acceptodds
Under review as a conference paper at ICLR 2027

Attribution-Guided Skill Evolution for Online Continual Agent Learning

Abstract

LLM agents increasingly face streams of tasks where experience from earlier tasks should be reused while adapting to new ones. We study this setting as online continual agent learning, where a shared external skill library is updated sequentially from completed tasks. Existing methods leverage agent experience to refine reusable skills through reflection, feedback, or reinforcement learning, but task-level outcomes provide limited insight into individual skill contributions. Failures may stem from missing activation, non-adherence, or incorrect application, while task success does not guarantee that every invoked skill is useful, making direct feedback-driven updates prone to inappropriate modifications. We propose AGSE (Attribution-Guided Skill Evolution), a plug-and-play framework that attributes skill–trajectory interactions along five dimensions: activation, adherence, correctness, coverage, and precision. These attributions provide structured, skill-specific evidence for guarded skill repair, expansion, and scope refinement, without model-weight updates or policy training. Across three heterogeneous CLI-agent frameworks, AGSE achieves exact accuracy of 72.9%, 59.4%, and 48.9% on OfficeQA Pro, and completion rates of 59.64%, 49.40%, and 48.19% on SkillFlow. In continual learning evaluation, AGSE achieves an FS of 67.7, FWT of +0.0227, and forgetting rate of 13.95%. Ablations demonstrate the importance of both multidimensional attribution and continual skill evolution.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
LLM, Agent, Continual Learning, Skill evolution

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_nXY4OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates continual agent learning with an online learning setting. Specifically, this paper proposes AGSE, which mainly consists of (1) skill-use attribution for more fine-grained credit assignment, (2) evidence-weighted utility updates for robust estimation of the skill utility, (3) guarded lifecycle evolution for selective triggering. Extensive experiments on OfficeQA Pro and SkillFlow demonstrate the effectiveness of the proposed framework.

Critical strengths
  • This paper does a strong engineering work.
  • Extensive harnesses are evaluated to demonstrate the general applicability of the proposed method.
Critical weaknesses
  • The benchmarks (OfficeQA Pro & SkillFlow) are overly small, with merely ~100 tasks, which cannot support the verification of “agent continual learning” in online stream.
  • Online agent continual learning (similar to RL) presents much higher variance in the final results than the traditional training-testing setting. Please consider report results with at least 5 repeat runs. Otherwise, the results can lead to biased findings.
  • Main results in Table 2 seems not convincing enough. Both baselines (Skill-Creator & SkillFlow-Evo) perform much worse than w/o skill, which makes me question the fairness in the empirical evaluation.
  • As shown in Table 6, tt is a very sensitive hyper-parameter. Setting t=2 or t=8 leads to significant deterioration, which makes the algorithm fragile. Meanwhile, Skill-Creator & SkillFlow-Evo are set with t=4 by default without careful hyper-parameter tuning, which may lead to unfair comparison.
  • The technical novelty of this paper is limited. Although the skill-use attribution seems to be well-motivated, evidence-weighted utility updates & guarded lifecycle evolution are just heuristic design choices without any empirical or theoretical insights.
Questions for authors

None

7 comments
AuthorsOpenReview ↗

Response to Reviewer nXY4(1/3)

We thank the reviewer for the thoughtful and constructive feedback. We respectfully clarify several concerns regarding benchmark scale, experimental reliability, baseline fairness, hyperparameter sensitivity, and technical novelty. To address these concerns, we conducted additional experiments on OfficeQA Pro using GPT-5.6-Sol due to resource constraints, including repeated runs, reversed task-order evaluation, baseline hyperparameter sensitivity analysis.

W1. The benchmarks (OfficeQA Pro & SkillFlow) are overly small, with merely ~100 tasks, which cannot support the verification of “agent continual learning” in online stream.

A1: We agree that evaluating online continual agent learning with larger and more diverse task streams is important. However, the scale of agent benchmarks should also be considered in light of the time cost and complexity of individual tasks.

In particular, each OfficeQA Pro task requires approximately 7–8 minutes of agent execution and consumes around 1M tokens on average. The benchmark contains 133 questions grounded in approximately 89K pages of Treasury Bulletins and over 26M numerical values, requiring substantial document retrieval, numerical reasoning, and multi-step problem solving. Consequently, scaling such evaluations to thousands of tasks incurs substantial time costs, especially when each task must be executed sequentially on continual learning settings to evaluate the evolution and reuse of skills.

