acceptodds
Under review as a conference paper at ICLR 2027

EvoMoSkill: Evolving a Mixture-of-Skills for LLM Agents

Abstract

Enabling agent skills to autonomously refine reusable knowledge is a central challenge in adapting large language model (LLM) agents to real-world applications. Recent work treats skills as trainable external parameters, following the iterative optimization paradigm of deep learning. However, these methods typically optimize a single skill across tasks, so revisions that benefit one task group can disrupt others. A skill library can limit this interference, but raises questions across skill creation, task routing, failure attribution, and evaluation. We introduce EvoMoSkill, a mixture-of-skills framework that jointly optimizes a share skill, expert skills, and a router. Each optimization step applies Credit-Assigned Gradient Descent (CreAGD) in five stages: (1) **Forward execution and credit assignment** attributes execution failures to skill text or router scope; (2) **Textual and routing gradient descent** edits skills, creates experts, and revises routing; (3) **Scope-conditioned paired evaluation** compares candidates with the original system on tasks they would serve; (4) **History-calibrated step acceptance** checks gains against past outcomes and accepts revisions after validation-set checks; and (5) **Optimizer memory accumulation** uses rejected drafts and feedback to improve later updates. Experiments with five LLMs on five benchmarks demonstrate state-of-the-art performance, substantially outperforming human-authored skills, LLM-generated skills, and baselines that optimize only a single skill. These results show that evolving mixture-of-skills yields more adaptable agents.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Self-Evolution, Mixture-of-Skills, Skill Optimization, Credit-Assigned Gradient Descent

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_jmVCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EvoMoSkill maintains a shared skill, task-family experts, and a router for frozen LLM agents. CreAGD diagnoses failures, proposes skill or routing edits, compares candidate updates, and validates them before changing the library. The paper reports improvements over single-skill baselines on five benchmarks and five student models.

Critical strengths

The paper looks at a practical problem that is easy to understand: a shared update can help one type of task while making another worse. Testing five benchmarks and five models gives the study useful breadth. I also found the acceptance-check ablations helpful.

Critical weaknesses

The main concern is that the experiments do not directly test the motivating claim. There is no controlled study of conflicting updates, and no clear measure of how often an update helps one family while hurting another. The full method also adds expert capacity, task-family structure, routing, several candidate drafts, and guarded acceptance at the same time. Because these factors are bundled together, it is hard to tell whether mixture-based joint evolution is the key reason for the improvement.

Most parts of the method are familiar on their own, and the paper does not include a closely matched dynamic skill-library or routing baseline. In some cases, static partitions or admission-time scopes perform as well as, or better than, dynamic routing. The final results also use only seed 42. Overall, the current experiments do not yet show enough novelty or robustness beyond a careful combination of existing ideas.

Questions for authors

Can the authors test the interference claim directly? For example, report how often a single-skill update improves one family but hurts another, then show whether EvoMoSkill prevents these regressions when skill capacity and optimization calls are matched.

Please compare against a modular skill-library baseline with the same family information, skill-token budget, candidate budget, and validation budget. It would also help to test noisy or inferred family labels, run independent seeds, and report total teacher calls, rollout cost, token use, and wall-clock time.

7 comments
AuthorsOpenReview ↗

Responses to Reviewer jmVC (1)

Responses to Weakness 1 and Question 1

Thank you for your valuable comments!

The main concern is that the experiments do not directly test the motivating claim. There is no controlled study of conflicting updates, and no clear measure of how often an update helps one family while hurting another. The full method also adds expert capacity, task-family structure, routing, several candidate drafts, and guarded acceptance at the same time. Because these factors are bundled together, it is hard to tell whether mixture-based joint evolution is the key reason for the improvement.

Can the authors test the interference claim directly? For example, report how often a single-skill update improves one family but hurts another, then show whether EvoMoSkill prevents these regressions when skill capacity and optimization calls are matched.

We appreciate your suggestion to measure how updates affect different task families directly. We added update replay experiments to test whether a change helps one family while harming another, together with modular ablations to distinguish the roles of credit assignment and scope updates.

1. Replaying updates to measure cross-family interference (Table R11)

To measure each update's effects on different families, we re-evaluate the saved libraries and candidates after training. This procedure assesses changes already generated during training without further skill optimization. It has three steps.

  1. Establish the reference before the update. For each training step, we retrieve the library and routing state at the start of that step as the common reference for every comparison.
  2. Construct the versions to evaluate. For a shared-skill or expert-text draft, we copy the old library and replace only the relevant text, holding all other skills and routing fixed. For an accepted new expert, we add its text and scope together. Every candidate is evaluated separately against the same old library. We also retrieve the library actually deployed after the step to assess the combined effect of all accepted changes.
  3. Compare family-level performance. The old library and each evaluated version execute the same tasks. We average repeated executions to obtain each family's success rate, then subtract the old score from the new score to measure the update's effect on that family.

