acceptodds
Under review as a conference paper at ICLR 2027

EASE: Behavior-Adaptive Skill Curation for Self-Evolving Agents

Abstract

Agent skills provide a lightweight mechanism for self-evolving agents to accumulate reusable procedural knowledge without updating model parameters. However, existing learned skill curators typically optimize curation without explicitly modeling the behavior of the downstream executor. We show that this limitation can lead to systematic cross-executor degradation: curators optimized with different executors perform best when paired with their own training executor, suggesting that effective skill curation is executor-dependent. To address this challenge, we formulate behavior-adaptive skill curation and introduce , a framework that learns a single curator capable of adapting its curation decisions to different executor behaviors. EASE maintains an online behavioral profile summarizing recent execution patterns and conditions the curator on this profile, together with the current trajectory and retrieved skills, to dynamically add, modify, or remove skills from an evolving repository. We train the shared curator jointly across multiple frozen executors with reinforcement learning, using retrieval-aware and behavior-aware temporal attribution to focus optimization on curation actions with observable downstream influence. Across multiple public agentic benchmarks, including ALFWorld, ScienceWorld, and WebShop, and heterogeneous executor families ranging from Qwen3-8B/32B and GPT-OSS-120B to previously unseen Kimi K2.6, DeepSeek V4 Flash, and Gemini 3.5 Flash, EASE outperforms strong skill- and memory-based baselines without per-executor finetuning. Beyond task performance, behavior-adaptive curation also enables more efficient skill evolution: across the three benchmarks, EASE maintains % fewer skills, improves skill retrieval by % and measured edit utility by %, and reduces deployment-time inference tokens by %. These results establish behavior-adaptive skill curation as an effective principle for building self-evolving agents. Anonymized code is available at https://anonymous.4open.science/r/EASE-skills.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
AI Agents, Large Language Models, Agent Skills

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_sxWbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to curate evolving skills so that when the executor model is changed, the skill curation policy is still effective. The paper finds that after training a skill curator for a certain executor, transferring such a curator to a new executor can cause a performance drop. Thus, the paper maintains an online behavioral profile of the executor and trains a shared curator that adapts its curation decisions based on this profile. Based on the proposed method, there are improvements on three benchmarks and various executor backbones.

Critical strengths
  1. The paper validates the dependency of the skill manager on its corresponding executor. Experiments show that the manager and executor are often paired best when they are trained together.
  2. The online behavioral profile is easy to construct and well supported by targeted ablations.
  3. The evaluation goes beyond task performance and retrieval rates. There is extensive analysis of the deployment-time token usage, environment steps, and paired edit-rollback diagnostics.
Critical weaknesses
  1. The motivation for a separately trained skill curator is not well supported. A frozen executor could also curate its own skills. The paper does not clearly explain why a separate curator is necessary, given the cost of additional training and maintenance, particularly for strong executors. Considering all that extra trouble brought by an external curator, the motivation is weak.
  2. Comparisons with plain-text skill optimization methods are missing. For example, SkillOpt and GEPA optimize skills without training a separate curator yet show high performance. Such comparisons are necessary to support the method.
Questions for authors
  1. What contributes to the token savings? Since token usage is averaged over all tasks, including the failed ones, the saved tokens may partly come from fewer unsuccessful steps rather than more compact skill context. Could the authors report both the token and step counts on both EASE and baselines so that we could obtain a clearer understanding?
  2. Is the deployment advantage worth the cost of training models? The reported token counts exclude curator training, quality judging, and skill-free reference generation. Could the authors report these costs and the trade-off between using a curator and using plain-text skill optimizers?
5 comments

We deeply appreciate the reviewer sxWb’s recognition of our preliminary analysis on the “dependency of the skill curator on its corresponding executor”, simple yet effective “behavioral profile which is well supported by targeted ablations” and our “extensive analysis of the deployment-time token usage, steps and edit-rollout back diagnosis”. These key points summarize some highlights of our research on learning a transferable skill curator for evolving agents in real environments. To further strengthen our manuscript with the help of reviewer’s insightful comments, we now carefully address and answer each one of them as follows:

W1: The motivation of this work should be further strengthened by comparing with frozen training-free executors that can self-edit their own skills without the effort of additional parameter fine-tuning.