We selected OfficeQA Pro and SkillFlow as our evaluation benchmarks because they involve challenging agent tasks that require substantial problem-solving capabilities and provide opportunities for cross-task skill reuse. In our evaluation, the agent incrementally updates a shared skill library after completing each task, allowing subsequent tasks to benefit from previously accumulated experience. We acknowledge that the number of tasks in these benchmarks limits the extent to which our findings can be generalized to substantially longer task streams, and will clarify this limitation in the paper.

W2.Online agent continual learning (similar to RL) presents much higher variance in the final results than the traditional training-testing setting. Please consider report results with at least 5 repeat runs. Otherwise, the results can lead to biased findings.

A2: We would first like to clarify why repeated experiments were not included in the original submission. Our evaluation involves long-horizon LLM agent interactions, where each task requires multiple rounds of tool execution, reasoning, and skill utilization. In particular, a single OfficeQA Pro task typically takes approximately 7–8 minutes to complete and consumes around 1M tokens on average. Consequently, evaluating the entire benchmark already incurs substantial time costs. Nevertheless, we agree that repeated experiments are important for assessing the robustness of our method. We conducted three additional repeated runs for all methods on OfficeQA Pro. All additional experiments reported in this response were conducted using GPT-5.6-sol. The results are summarized below.

MethodRun 1Run 2Run 3Mean ± Std
w/o Skill62.463.962.462.9 ± 0.87
SkillFlow-Evo65.465.463.264.7 ± 1.30
Skill-Creator65.463.966.965.4 ± 1.50
AGSE67.770.768.468.9 ± 1.57

AGSE consistently outperforms the w/o Skill baseline across all three runs, with an average improvement of 6.0 percentage points. These results provide additional evidence that the observed gains are not attributable to a single run.

We also conducted a task-order robustness check by reversing the original task order. AGSE obtains 67.8% accuracy under the reversed order, compared with 68.9% under the original order. The relatively small difference suggests that the performance is not strongly dependent on the particular task order evaluated in this additional check.

AuthorsOpenReview ↗

Response to Reviewer nXY4(2/3)

W3. Main results in Table 2 seems not convincing enough. Both baselines (Skill-Creator & SkillFlow-Evo) perform much worse than w/o skill, which makes me question the fairness in the empirical evaluation.

A3: We respectfully clarify that the claim that both skill-evolution baselines consistently underperform the w/o Skill baseline does not accurately reflect the results in Table 2. For example, on Qwen Code (Qwen3.6 Plus), Skill-Creator achieves 47.4% on OfficeQA @0, outperforming w/o Skill (44.4%), and 51.2% on OfficeQA @0.001, slightly exceeding w/o Skill (51.1%). On Codex (GPT-5.5), SkillFlow-Evo achieves 71.4% on OfficeQA @0.01, outperforming w/o Skill (70.7%). On SkillFlow, both Skill-Creator (44.31%) and SkillFlow-Evo (43.37%) outperform w/o Skill (42.17%) in completion rate under Qwen Code. More importantly, the effectiveness of skill evolution is not guaranteed merely by creating, modifying, or using skills.

This is also supported by recent empirical studies. SkillsBench (Li et al., 2026) reports that self-generated skills provide no average benefit, with 16 of 84 tasks exhibiting negative performance changes even when skills are provided. Similarly, SkillFlow (Zhang et al., 2026) finds that lifelong skill evolution yields substantially different outcomes across models: while Claude Opus 4.6 improves by 8.43 percentage points, Kimi K2.5 gains only 0.60 points despite a high skill usage rate, and Qwen-Coder-Next even regresses relative to its vanilla setting. These findings demonstrate that skill evolution can introduce ineffective or harmful modifications, and that skill usage alone is not a reliable indicator of utility.

This is precisely the problem AGSE aims to address: rather than assuming that every skill update is beneficial, AGSE identifies how skills are used and whether they contribute effectively to task completion, then uses the resulting evidence to guide selective evolution.

W4. As shown in Table 6, tt is a very sensitive hyper-parameter. Setting t=2 or t=8 leads to significant deterioration, which makes the algorithm fragile. Meanwhile, Skill-Creator & SkillFlow-Evo are set with t=4 by default without careful hyper-parameter tuning, which may lead to unfair comparison.