ALFWorld uses 140 tasks not involved in candidate acceptance, with two executions per task, Qwen3-32B, and seed 42. SearchQA uses a fixed set of 300 test questions, with three executions per question, covering three training seeds each for DeepSeek-V4-Flash and Kimi-K2.6. We compare Single, which optimizes one shared skill, with EvoMoSkill, which uses a shared skill and family experts. On ALFWorld, we also include Combo, the whole-library acceptance variant whose rollback rules are given in Table R5. All replay evaluations use gpt-5.6-sol for routing.

To avoid counting random fluctuations as regressions, we independently re-evaluate the unchanged old library. For each benchmark and student model, we pool the family-level differences between repeated evaluations across steps. Each family's noise threshold is the 95th percentile of the absolute differences. Only improvements or declines exceeding the corresponding threshold are counted.

In Table R11, Updates counts evaluated changes, and Drops counts changes that cause at least one family to regress. Conflicts requires the same change to improve one family and harm another, with both effects exceeding their respective thresholds. Mean worst-family change is computed only over regressing updates. For each such update, we take the change in the most adversely affected family, then average these changes in percentage points.

Table R11. Cross-family interference in candidate and deployed updates. Drops and conflicts exceed family-specific noise thresholds. Mean worst-family change covers regressing updates only.

BenchmarkModelMethodUpdate typeUpdatesDropsConflictsMean worst-family change (pp)
SearchQADeepSeek-V4-FlashSingleShared-skill candidate39172−4.4
SingleDeployed full step1920−2.6
EvoMoSkillShared-skill candidate932−3.4
EvoMoSkillExpert-text candidate3950−2.6
EvoMoSkillAccepted new expert900—
EvoMoSkillDeployed full step1211−4.8
Kimi-K2.6SingleShared-skill candidate3590−6.8
SingleDeployed full step1930−6.7
EvoMoSkillShared-skill candidate722−7.7
EvoMoSkillExpert-text candidate2421−4.8
EvoMoSkillAccepted new expert411−6.0
EvoMoSkillDeployed full step1044−6.7
ALFWorldQwen3-32BSingleShared-skill candidate410−15.7
SingleDeployed full step100—
EvoMoSkillShared-skill candidate410−12.5
EvoMoSkillExpert-text candidate1854−11.6
EvoMoSkillAccepted new expert433−21.2
EvoMoSkillDeployed full step410−11.5
ComboShared-skill candidate420−11.2
ComboExpert-text candidate1662−19.5
ComboAccepted new expert310−12.5
ComboDeployed full step200—

Table R11 shows three findings.

  1. Shared-skill revisions can cause family-level regressions, including conflicts that help one family while harming another. On SearchQA with DeepSeek-V4-Flash, 17 of Single's 39 shared-skill candidates cause a family-level regression, or 43.6%. Two also improve one family while harming another. With Kimi-K2.6, 9 of 35 candidates regress, or 25.7%. This directly shows that a local repair to a shared skill can disrupt procedures that already work for other tasks.
  2. Expert-text candidates have lower family-level regression rates in these two SearchQA settings. With DeepSeek-V4-Flash, 5 of EvoMoSkill's 39 expert-text candidates regress, or 12.8%. With Kimi-K2.6, 2 of 24 regress, or 8.3%. Compared with Single's corresponding rates of 43.6% and 25.7%, expert revisions less frequently harm other families. This supports retaining separate procedures for different tasks and placing family-specific repairs in the corresponding expert.
  3. Whole-library acceptance further filters regressions in combined updates. On ALFWorld, 1 of the original EvoMoSkill's 4 deployed updates causes a family-level regression. Neither of the 2 deployed updates retained by Combo regresses. Combo still generates candidates that reduce family performance, but its accepted complete updates preserve family performance. Local candidate screening and whole-library execution checks can therefore work together to control the effects of updates.

2. Ablating credit assignment and scope updates (Table R1)

To distinguish the contribution of the expert structure from that of the update mechanism, we retain the same shared-skill and family-expert structure and vary only credit assignment and scope updates. All variants use the same teacher, student model, and data split, together with the same family information, expert-count limit, candidates per step, acceptance procedure, and training schedule. This ablation uses seed 42 throughout. Each variant is trained independently from scratch, selects its best checkpoint on validation data, and is evaluated on the same test tasks.

