I do think the corrector is a solid component, and I hope these comments are useful for future revisions.

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

SC2SE uses a loop of four agent roles to turn mathematical problem-solving experience into reusable guidance. A solver generates an answer, and a correction agent checks it using mathematical tools and revises it until verification passes or the correction budget is exhausted. A skill-evolution agent extracts guidance from both successful and failed trajectories and merges it with an existing skill library. Role calibration checks the skills before and after aggregation to ensure they are usable during solving. The updated library guides subsequent answer generation, while model weights remain frozen.

Experiments use four models and five mathematics datasets, total 170 problems per model. The main evaluation learns and tests on the same problem pool over four update rounds, then evaluates initial and further-corrected answers separately. The corrected system reports 64.6% mean accuracy versus 50.3% for the strongest baseline by mean accuracy. A separate cross-dataset experiment examines skill transfer on one model.

Critical strengths
  1. A concrete and interpretable integration of correction and experience reuse.The method clearly specifies how correction evidence becomes solver-oriented guidance, how skills are maintained, and how they influence subsequent answers. Learning without ground-truth feedback makes the setting relevant when reference answers are unavailable.
  2. Positive results supported by several complementary analyses.The full system improves accuracy across all four models. Separating initial and corrected answers, including a no-skill reference, and reporting answer-level repairs and regressions makes the results more informative. The cross-dataset experiment provides useful preliminary evidence that learned guidance can transfer beyond the problems used to construct it.
Critical weaknesses
  1. Limited novelty. SC2SE mainly records previous mistakes, converts them into textual guidance, and reuses them through a multi-agent workflow. Similar reflection, memory, and skill-editing mechanisms already exist in Reflexion, ACE, and SkillOpt. Its main addition, role calibration before and after aggregation is not isolated experimentally. Therefore, the contribution appears primarily an engineering combination of existing components, without clearly established methodological novelty.
  2. Insufficiently controlled comparisons. Firstly, closely related methods such as SkillOpt and SkillAudit are not compared experimentally. Secondly, SC2SE performs four additional skill-learning rounds without matching total computation across methods. Finally, the skill component improves accuracy over the independently implemented no-skill reference by only approximately 1-4 percentage points, without uncertainty estimates. These comparisons leave its benefit over strong correction and additional computation unclear.
Questions for authors
  1. What changes when role calibration is disabled while the rest of SC2SE remains fixed? How many skills are rejected or modified, and what effect do those changes have on answer accuracy?
  2. What accuracy do SC2SE and the baselines achieve when total LLM calls, and tool-use budgets are matched including all learning and correction rounds with the same base model, decoding settings, and evaluation protocol? Please report results under matched conditions.
  3. Table 2’s improvements over the no-skill reference correspond to only 4, 6, and 2 additional correct answers out of 170, while Table 9 specifies seed 42. Have you repeated skill learning with different seeds and trajectory orders? Please report variability and paired uncertainty estimates.
  4. In Table 5, Qwen’s initial accuracy on AIME 2025 drops from 60.0% without skills to 53.3% with skills. Can you trace this regression to particular skill updates and explain when accumulated guidance helps or hurts?
  5. Section 3.3 verifies candidate-library answers, but Equation 4 and Algorithm 1 appear to adopt every structurally valid candidate. Can failed verification reject or roll back an update? What exactly does “valid” mean, and how do the recorded verification results affect acceptance?
3 comments
AuthorsOpenReview ↗

Response to Reviewer JedB (1/3)

Thank you for your suggestions. We have added results on role calibration and computational budgets and explained how the skill library is updated and how accuracy changes across updates. We address each question below.

Q1. How SC2SE relates to existing methods and what role calibration does

We use structured skill edits to turn errors found during correction into steps the solver can use directly on the next problem. Role calibration checks whether those steps can be used on a new problem.

For example, dividing z2=zz^2=z by zz causes the solver to miss z=0z=0. The resulting skill is: “Before dividing by an expression that contains an unknown, check whether the expression can be zero.” Before and after merging skills, the system checks whether the solver can use this guidance from the new problem's conditions alone. It keeps usable guidance, rewrites and rechecks unclear guidance, and discards candidates that remain unsuitable. Skills with identical content can reuse an earlier check.

The table gives the numbers of new checks, reused checks, and repair attempts during role calibration for four models. A given skill can be checked more than once.