A4: We agree that the evolution interval (tt) affects performance, since it determines how frequently accumulated experience is translated into skill updates. To examine whether the baseline configurations were fairly selected, we conducted an additional sensitivity analysis for both Skill-Creator and SkillFlow-Evo on OfficeQA Pro using GPT-5.6-sol.

Methodt=2t=4t=8
SkillFlow-Evo60.964.761.7
Skill-Creator62.465.463.9
AGSE64.268.964.9

The results show that (tt=4) achieves the best performance among the tested configurations for both baselines. Importantly, both baselines exhibit performance changes under different evolution intervals, indicating that this sensitivity is not unique to AGSE. It reflects the general trade-off between accumulating sufficient experience and updating the skill library frequently enough to benefit subsequent tasks.

We will include these additional results and clarify the hyperparameter selection procedure to make the comparison more transparent.

AuthorsOpenReview ↗

Response to Reviewer nXY4(3/3)

W5. The technical novelty of this paper is limited. Although the skill-use attribution seems to be well-motivated, evidence-weighted utility updates & guarded lifecycle evolution are just heuristic design choices without any empirical or theoretical insights.

A5: We appreciate the reviewer’s comments on the technical novelty. We would like to further clarify the central contributions of AGSE and the importance of the problem it addresses.

As LLM agents increasingly operate over streams of related tasks, enabling them to continuously accumulate, retain, and reuse experience becomes an important direction toward more capable and adaptive agents. Rather than treating each task as an isolated problem, continual learning agents should be able to improve their reusable capabilities based on previous interactions. However, an important challenge is how to turn task execution experience into meaningful and actionable updates to the agent's knowledge and skills.

AGSE addresses this challenge by introducing attribution-guided skill evolution, which gives skill self-evolution an explicit direction rather than relying on undirected reflection or task-level success/failure signals. Its central contribution is a five-dimensional skill-use attribution mechanism that distinguishes Activation, Adherence, Correctness, Coverage, and Precision. These dimensions identify different aspects of skill usage and provide actionable evidence about why a skill may fail to contribute effectively to task completion. For example, a skill that is not activated requires a different response from one that is activated but contains incorrect guidance or insufficient coverage.

Building on this attribution mechanism, AGSE integrates two complementary components:

ComponentRole in continual skill evolution
Evidence-weighted utility updatesAccumulates dimension-specific feedback across tasks to track the evolving utility of each skill, reducing reliance on individual task outcomes.
Guarded lifecycle evolutionUses the attributed evidence and utility changes to guide selective skill preservation, repair, expansion, scope adjustment, and other lifecycle operations.

The key contribution is therefore not simply to make agents revise their skills, but to enable them to understand how their existing skills are used, identify what needs to improve, and evolve skills in a corresponding direction over time. This provides a structured mechanism for connecting task-level experience with skill-level improvement in online continual agent learning.

Reviewer_fST8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces AGSE, a skill evolution framework designed for the online continual learning of LLM agents. Addressing the limitation that existing methods struggle to pinpoint specific skill issues based on task-level feedback, the authors propose a five-dimensional attribution mechanism that analyzes skill execution across five aspects: Activation, Adherence, Correctness, Coverage, and Precision. AGSE accumulates historical feedback through an evidence-weighted utility update mechanism and triggers constrained skill modifications when utility declines. Experiments conducted on OfficeQA Pro and SkillFlow demonstrate that the method improves task performance and achieves favorable results on continual learning metrics, such as forward transfer and forgetting rates.

Critical strengths

S1. The research motivation is clear, and the design for attributing skill failures offers practical value.

The paper notes that task-level feedback struggles to distinguish between issues such as skills failing to activate, skills not being followed, and content errors. It proposes a five-dimensional attribution mechanism to guide targeted revisions, thereby helping to avoid unnecessary skill rewriting.

S2. The experimental results are significant.

AGSE achieved notable performance improvements across two benchmarks and three agent frameworks. Ablation studies further demonstrate that multi-dimensional attribution and continuous skill evolution positively impact final performance.

Critical weaknesses

W1. Lack of direct validation regarding the effectiveness of Attribution-guided Skill Evolution.