Here, credit assignment identifies whether a failure calls for a change to the shared skill, an expert's text, or a routing scope, and provides the teacher with the corresponding revision feedback. When it is disabled, the teacher still receives raw failure trajectories and can propose revisions. Scope updates allow an expert's applicable task range to change after its creation. Fixing scopes disables only this adjustment. Skill text continues to be optimized.

  • M0: Fixed scopes, no credit assignment. This retains the modular library and measures the joint benefit of the two mechanisms when compared with the full method.
  • M1: Editable scopes, no credit assignment. This removes only attribution feedback from the full method to test the contribution of credit assignment.
  • M2: Fixed scopes with credit assignment. This removes only scope adjustment from the full method to test the contribution of dynamic scopes.

Full EvoMoSkill retains both mechanisms. We compare all four settings on ALFWorld with Qwen3-32B and compare the full method with M0 on SearchQA using Kimi-K2.6. In Table R1, the 95% CI estimates the full method's success rate minus that of the comparator in the same row. The p-values use paired McNemar tests.

Table R1. Modular ablations with the same task-family information. Bold marks the highest success rate within each setting. CIs estimate EvoMoSkill minus the comparator.

BenchmarkModelMethodScope updatesCreditSuccess (%)95% CI (pp)p
ALFWorldQwen3-32BEvoMoSkillEnabledEnabled86.57——
M0FixedDisabled85.07[−2.08,+5.07]0.6875
M1EnabledDisabled77.61[+2.59,+15.32]0.0118
M2FixedEnabled84.33[−1.61,+6.09]0.4531
EvoSkill (task-wise selection)——75.37——
SearchQAKimi-K2.6EvoMoSkillEnabledEnabled84.69——
M0FixedDisabled83.98[+0.19,+1.24]0.0129

Credit assignment provides a significant gain with the same expert capacity and family information. On ALFWorld, the full method reaches 86.57%, while M1 without attribution reaches 77.61%, a decrease of 8.96 percentage points, with p = 0.0118. Both allow scope updates and use the same teacher and acceptance procedure. Identifying the update target associated with a failure is therefore more effective than asking the teacher to process unassigned failure trajectories directly.

The full mechanism also outperforms fixed specialization on question answering. On SearchQA, the full method and M0 reach 84.69% and 83.98%, respectively, with p = 0.0129. Targeted content and scope revisions provide an additional gain with the same family information and expert capacity. Together with the replay results in Table R11, these findings show that local experts provide separate places to revise task-specific procedures, while credit assignment makes those revisions more effective. We will include these settings and results in the revision. Thank you again for your suggestions.

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

The paper propose EvoMoSkill, a framework that jointly optimizes a shared skill document, a set of task expert skills, and a router for LLM agents. The core is the Credit-Assigned Gradient Descent (CreAGD).

A teacher stronger LLM is responsible for 1.examines failed student trajectories 2. assigns failures to skill text or router scope 3.proposes bounded edits or new experts 5.accepts revisions only after paired evaluation, historical calibration, regression checks, and validation.

The student model remains frozen. Experiments cover five benchmarks and five student models.

Critical strengths
  1. Treating skill evolution as joint optimization of skill content and applicability is good beyond single-skill training. The router scope is part of the editable object. It is a useful design choice for heterogeneous task families. This shifts the method from "which skill text is better" to "which skill should serve which task."

  2. The second is the experimental breadth. Five benchmarks and five frozen students. The ablation set is extensive. The acceptance-check ablation, the routing ablation, and the expert-lifecycle ablation are meaningful components.

Critical weaknesses
  1. The gap between the claimed contribution and the evidence. The method is essentially a strong teacher LLM that edits skill text and router scopes, wrapped in a multi-stage acceptance pipeline. The paper does not show that this pipeline outperforms a simple strong-teacher baseline under the same token and rollout budget. Without an equal-budget comparison, the gains may come from more teacher calls and more student rollouts rather than from the mixture-of-skills design. Since the central claim is the joint optimization of a skill library and router is better than single-skill optimization.

  2. lack of routing analysis. The router is a central component, but accuracy and failure modes are not reported. The credit assignment step is also central, but still the reliability is not measured. These make it hard to tell whether the reported improvements are robust or reflect noise, validation overfitting, or unequal compute. Also the paper reports point estimates without seeds, variance, or significance tests. Several benchmarks are small.

Questions for authors
  1. Can you provide an equal-budget comparison? Specifically, compare EvoMoSkill with a strong single-skill baseline that uses the same teacher model and the same total number of teacher tokens and student rollouts. Report cost-performance curves.

  2. How is the router implemented? Is it a prompted LLM that reads scope descriptions, or a trained classifier? What is its routing accuracy on validation tasks? What fraction of tasks are misrouted, and how does misrouting affect final success?