We appreciate the reviewer’s suggestions on strengthening and clarifying our motivations. As the reviewer mentioned, the training-free executor with the skill-using capability alone is an intuitive and nonnegligible baseline. Here are also some of our observations that build up our motivations: 1) Although training-free baselines are more seemingly accessible, they do require either a carefully designed prompt for every domain or an iterative loop that enables automated text optimization during runtime, which introduces sensitivity to prompt or unstability for the stochastic process; 2) These baselines are also largely depend on the strong LLM’s native capability to self-reflect and self-manage its existing skills, which may not necessarily be budget-friendly when using top-tier models due to long-context inference and high API/inference-cost. Given these reasons, learning a separate and generalizable skill curator that works for even lightweight but much more affordable agents becomes natural to us.

EASE framework goes one step forward and targets a harder optimization problem: how to curate executor-adaptive skills that facilitate any agent’s performance. This path offers a plug-and-play solution for any agentic system that supports agent skills. (To our knowledge, many popular and well-developed agent harness systems like Codex, Hermes, Pi, DeepSeek Harness, etc., do support skills natively.) Additionally, since we focus on executor-adaptive skill curation policy learning for any, from weak to strong, executors, even lightweight open-source models (e.g., Qwen3-8B) get benefits from our solution. Lastly, we do believe that by offloading skill curation from the main executor itself, it provides more flexibility for researchers to study stronger and more transferable skill curators.

Empirically, in Table 1, we also compare with several existing approaches (AWM, ReasoningBank, MemP, and Memento-Skills) that follow a similar training-free single-executor configuration as the reviewer suggested. According to the results, these methods are outperformed by training-based methods, including SkillOS and ours, under the same configurations.

Given all these reasons, we believe learning an executor-behavior adaptive skill curation policy is nontrivial and worthwhile. We have updated these discussions in our manuscript to show a more holistic motivation for our research.

W2: To support the method, comparisons with plain-text skill optimization methods are necessary. (e.g., SkillOpt and GEPA optimize skills instead of training a separate curator)

We thank the reviewer for suggesting more advanced training-free prompt optimization methods like SkillOpt[1] and GEPA [2]. GEPA [2] employs natural-language reflection and evolutionary prompt optimization, providing an effective alternative to reinforcement learning for improving LLM systems without parameter updates. SkillOpt[1] applies bounded add/delete/replace edits gated by held-out validation to a single skill document, offering a solution for utilizing frozen LLM agents without parameter training. As the reviewer suggested, we have included these two new context optimization training-free baselines for comparison.

According to our results, GEPA and SkillOpt’s performances are sensitive to different executors: when working with the weak and lightweight agent Qwen3-8B, they are outperformed by the training-based methods (e.g., SkillOS and EASE); However, for a more capable model like Kimi K2.6, they can match the performance of the training-based approach. Our take on this is that when only lightweight executors are accessible, these plain-text skill optimization methods may not be the optimal choices.

Beyond performance, we also observed that these prompt optimization methods are very sensitive to the seed prompt and therefore require lots of heuristic prompt-tuning in the first place. Additionally, the stochastic hill-climbing nature of these text optimization approaches makes them extremely difficult to optimize and analyze retrospectively.

Nonetheless, we do agree that taking parameter-free optimization methods into the discussion will greatly increase the impact of our paper. We greatly appreciate the reviewer for these insightful suggestions and have updated our manuscript accordingly.

MethodID (Qwen3-8B)OOD (Kimi K2.6)
No Memory39.672.1
MemP42.776.3
Memento44.974.9
GEPA43.775.6
SkillOpt44.576.9
SkillOS50.075.3
EASE5581.1

[1] Y. Yang, Z. Gong, W. Huang, Q. Yang, Z. Zhou, Z. Huang, Y. Li, X. Gao, Q. Dai, B. Liu, K. Qiu, Y. Yang, D. Chen, X. Yang, and C. Luo, "SkillOpt: Executive Strategy for Self-Evolving Agent Skills," arXiv preprint arXiv:2605.23904, May 2026.

[2] Agrawal et al. GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning. ICLR, 2026.

Q1: To better explain why the method is token-efficient, the author should consider reporting both the token and step counts on both EASE and baselines to build a clearer picture. Is it fewer unsuccessful steps or a compact skill context that helps the most?