The paper relies primarily on LLM-based assessments across five dimensions to evaluate skills; however, high inter-annotator agreement does not prove that the attribution results accurately reflect the actual causes of task failures. Furthermore, the absence of a systematic patch-level evaluation makes it difficult to determine whether individual skill modifications genuinely improved performance or compromised existing capabilities. Existing ablation studies also fail to isolate the specific contribution of the "Guarded Evolution" constraint. These issues undermine the persuasiveness of the paper's core methodology.

W2. Insufficient robustness validation for continual learning experiments.

Evaluation is conducted primarily using a fixed task sequence, lacking independent repeat experiments across multiple random seeds and varying task orders. Given the strong path dependency inherent in Skill Evolution, task-level bootstrap confidence intervals are inadequate to fully capture performance fluctuations resulting from different evolutionary trajectories. Consequently, the paper's conclusions regarding stable skill accumulation, forward transfer, and the mitigation of forgetting require further validation.

Questions for authors

Q1. Could the authors provide a performance comparison of the Skills before and after modification, along with the proportions of patches that yielded positive gains, showed no significant effect, or resulted in performance degradation?

Q2. Have the authors conducted ablation studies that retain the five-dimensional attribution but remove the "Guarded Editing" constraints, in order to verify the independent contribution of these constraints?

Q3. Did the authors repeat the experiments using different random seeds or task sequences? This is a crucial validation step in continual learning. If so, could the mean and variance for Accuracy, FWT, and Forgetting Rate be provided?

Q4. The paper triggers evolution based on the magnitude of the drop in Utility. How does the system handle a scenario where a Skill's Utility remains consistently low but does not exhibit a significant decline?

4 comments
AuthorsOpenReview ↗

Response to Reviewer fST8(1/3)

We thank the reviewer for the careful and constructive evaluation of our work. We are encouraged that the reviewer recognizes the motivation of AGSE, the practical value of five-dimensional skill attribution, and the improvements demonstrated across two benchmarks and three agent frameworks. We particularly appreciate the concerns regarding the direct validation of attribution-guided evolution and the robustness of continual learning under different execution trajectories.

We agree that these aspects are important for establishing the reliability of an online skill evolution framework. In response, we conducted additional experiments and analyses using GPT-5.6-sol because of resource constraints, including (1) a paired task-level comparison between AGSE and the agent without skills, (2) an ablation study removing Guarded Editing while retaining five-dimensional attribution, (3) three independent experimental runs, and (4) a reverse-order evaluation to examine task-order sensitivity. We address the questions individually below.

W1. Lack of direct validation regarding the effectiveness of Attribution-guided Skill Evolution. & Q1: Could the authors provide a performance comparison of the Skills before and after modification, along with the proportions of patches that yielded positive gains, showed no significant effect, or resulted in performance degradation?

We agree that inter-annotator agreement alone measures the consistency of attribution judgments rather than establishing their causal correctness. To provide additional outcome-based evidence, we conducted a paired comparison between AGSE and the w/o Skill baseline on 133 identical task UIDs, aligning their outcomes using exact accuracy.

The paired results are as follows:

OutcomeNumber of tasks
Both methods correct76
Both methods incorrect33
w/o Skill incorrect, AGSE correct21
w/o Skill correct, AGSE incorrect3

AGSE achieves 72.93% accuracy (97/133), compared with 59.40% (79/133) for w/o Skill, yielding an absolute improvement of 13.53 percentage points. More importantly, among the 24 tasks where the two methods produce different correctness outcomes, AGSE improves 21 tasks (87.5%) and regresses on 3. Overall, the number of incorrect answers decreases from 54 to 36, corresponding to a 33.33% relative reduction in errors.

These paired results provide further evidence that the skill library maintained by AGSE delivers consistent task-level benefits, with substantially more improvements than regressions.

AuthorsOpenReview ↗

Response to Reviewer fST8(2/3)

W2. Insufficient robustness validation for continual learning experiments. & Q3. Did the authors repeat the experiments using different random seeds or task sequences? This is a crucial validation step in continual learning. If so, could the mean and variance for Accuracy, FWT, and Forgetting Rate be provided?

We appreciate the reviewer’s concern about path dependency in continual learning. Our original evaluation involves long-horizon agent interactions, in which each OfficeQA Pro task requires multiple rounds of reasoning, tool execution, and skill utilization. Each task takes approximately 7–8 minutes and consumes around 1 million tokens on average, making repeated evaluation over the entire task stream computationally expensive. This was the reason that the original submission primarily reported a fixed task sequence.