  3. Can you report multi-seed results, and significance tests for the main? For small benchmarks such as SkillLearnBench, how many tasks change between the best baseline and EvoMoSkill?

  4. How accurate is the teacher's failure diagnosis? Can you provide a human-annotated sample or an independent evaluation of the credit assignment labels? What happens when the diagnosis is wrong?

9 comments
AuthorsOpenReview ↗

Responses to Reviewer CVEN (1)

Responses to Weakness 1 and Question 1

Thank you for your valuable comments!

The gap between the claimed contribution and the evidence. The method is essentially a strong teacher LLM that edits skill text and router scopes, wrapped in a multi-stage acceptance pipeline. The paper does not show that this pipeline outperforms a simple strong-teacher baseline under the same token and rollout budget. Without an equal-budget comparison, the gains may come from more teacher calls and more student rollouts rather than from the mixture-of-skills design. Since the central claim is the joint optimization of a skill library and router is better than single-skill optimization.

Can you provide an equal-budget comparison? Specifically, compare EvoMoSkill with a strong single-skill baseline that uses the same teacher model and the same total number of teacher tokens and student rollouts. Report cost-performance curves.

We appreciate your questions about computational budgets and performance gains. Figure 3 and Appendix Figure A1 in the submission show that continued skill optimization raises the attainable performance level. In particular, EvoMoSkill outperforms other iterative skill methods on ALFWorld at similar or lower token cost. We added the following comparisons to test whether these gains persist with the same strong teacher and the same training budget.

1. Performance under matched training budgets (Table R2)

To isolate the contribution of the skill mixture under a shared training budget, we added Single, a single-skill control using the same strong teacher. Single uses the same gpt-5.6-sol teacher and CreAGD optimization procedure as EvoMoSkill and can revise skills based on execution failures. It always retains only one shared skill and does not create experts. We also compare SkillOpt, a single-skill optimizer, and the dynamic skill-library methods EvoSkill and Memento-Skills. All methods use the same teacher, student model, training and test splits, and evaluator within each setting. EvoSkill retains task-specific skill selection, and Memento-Skills retains its native retrieval and update procedure.

We first accumulate the tokens consumed by each method's native training procedure, including teacher feedback, student execution, routing, skill selection, and its own evaluation calls. We then define a shared budget cap B separately for EvoMoSkill and each baseline. For each method, we take the maximum training cost across its three seeds, then use the smaller of the two methods' maxima. This keeps the comparison within the available training range of both methods. At 25%, 50%, 75%, and 100% of B, we consider only checkpoints whose cumulative cost is within the cap, select the best by validation performance, and evaluate it on the same test set. Table R2 reports test means and standard deviations across three seeds. Each adjacent pair shares the same budget cap.

These comparisons match total native training-token caps. Teacher and student resources follow each method's own allocation procedure, and additional offline validation and testing are excluded. Actual teacher-token consumption, student executions, and candidate evaluations are reported in Table R3.

Table R2. Test success under matched training-token caps (mean±SD, %). B is in millions of tokens. Bold marks the higher mean at each budget point within each adjacent matched-cap pair.