ModelNew checksReused checksRepair attempts
DeepSeek Flash5311621
DeepSeek Pro2932350
Qwen 35B2942391
GPT Luna512580

Across the four models, role calibration attempted two repairs. The Qwen 35B repair passed rechecking. The DeepSeek Flash repair did not succeed.

By the time we ran the ablation, DeepSeek had updated its API and the Flash version used in the submission could no longer be selected. We compared role calibration on and off using the updated DeepSeek Flash on five datasets with a fixed random seed of 42. With calibration off, the system skipped the role checks before and after skill aggregation. All other settings were the same. The table gives the ablation results.

Role calibrationInitial accuracyAccuracy after correction
Off58.2%72.9%
On65.3% (+7.1)76.5% (+3.5)

Parentheses give the gain in percentage points with calibration on relative to calibration off.

With role calibration on, initial accuracy increased by 7.1 percentage points and accuracy after correction increased by 3.5 points. Calibration makes the skills easier for the solver to use on new problems and thus directly affects the initial answer. The corrector does not read the skill library. The gain after correction may reflect some additional correct initial answers remaining correct after verification.

Relation to existing skill-learning methods

The related work in the original manuscript discussed SkillOpt and SkillAudit. We did not evaluate SkillOpt empirically because it optimizes skills using scored training trajectories and uses scores on a separate validation set to decide whether to accept an update. We study test-time skill evolution: SC2SE learns skills directly from solving and correcting math problems at test time, without a separate training set or training phase. SkillOpt paper

For SkillAudit, the authors state in the revised paper that they plan to release the code upon publication. As of this response, we have not found a publicly available, runnable implementation from the authors, so we could not evaluate the method directly. Authors' code release statement

We compared SC2SE with Dynamic Cheatsheet (DC) and ACE. Your suggestion to include another related method helps test whether learning skills from solving problems is effective. We therefore also evaluated EvoLib, which extracts knowledge from problem solving for use on later problems. The table reports the four-model averages and the additional results on the updated DeepSeek Flash separately. In the additional comparison, we scored the final answers produced by each method's own learning and answering procedure. EvoLib paper

ExperimentMethodAccuracy
Four-model meanDC42.8%
ACE44.1%
SC2SE64.6%
Additional test on updated DeepSeek FlashEvoLib65.9%
SC2SE76.5%

Across four models, SC2SE averaged 64.6% accuracy, compared with 42.8% for DC and 44.1% for ACE. On the updated DeepSeek Flash, SC2SE achieved 76.5% accuracy, 10.6 percentage points above EvoLib's 65.9%.

AuthorsOpenReview ↗

Response to Reviewer JedB (2/3)

Q2. How accurate are the methods under the same model-call and tool budgets?

The three methods started from the same initial answers and used the updated DeepSeek Flash for subsequent calls. We replayed the runs on five datasets, took the last complete answer to each problem at each budget, and computed accuracy using the three-step scoring procedure. The SC2SE budget includes skill learning, evaluation of candidate libraries, and answer correction. Shared costs were divided among the problems within each dataset. Scoring calls were counted separately.

With at most 250 model calls and 68 Python executions per problem, accuracy was 76.5% for SC2SE, 67.6% for ProgCo-Py, and 65.9% for Self-Refine. SC2SE outperformed ProgCo-Py, the stronger of the two baselines, by 8.8 percentage points. With the limits reduced to 64 model calls and 16 Python executions, accuracy was 77.6%, 66.5%, and 65.9%, respectively.

The next two tables cap model calls and Python executions separately. When only model calls were capped, Self-Refine had the highest accuracy at 16 calls per problem. At 32 calls, SC2SE reached 71.8%, compared with 65.9% for both baselines. For Self-Refine, we used the recorded eight-round correction run.

Model-call limit per problemSelf-RefineProgCo-PySC2SE
1665.9%60.0%55.3%
3265.9%65.9%71.8%
6465.9%67.6%79.4%
12865.9%67.6%79.4%
24065.9%67.6%76.5%
25065.9%67.6%76.5%

When only Python executions were capped, SC2SE was more accurate at every limit shown than ProgCo-Py, which also uses Python. Figure 10‾\color{#174EA6}{\underline{\textsf{Figure 10}}} in Appendix E.4 of the revised manuscript shows the accuracy curves for both types of limits.