Nevertheless, we agree that repeated experiments are important for assessing the stability of continual skill evolution. We therefore conducted three independent runs using GPT-5.6-sol on OfficeQA Pro. The results are summarized below:

MethodRun 1Run 2Run 3Mean ± Std
w/o Skill62.463.962.462.9 ± 0.87
SkillFlow-Evo65.465.463.264.7 ± 1.30
Skill-Creator65.463.966.965.4 ± 1.50
AGSE67.770.768.468.9 ± 1.57

AGSE consistently outperforms the w/o Skill baseline across all three runs, with an average improvement of 6.0 percentage points. These results provide additional evidence that the observed gains are not attributable to a single run.

We also conducted a task-order robustness check by reversing the original task order. AGSE obtains 67.8% accuracy under the reversed order, compared with 68.9% under the original order. The relatively small difference suggests that the performance is not strongly dependent on the particular task order evaluated in this additional check.

AuthorsOpenReview ↗

Response to Reviewer fST8(3/3)

Q2: Have the authors conducted ablation studies that retain the five-dimensional attribution but remove the "Guarded Editing" constraints, in order to verify the independent contribution of these constraints?

To isolate the contribution of Guarded Editing, we conducted an additional ablation using GPT-5.6-sol on OfficeQA Pro that retains the five-dimensional attribution mechanism and utility-based evolution trigger, while removing the constraints imposed on the editing stage.

MethodAccuracy
w/o Guarded Editing66.9
AGSE68.9

AGSE outperforms the unconstrained editing variant by 2.0 percentage points. This result indicates that identifying a skill problem is not sufficient by itself: the editing process must also be constrained by the available attribution evidence. Without Guarded Editing, the evolution module has greater freedom to modify skill content, potentially introducing unnecessary or insufficiently supported changes. The improvement therefore supports the independent contribution of Guarded Editing in making skill evolution more reliable.

Q4: The paper triggers evolution based on the magnitude of the drop in Utility. How does the system handle a scenario where a Skill's Utility remains consistently low but does not exhibit a significant decline?

The reviewer raises an important distinction between relative utility degradation and persistently low absolute utility. AGSE uses the degradation threshold to determine whether a skill should undergo evolution. However, a skill whose utility remains consistently low may not trigger evolution if its utility no longer decreases substantially.

To handle this case, our system additionally employs an absolute utility threshold of 0.3. If any dimension of a skill's utility falls below this threshold, the skill is no longer exposed to the agent for subsequent task execution. This operation prevents a persistently unreliable skill from continuing to affect task execution.

Thank you very much for reconsidering our submission and raising the score. We sincerely appreciate your time and constructive feedback. Your comments have helped us strengthen the presentation and empirical validation of our work.

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

This paper proposes AGSE, a framework for online continual agent learning through attribution-guided skill evolution. It evaluates skill–trajectory interactions along five dimensions: activation, adherence, correctness, coverage, and precision. These signals guide targeted skill updates through evidence-weighted utility estimation and constrained editing. Experiments across two benchmarks and three agent frameworks demonstrate improvements in task performance and knowledge retention.

Critical strengths
  • The five-dimensional attribution framework distinguishes different skill failure modes, enabling more targeted updates than conventional outcome-driven refinement.

  • The separation of attribution and skill editing provides a practical mechanism for maintaining reusable skills without model retraining.

  • The evaluation covers multiple agent frameworks and benchmarks, with ablation studies, human validation, and continual-learning metrics providing complementary empirical evidence.

Critical weaknesses
  • The performance gains are highly sensitive to the evolution interval, weakening the robustness claims. Table 6 shows that changing t=4t=4 to t=2t=2 or t=8t=8 reduces Codex accuracy from 72.9% to 60.2% or 60.9%, nearly eliminating the improvement over the 59.4% no-skill baseline. Evaluation under multiple task orders and random seeds is needed to establish whether the gains generalize beyond the selected configuration.
  • Appendix A.6 reports that only coverage crosses the evolution threshold in round 53, yet the resulting patch also corrects the identified Type-7 H-spread rounding error. This appears inconsistent with the core-content editing restriction in Appendix A.3, which requires explicit correctness-related authorization. The authors should clarify how this modification satisfies the proposed safeguards.
Questions for authors

see wekaness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer EqTc(1/2)

We sincerely thank the reviewer for the thoughtful and constructive feedback, as well as for recognizing the strengths of our work, particularly the multidimensional attribution mechanism and the separation between skill attribution and skill editing. We appreciate the reviewer's suggestions, which have helped us identify important aspects that require further validation and clarification.