BenchmarkModelMethodB (M)25% B50% B75% B100% B
ALFWorldQwen3-32BEvoMoSkill45.1866.42±3.7370.90±4.4870.90±2.9978.36±9.87
Single45.1855.22±0.0060.95±6.3567.41±4.8076.12±4.89
EvoMoSkill126.8370.90±3.7378.86±11.9983.58±7.2086.57±3.73
SkillOpt126.8372.64±4.3774.13±1.5575.12±6.6976.12±6.38
EvoMoSkill28.4760.45±3.7366.42±3.7366.42±2.2470.90±2.24
EvoSkill (task-wise selection)28.4766.67±4.5069.40±4.1672.14±4.8674.63±4.89
EvoMoSkill130.3670.90±2.2478.61±3.0283.33±7.1786.57±3.73
Memento-Skills130.3648.76±5.9957.71±6.7765.92±4.9775.12±5.65
DeepSeek-V4-FlashEvoMoSkill36.3857.96±2.6270.90±0.7575.37±2.2485.57±1.88
Single36.3864.68±3.3770.90±1.9776.37±1.5582.84±2.24
EvoMoSkill71.7070.40±0.8685.82±0.7585.82±1.4985.82±0.75
SkillOpt71.7062.94±8.2571.14±8.3976.37±4.5076.87±4.54
EvoMoSkill19.2257.71±0.8658.46±2.2870.40±0.8670.90±1.29
EvoSkill (task-wise selection)19.2270.15±2.9973.88±2.6976.37±2.8377.61±2.99
EvoMoSkill88.8081.84±1.1485.82±0.7585.57±1.1490.30±1.29
Memento-Skills88.8055.97±2.9963.18±3.1167.91±2.9979.35±2.28
SearchQAKimi-K2.6EvoMoSkill4.5280.69±1.2278.83±0.9378.83±1.1581.04±0.73
Single4.5279.26±1.0079.90±1.8681.90±1.9983.29±1.91
EvoMoSkill36.1280.85±1.1583.33±1.9683.33±1.0883.48±1.04
SkillOpt36.1282.45±0.4883.67±0.5783.67±1.0485.62±0.61
EvoMoSkill8.5378.83±0.0080.97±1.0081.76±0.9381.90±1.78
EvoSkill (task-wise selection)8.5381.47±2.1582.52±1.0483.24±0.9784.53±0.97
EvoMoSkill15.8780.97±1.1481.76±1.0780.81±0.8980.95±0.89
Memento-Skills15.8779.97±0.7280.52±1.2580.42±0.2382.62±1.30
DeepSeek-V4-FlashEvoMoSkill4.0377.57±0.0077.57±0.8279.38±2.1080.81±0.79
Single4.0377.62±0.6678.31±1.1578.79±0.5481.69±0.54
EvoMoSkill32.8180.79±0.0782.38±0.7983.64±0.8683.64±0.89
SkillOpt32.8181.69±0.9682.10±0.4982.14±0.3782.79±0.21
EvoMoSkill8.1477.57±0.7980.90±0.8280.79±2.1480.79±2.14
EvoSkill (task-wise selection)8.1481.86±1.5083.43±1.2785.17±0.5987.14±0.80
EvoMoSkill17.8680.86±0.9380.79±2.1482.24±0.4483.69±0.82
Memento-Skills17.8677.43±2.3579.21±0.4380.14±1.0080.93±1.19

Table R2 shows three findings.

  1. With the same strong teacher and training budget, the skill mixture still outperforms strong single-skill optimization. On ALFWorld with Qwen3-32B, EvoMoSkill and Single reach 78.36% and 76.12%, respectively, under the same 45.18M-token cap. With DeepSeek-V4-Flash at 36.38M tokens, they reach 85.57% and 82.84%. Relative to SkillOpt, the gains at the matched caps are 10.45 and 8.95 percentage points for the two students. These results show that the ALFWorld gains persist after controlling total training expenditure, supporting the value of shared and specialized skills in using feedback from the same teacher.
  2. Continued library optimization produces larger gains than revising a single skill. On ALFWorld with Qwen3-32B, EvoMoSkill improves from 70.90% to 86.57% as the budget increases, a gain of 15.67 percentage points. SkillOpt improves from 72.64% to 76.12%, a gain of 3.48 points. With DeepSeek-V4-Flash, EvoMoSkill reaches 85.82% at 50% of B, already exceeding SkillOpt's 76.87% at 100% of B. In these two settings, the skill mixture both reaches higher performance with continued optimization and surpasses the single-skill baseline at a lower budget.
  3. The budget-efficiency advantage also appears on SearchQA with DeepSeek-V4-Flash. EvoMoSkill reaches 83.64% at only 75% of the shared budget with SkillOpt, exceeding SkillOpt's 82.79% at 100%. Under the same 17.86M-token cap as Memento-Skills, success increases from 80.93% to 83.69%. Thus, in this question-answering setting, the gains over both single-skill optimization and dynamic skill memory remain under a shared budget.

2. Breakdown of training costs (Table R3)

To clarify where computation is spent, we also measured resource use over complete training runs. In Table R3, Teacher tokens and Student tokens count tokens consumed by the teacher and student executions. Student rollouts counts complete task executions. Scored proposals (C) counts candidate proposals that receive a score, and Full-val calls (V) counts evaluations of the entire validation set within the native procedure. One candidate can require multiple task executions, so these counts describe different aspects of the workload. All entries are means across seeds except the maximum total-token count. C and V can therefore be fractional.

Table R3. Native training costs and evaluation calls. Entries are means across seeds except maximum total tokens. M denotes one million tokens.