Python execution limit per problemProgCo-PySC2SE
457.6%61.2%
862.4%72.4%
1666.5%77.6%
3267.6%77.6%
6467.6%75.9%
6867.6%76.5%

Q3. How stable are the accuracy gains from the skill library?

We used a fixed random seed of 42 in the paper. Changing the seed would require learning and evaluating the skill library again and then answering the problems with the new library. We leave independent runs with multiple seeds to future work.

Thank you for suggesting that we examine the stability of these gains. We added paired 95% confidence intervals to measure sensitivity to the choice of test problems. We resampled problems within each dataset, drawing the results with and without skills together for each problem, and computed the accuracy difference.

ModelInitial accuracy gain [95% CI]Gain after correction [95% CI]
DeepSeek Flash+7.1 [1.2, 13.5]+2.4 [−2.9, 7.6]
DeepSeek Pro+5.3 [−1.2, 11.8]+3.5 [−3.5, 10.6]
Qwen 35B+4.1 [−2.4, 10.6]+1.2 [−4.1, 7.1]
Three-model mean+5.5 [1.4, 9.6]+2.4 [−1.2, 5.9]

All values in the table are in percentage points.

Using the skill library increased initial accuracy by 5.5 percentage points on average across the three models. When generating its initial answer, the solver can use methods learned from earlier corrections. The 95% confidence interval for the mean gain lies entirely above zero.

The corrector does not read the skill library. It uses the problem to check and revise the existing initial answer. We still found that mean accuracy after correction was 2.4 percentage points higher with the skill library, with a 95% confidence interval of [−1.2,5.9][-1.2, 5.9], which includes zero. One possible reason is that some additional correct initial answers remain correct after verification.

AuthorsOpenReview ↗

Response to Reviewer JedB (3/3)

Q4. Skill library updates and Qwen's initial accuracy on AIME 2025

AIME 2025 was the only one of the five datasets on which Qwen's initial accuracy was lower with the final skill library than with the empty library. Across all five datasets, each of the four library versions improved initial accuracy over the empty library. Accuracy rose from 53.5% with the empty library to 57.6% with the final library. On AIME 2025, initial accuracy rose from 60.0% with the empty library to 70.0% after the third update, then fell to 53.3% after the fourth.

Initial accuracyEmpty libraryAfter update 1After update 2After update 3After update 4
Qwen, all five datasets53.5%56.5%57.1%56.5%57.6%
AIME 2025 subset60.0%63.3%63.3%70.0%53.3%

From the empty to the final library, 6.7% of AIME 2025 problems changed from incorrect to correct and 13.3% from correct to incorrect. We checked these changes problem by problem. For example, a digit arrangement problem required listing combinations of four digits summing to 18. Qwen answered correctly with the empty library and after each of the first three updates, but omitted {1,4,5,8}\{1,4,5,8\} on its final attempt and thus missed a class of arrangements.

The fourth update revised a geometric pairing rule. An earlier solution had double-counted pairings along a diameter. The revised rule explicitly states that “there is only one way to pair along a diameter,” and the model cited this guidance in its solution. It answered this geometry problem correctly in both of the last two rounds. Changes in initial answers on other problems may reflect the influence of the updated prompt on the reasoning or variation when the model answers again.

Initial accuracy on AIME 2025 peaked after the third update, while overall accuracy across the five datasets peaked after the fourth. In future work, we plan to track the libraries on independent validation problems to decide when to stop updating and retain the better-performing version.

Q5. Does per-problem verification affect whether a candidate skill library is adopted?

No. In our experiments, after the solver generates initial answers using the candidate library, the system verifies them problem by problem and records the results to track performance across rounds. Even if some answers fail verification or the pass rate falls relative to the previous round, the system still uses the candidate library in the next round.

The calibration agent first checks whether new or revised skills can guide the solver. The program then checks the type and target of the selected edits. A “valid edit” in the paper is one that passes these checks and can be applied to the current library. If there are no valid edits in a round, the system keeps the previous library.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Manuscript revisions

For ease of reading, we highlight the key points in bold in our responses. In the revised manuscript, the main additions to the text and appendix are marked in blue.

The supplementary material includes the data for the additional experiments reported in our responses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.