In response, we have conducted additional experiments to examine the robustness of AGSE under repeated runs and different task orders. We have also revisited the evolution and editing procedure and will clarify the distinction between the condition that triggers an evolution round and the evidence required to authorize specific skill modifications. We address each concern below.

W1: Sensitivity to the Evolution Interval and Robustness. The performance gains are highly sensitive to the evolution interval, weakening the robustness claims. Table 6 shows that changing t=4 to t=2 or t=8 reduces Codex accuracy from 72.9% to 60.2% or 60.9%, nearly eliminating the improvement over the 59.4% no-skill baseline. Evaluation under multiple task orders and random seeds is needed to establish whether the gains generalize beyond the selected configuration.

A1: We would first like to clarify why repeated experiments were not included in the original submission. Our evaluation involves long-horizon LLM agent interactions, where each task requires multiple rounds of tool execution, reasoning, and skill utilization. In particular, a single OfficeQA Pro task typically takes approximately 7–8 minutes to complete and consumes around 1M tokens on average. Consequently, evaluating the entire benchmark already incurs substantial time costs. Nevertheless, we agree that repeated experiments are important for assessing the robustness of our method.

We agree that the evolution interval tt is an important design choice. It controls the trade-off between accumulating sufficient task-level evidence and adapting the skill library promptly. An overly short interval may lead to premature updates frequently based on limited evidence, whereas an overly long interval may delay useful skill improvements.

Following the reviewer's suggestion, we conducted three additional runs on OfficeQA. Due to computational resource constraints, all additional experiments reported in this response were conducted using GPT-5.6-sol. We report the mean accuracy and sample standard deviation across the three runs.

MethodRun 1Run 2Run 3Mean ± Std
w/o Skill62.463.962.462.9 ± 0.87
SkillFlow-Evo65.465.463.264.7 ± 1.30
Skill-Creator65.463.966.965.4 ± 1.50
AGSE(tt=2)63.963.964.764.2 ± 0.46
AGSE(tt=8)64.766.263.964.9 ± 1.17
AGSE(tt=4)67.770.768.468.9 ± 1.57

AGSE consistently outperforms the w/o Skill baseline across all three runs, with an average improvement of 6.0 percentage points. These results provide additional evidence that the observed gains are not attributable to a single run.

We also conducted a task-order robustness check by reversing the original task order. AGSE obtains 67.8% accuracy under the reversed order, compared with 68.9% under the original order. The relatively small difference suggests that the performance is not strongly dependent on the particular task order evaluated in this additional check.

AuthorsOpenReview ↗

Response to Reviewer EqTc(2/2)

W2: The Coverage-Triggered Evolution in the Round-53 Case Study Appendix A.6 reports that only coverage crosses the evolution threshold in round 53, yet the resulting patch also corrects the identified Type-7 H-spread rounding error. This appears inconsistent with the core-content editing restriction in Appendix A.3, which requires explicit correctness-related authorization. The authors should clarify how this modification satisfies the proposed safeguards.

A2: Thank you for pointing out the apparent inconsistency between the Round-53 case study in Appendix A.6 and the editing constraint described in Appendix A.3. We agree that the distinction between triggering an evolution round and authorizing a particular edit should be made more explicit.

In AGSE, the evolution threshold determines when to initiate an evolution round, rather than restricting the evolution process to modifying only the dimension that crosses the threshold. Importantly, we intentionally adopt threshold-based triggering to avoid overly frequent updates to the skill library. Since individual task outcomes may introduce noisy or transient changes in utility scores, triggering evolution after every task could lead to unnecessary modifications and destabilize previously acquired skills. The threshold therefore serves as a mechanism for accumulating sufficient evidence before initiating skill evolution. Once an evolution round is triggered, the evolution agent receives all available proposals and the corresponding changes in the skill's multidimensional utility scores. It can therefore consider evidence from multiple dimensions when deciding how to improve the skill.