BenchmarkModelMetricEvoMoSkillSingleSkillOptEvoSkillMemento-Skills
ALFWorldQwen3-32BTeacher tokens (M)0.980.502.070.035.46
Student tokens (M)127.0939.87118.7425.94116.54
Total tokens: mean (M)128.8140.45120.8126.11122.31
Total tokens: max (M)133.2745.18126.8328.47130.36
Teacher calls357362223211,035
Student rollouts2,3878032,189731593
Wall time (h)22.03.620.44.718.3
Scored proposals (C)31.674.0028.6710.00187.00
Full-val calls (V)4.004.0026.6711.005.00
DeepSeek-V4-FlashTeacher tokens (M)0.620.371.010.033.98
Student tokens (M)84.4035.9465.9017.3691.60
Total tokens: mean (M)85.6036.3666.9117.4695.74
Total tokens: max (M)88.8036.3871.7019.2295.77
Teacher calls22328421321716
Student rollouts1,7861,0371,387664497
Wall time (h)22.56.420.45.815.2
Scored proposals (C)29.334.0024.6710.00178.33
Full-val calls (V)4.004.008.6711.005.00
SearchQAKimi-K2.6Teacher tokens (M)0.490.472.630.072.85
Student tokens (M)32.983.6231.056.7610.30
Total tokens: mean (M)36.994.2033.688.0115.21
Total tokens: max (M)40.764.5236.128.5315.87
Teacher calls17818554531841
Student rollouts9,7631,5279,5833,8003,986
Wall time (h)15.52.714.02.613.3
Scored proposals (C)23.6711.6739.3315.00195.67
Full-val calls (V)4.0013.6732.6716.005.00
DeepSeek-V4-FlashTeacher tokens (M)0.590.462.580.073.90
Student tokens (M)30.143.1628.066.8012.76
Total tokens: mean (M)35.483.6930.647.7317.34
Total tokens: max (M)36.394.0332.818.1417.86
Teacher calls21717254835963
Student rollouts9,8271,5799,4734,2803,991
Wall time (h)4.60.74.01.511.8
Scored proposals (C)28.6713.0039.3317.00311.67
Full-val calls (V)4.0014.0032.0018.005.00

The gains over SkillOpt in Table R2 do not require more teacher tokens or more frequent full-validation calls. EvoMoSkill uses fewer teacher tokens over complete training runs in all four settings. On ALFWorld with Qwen3-32B, the counts are 0.98M for EvoMoSkill and 2.07M for SkillOpt, so EvoMoSkill uses approximately 47% as many teacher tokens. With DeepSeek-V4-Flash, the counts are 0.62M and 1.01M. The corresponding matched-budget gains in Table R2 remain 10.45 and 8.95 percentage points. The skill mixture therefore achieves better performance with less teacher-generated text.

EvoMoSkill allocates more optimization computation to student execution and local candidate checks. On ALFWorld with Qwen3-32B, EvoMoSkill and Single make 357 and 362 teacher calls, respectively, but 2,387 and 803 student executions, and score 31.67 and 4.00 proposals. A similar number of teacher calls supports exploration of more candidates, accompanied by greater investment in execution-based checks. Compared with SkillOpt, EvoMoSkill scores a similar number of candidates, 31.67 versus 28.67, while reducing full-validation calls from 26.67 to 4.00. Paired executions within the relevant scope allow the optimizer to assess local revisions before retaining candidates that pass validation. We will include these cost and performance results in the revision. Thank you again for your suggestions.

Reviewer_pxL2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper propsoes EvoMoSkill to adapt frozen LLM agents by evolving an external library containing a shared skill, task-family-specific expert skills, and a router that selects at most one expert per task. The optimizer CreAGD uses teacher–student trajectory comparisons to assign failures to skill content, missing expertise, or routing scope. It generates candidate revisions, evaluates them against the previous system, applies history-based and validation-based acceptance checks, and retains rejected proposals as optimizer memory. Experts can be admitted, refined, and retired. Experiments with five LLMs on five benchmarks demonstrate state-of-the-art performance.

Critical strengths

he draft presents a promising system with substantial empirical evaluation. It evaluates five student models on five benchmarks spanning interactive environments, question answering, spreadsheet manipulation, and reusable task skills. It reports the best result outright in 23 of 25 benchmark–student settings, with average improvements of 4.0–6.8 points over the strongest competing baseline for each student. It provides strong evidence that external skill-library optimization can be practically useful across different agent capabilities and task interfaces. The acceptance-check ablations are also informative to investigates various design choices rather than reporting only aggregate gains.

Critical weaknesses

The results show that the complete system works well, but less clearly establish which new methodological idea accounts for its advantage over a well-tuned, budget-matched library optimizer. The cited EvoSkill already creates and revises reusable skills with validation-based selection, while Memento-Skills combines an adaptive skill library with a trainable router. So the contribution of the draft should go beyond evolving multiple skills or jointly adapting selection and content.

The draft does not clearly establish whether the combined system, including simultaneously accepted edits and routing changes, is subsequently checked. Section 3.3 evaluates candidate skill edits individually against the previous library, while routing-scope revisions enter the next step without additional rollouts. Those details affect whether the acceptance checks actually protect the deployed system, which need to be clarified. Please be aware that individually beneficial changes may not remain beneficial when combined.

Questions for authors

Does the EvoSkill baseline retain its native multi-skill creation and selection behavior? For the main baselines and static-partition ablation, what are the total teacher tokens, student rollout tokens, candidate evaluations, and validation calls? Does EvoMoSkill retain its advantage under matched total optimization budgets? When multiple skills and routing scopes change in one step, is the final combined library/router evaluated before deployment? What rollback mechanism applies to a harmful routing-only update?

8 comments
AuthorsOpenReview ↗

Responses to Reviewer pxL2 (1)

Responses to Weakness 1

Thank you for your valuable comments!

The results show that the complete system works well, but less clearly establish which new methodological idea accounts for its advantage over a well-tuned, budget-matched library optimizer. The cited EvoSkill already creates and revises reusable skills with validation-based selection, while Memento-Skills combines an adaptive skill library with a trainable router. So the contribution of the draft should go beyond evolving multiple skills or jointly adapting selection and content.

We appreciate your suggestions on the methodological contribution and ablation studies. EvoMoSkill introduces CreAGD, which integrates failure attribution, specialization within the skill library, and execution-based checks tailored to each update's scope. We clarify the differences from prior methods and evaluate this mechanism through controlled modular ablations.

1. Direct comparison with related methods

The central contribution is CreAGD: it assigns failure evidence to specific skill text or routing scopes within a library of shared and specialized skills, then evaluates revisions on the tasks that those skills serve. This involves two connected decisions. The first is what to change. Problems shared across families are assigned to the shared skill, family-specific procedural failures to the corresponding expert, coverage gaps to a new expert, and incorrect expert selection to a scope revision. The second is how to decide whether to retain the change. Shared-skill candidates are checked for regressions across the entire batch. Revisions to existing experts are evaluated through paired executions against the old library on their target family. New experts are tested on family tasks actually selected by their scope, with an additional check for incorrectly attracting tasks from other families. Expert candidates also undergo calibration against historical performance and held-out validation, so that an unusually weak execution by the old library does not create a spurious apparent gain.

The value of this mechanism is that family-specific repairs have a dedicated place, shared knowledge remains reusable, and the tasks used for acceptance match the intended scope of each update. For heterogeneous tasks that require different procedures, the library can therefore refine how knowledge is specialized and shared as it accumulates new knowledge. The differences from three closely related methods are as follows.

  1. SkillOpt optimizes a single domain skill. EvoMoSkill additionally optimizes the division of responsibilities among skills. SkillOpt already uses bounded text edits, validation gates, and feedback on rejected edits to produce one general-purpose skill document. EvoMoSkill organizes the same set of failures into revisions for different responsible components. A repair specific to one task family can be assigned to its expert, while changes to shared rules are checked across families. Each task uses the shared skill and at most one expert, so this division also determines which learned procedures enter the execution context.
  2. EvoSkill already diagnoses failures and decides whether to create or edit a skill. EvoMoSkill further connects responsibility assignment to the scope of acceptance tests. EvoSkill's proposer identifies capability gaps, its builder creates or edits skills, and its candidate pool is maintained using held-out validation of complete candidate programs. CreAGD explicitly distinguishes shared rules, family-specific procedures, coverage gaps, and routing errors. It selects paired evaluation tasks and regression checks according to the component being updated. An expert-text revision is first evaluated on its target family, whereas a shared-rule revision is checked across the entire batch. Local expert repairs and global shared-skill repairs thus receive acceptance tests appropriate to their scope.
  3. Memento-Skills already performs skill-level credit assignment. EvoMoSkill extends responsibility assignment to shared skills, experts, and routing scopes, with corresponding task-level acceptance tests. Memento-Skills identifies the skill most responsible for a failure, makes a targeted file revision, and checks the update using generated unit tests. Its Read–Write framework also includes a trainable behavioral router. EvoMoSkill further distinguishes incorrect skill content from selection of an unsuitable expert, using the latter to revise natural-language scopes. Skill-text candidates undergo repeated paired executions against the old system on the same tasks to assess local gains and cross-family regressions. The contribution is this specific mechanism for responsibility assignment, shared and local updates, and scope-dependent acceptance.