In the Round-53 case, the decrease in Coverage exceeded the evolution threshold and triggered the evolution round. The agent then considered the available proposals and score changes, including the evidence related to the Type-7 H-spread rounding error. Thus, the fact that Coverage triggered the round does not, by itself, imply that the subsequent edit must be limited to Coverage.

Importantly, the evolution trigger and the authorization for editing core skill content serve different purposes: the former determines whether an evolution round should take place, while the latter constrains which modifications are permitted. We will revise the description of the case study and the corresponding methodological explanation to make this distinction explicit and avoid the impression that crossing the Coverage threshold automatically authorizes an unrestricted correctness-related edit.

Dear Reviewer EqTc,

Thank you for your time. We would appreciate it if you could take a moment to review our rebuttal and let us know whether our responses address your concerns.

Thank you again for your consideration.

Best regards, The Authors

Thank you to the authors for their detailed response, which has addressed some of my concerns. Taking into account the other reviewers' comments and the authors' rebuttal, I have decided to maintain my original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Summary of Contributions and Responses

Dear PCs, SACs, and ACs,

We sincerely thank all three reviewers for their thoughtful and constructive feedback. Below, we summarize the core contribution of our work and our responses to the major concerns.

Core contribution

AGSE (Attribution-Guided Skill Evolution) enables LLM agents to continuously improve reusable skills through online task experience. Its central contribution is a five-dimensional skill-use attribution mechanism that distinguishes different skill failure modes across Activation, Adherence, Correctness, Coverage, and Precision. Combined with evidence-weighted utility updates and guarded lifecycle evolution, AGSE translates task-level feedback into targeted, evidence-supported skill modifications, rather than relying on undirected reflection or task-level success/failure signals alone.

The main concerns raised by the reviewers and our responses are summarized as follows:

  1. Effectiveness and reliability of attribution-guided skill evolution.

    Response: We conducted additional paired task-level comparisons and ablation studies. The paired comparison shows that AGSE improves 21 tasks while regressing on only 3, compared with the agent without skills. We also isolate the contribution of Guarded Editing: removing its constraints reduces accuracy by 2.0 percentage points. These results provide further evidence that attribution-guided and constrained skill modifications contribute to more reliable skill evolution.

  2. Robustness to random runs and task order.

    Response: We conducted three independent runs across AGSE and the baselines, reporting mean accuracy and standard deviation. AGSE consistently outperforms the no-skill baseline across all three runs. We additionally evaluated AGSE under reversed task order and observed only a small performance difference, providing further evidence that its gains are not specific to a single execution trajectory.

  3. Sensitivity to the evolution interval and fairness of baseline comparisons.

    Response: We performed additional hyperparameter sensitivity analyses for both Skill-Creator and SkillFlow-Evo under the same evolution intervals. Both baselines also exhibit sensitivity to the evolution interval, with the intermediate setting achieving their best results. This suggests that the observed sensitivity reflects a general trade-off in continual skill evolution, rather than a limitation unique to AGSE.

  4. Handling persistently low-utility skills and controlling skill modifications.

    Response: We clarified the distinction between evolution triggering and edit authorization. The evolution threshold determines when an evolution round begins, while all available attribution evidence is considered to guide the subsequent modifications. We also introduce an absolute utility threshold: a skill whose utility falls below this threshold in any dimension is excluded from subsequent task execution, preventing persistently unreliable skills from being reused.

  5. Benchmark scale, empirical evaluation, and technical novelty.

    Response: We clarified the characteristics and computational costs of the evaluated benchmarks, as well as the challenges of conducting long-horizon online continual learning experiments. We also clarified the role of each component in AGSE: multidimensional attribution diagnoses skill-use issues, evidence-weighted utility updates accumulate skill-specific feedback, and guarded lifecycle evolution translates this evidence into targeted skill improvements. Together, these components provide a structured approach to online continual agent learning.

We sincerely appreciate the reviewers' time and consideration. Reviewer fST8 increased their score from 2 to 3 after reviewing our rebuttal. For the two reviewers who assigned a score of 2, we have carefully addressed each of their concerns through detailed point-by-point responses and additional empirical evidence. We hope our clarifications and additional results help the Area Chairs better assess the contributions and empirical support of AGSE.

We respectfully ask the Area Chairs to consider our responses and the additional evidence when making their final assessment.

Thank you for your time and consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.