2. Ablating the key components (Table R1)

To identify which components drive the gains, we added ablations with the same library structure and family information. On ALFWorld with Qwen3-32B, the full method and M0, M1, and M2 share the same shared-skill and expert structure, family partition, expert-count limit, candidate-count settings, acceptance procedure, and training schedule. Each variant is trained from scratch, selects its checkpoint on validation data, and is evaluated on the same test tasks. We change only two settings: whether the optimization teacher receives credit-assignment feedback organized by the responsible component, and whether an expert's scope can change after creation. M0 disables both. M1 allows scope evolution but gives the teacher only raw failure trajectories. M2 retains credit assignment while fixing scopes. We also compare the full method with M0 on SearchQA using Kimi-K2.6. The intervals in Table R1 estimate the full method's success rate minus that of the comparator, and p-values come from paired McNemar tests.

Table R1. Modular ablations with the same task-family information. Bold marks the highest success rate within each setting. CIs estimate EvoMoSkill minus the comparator.

BenchmarkModelMethodScope updatesCreditSuccess (%)95% CI (pp)p
ALFWorldQwen3-32BEvoMoSkillEnabledEnabled86.57——
M0FixedDisabled85.07[−2.08,+5.07]0.6875
M1EnabledDisabled77.61[+2.59,+15.32]0.0118
M2FixedEnabled84.33[−1.61,+6.09]0.4531
EvoSkill (task-wise selection)——75.37——
SearchQAKimi-K2.6EvoMoSkillEnabledEnabled84.69——
M0FixedDisabled83.98[+0.19,+1.24]0.0129

First, credit assignment makes the largest contribution in this ablation. With scope updates enabled in both variants, the full method reaches 86.57%, whereas M1 without credit assignment reaches 77.61%, a decrease of 8.96 percentage points. The 95% CI is [+2.59, +15.32], with p = 0.0118. Both retain the same multi-expert structure, teacher, and candidate acceptance procedure. The difference is whether the teacher receives diagnosis organized by the component to update or unassigned raw failure trajectories. Identifying which skill to revise and what to repair therefore significantly improves task success after skill optimization.

M0 with fixed specialization already reaches 85.07%, whereas M1, which permits scope changes without attribution feedback, reaches only 77.61%. Together with the full method's 86.57%, this shows that allowing more types of edits does not automatically improve performance. Effective optimization depends on choosing the update target based on the cause of failure. By separating content errors from routing errors, CreAGD enables revisions that address the diagnosed failure.

Second, combining credit assignment with scope updates also outperforms fixed specialization on question answering. On SearchQA with Kimi-K2.6, the full method achieves 84.69%, compared with 83.98% for M0, with a 95% CI of [+0.19, +1.24] and p = 0.0129. Thus, with the same family information and expert capacity, targeted content and scope revisions provide additional gains. We will include these methodological comparisons and ablation results in the revision. Thank you again for your suggestions.

AuthorsOpenReview ↗

Responses to Reviewer pxL2 (2)

Responses to Question 1

Thank you for your valuable comments!

Does the EvoSkill baseline retain its native multi-skill creation and selection behavior? For the main baselines and static-partition ablation, what are the total teacher tokens, student rollout tokens, candidate evaluations, and validation calls? Does EvoMoSkill retain its advantage under matched total optimization budgets?

We appreciate your questions about baseline implementation and optimization budgets. We added baseline runs that retain native skill selection and compared performance under the same training-token caps.

1. Dynamic skill-library baseline implementations

EvoSkill's multi-skill creation and task-specific selection. The submitted implementation retained the ability to create and edit multiple skills, but concatenated the available skills during execution rather than fully preserving task-specific selection. The additional experiments enable this selection mechanism. Each skill has a description, and the student selects zero or more skills for the current task. Skill creation, editing, and native early stopping are retained. The programs ultimately selected in these runs contain only one skill as an outcome of evolution and selection, not because the library was restricted to one skill in advance. EvoSkill in Tables R2 and R3 uses this task-specific selection implementation. Memento-Skills uses the public v0.4.0 Read–Write implementation, retaining hybrid retrieval, student skill selection, and its update checks.

2. Performance under matched training budgets (Table R2)

To compare performance at the same training cost, all methods use the same teacher, gpt-5.6-sol, and the same student model, data split, and evaluator for each setting. Single uses EvoMoSkill's optimization procedure but allows only one shared skill. SkillOpt is a single-skill optimization baseline. We count tokens used by the teacher, student executions, routing, skill selection, and evaluations within each method's native optimization procedure. Additional offline validation and testing are excluded. For each method pair, we take each method's maximum cumulative training cost across seeds and use the smaller of these two values as the shared cap B. At 25%, 50%, 75%, and 100% of B, we select the validation-best checkpoint among those within the cap and report test mean and standard deviation across three independent seeds. Each adjacent pair in Table R2 shares the same total-token cap. Teacher and student resources are allocated according to each method's native procedure.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.