We thank the reviewer for suggesting us to showcase a more detailed analysis of our token usage and efficiency. As we visualized in Figure 3(b) and reported in Table 7, 8, 9, we have shown and compared EASE’s component-wise token consumption against a strong SkillOS baseline. Specifically, EASE’s inference token advantage comes from two parts: i) shorter skill-context tokens: our learned skill curation policy is explicitly optimized to generate concise rather than verbose skills that help executors. ii) fewer agentic steps: EASE-generated skills facilitate executor agents to take a shorter path towards the answer on average. For more clarity, we recollect this evidence from the manuscript and summarize it as follows:

BenchmarkMetricSkillOS-baseSkillOSEASE (ours)
ALFWorldSkill-context tokens3.4k2.8k1.9k
Steps28.52522
Exec. tokens / step315.8300313.6
ScienceWorldSkill-context tokens7.8k6.2k4.2k
Steps32.528.724.4
Exec. tokens / step830.8801.4778.7
WebShopSkill-context tokens8k6.5k4.5k
Steps18.51613.5
Exec. tokens / step1513.51562.51592.6

For long-horizon agentic tasks, it is the agent’s interaction steps that dominate the total token consumption, and the skill context is relatively negligible. Ideally, higher-quality agent skills should not only improve the agent’s success rates, but also reduce redundant steps, repeated tool interactions, and unproductive exploration by allowing agents to reuse previously acquired procedural knowledge [3, 4]. In this sense, our EASE has successfully fulfilled this expectation and therefore contributes to more efficient token inference when compared with baselines.

We hope these empirical results and discussion could better clarify our token efficiency and build up a clearer understanding for future readers and researchers.

[3] S. Chen et al., “SkillCraft: Can LLM Agents Learn to Use Tools Skillfully?,” arXiv preprint arXiv:2603.00718, 2026, doi: 10.48550/arXiv.2603.00718.

[4] X. Li et al., “SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks,” arXiv preprint arXiv:2602.12670, 2026, doi: 10.48550/arXiv.2602.12670.

Q2: Could the authors report more on training costs, the trade-off between using a curator and using plain-text skill optimizers, and justification for training a skill curator at all?

We appreciate the reviewer for the in-depth question, and we’d love to disclose more for reproducibility and facilitate follow-up research. During the training stage, following the hyperparameters we reported in Tables 2 and 3, EASE calls all executors and the curator about 307,200 times for the entire RL optimization process, resulting in a total token cost of approximately 3x10^9, including quality judging and skill-free reference generation. Please refer to the following table for more details on training cost.

StatisticsValue
Training updates60
Groups per sampling batch B32
Rollouts per group N8
Tasks per stream T10
CuratorQwen3-8B + LoRA (r=16)
Skill-free references m4 per (task, executor)
Executor episodes153,600
Curator calls153,600
Executor + curator calls307,200
Judge calls≤ 153,600
Reference episodes≤ 42636
Executor tokens1.41B – 1.95B
Curator tokens0.51B – 0.61B
Judge tokens0.15B – 0.46B
Reference tokens0.34B – 0.51B
Total tokens≈ 3B (3x10^9)

The training-based skill curator optimization and plain-text online skill optimizers represent two different paradigms. We realized that the training-free plain-text skill optimizer enjoys a great advantage: without any parameter finetuning cost, the strong agentic executor itself can handle problem solving, self-reflection, and skill curation altogether.

However, the advantage is not a free lunch. First, using these strong agent models for self-skill curation can be very costly. Second, the self-reflective behavior largely depends on the carefully designed seed prompt that only heuristically guides the models to do so. Beyond the model’s autonomy, how to monitor, analyze, and improve the self-driven hill-climbing process is also an open research problem. Third, it is less effective with lightweight models (as we have seen in Table 1 and in our answers in W2 above)

In contrast, if we can offload the skill-learning job from the main executor itself to another fine-tuned skill curator, the amortized inference cost can actually be lower when compared with pure plain-text skill optimization for long-running agentic systems. More importantly, with a learned transferable skill curator, it becomes feasible to use more affordable executors rather than only top-tier LLMs for easier and more repetitive agentic tasks, which can further reduce the inference cost.

Therefore, although plain-text skill optimizers show their unique training-free advantages, our training-based skill curation learning provides a transferable solution for long-running agentic systems with different executors. We hope the full cost breakdown, comparison, and justification can help clarify our advantages and strengthen our work. We have added these reviewer-inspired discussions to our revised manuscript for future readers.

In summary, we greatly appreciate the reviewer’s constructive comments and insightful suggestions. We hope the updated version can better explain our work’s motivation, contribution, and academic value for the community.

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

This paper identifies that efficient skill management is executor-dependent. Based on this observation, it designs an adaptive skill management method built upon executor behavior profiling. Across experiments covering multiple executors and agent benchmarks, the proposed approach consistently outperforms the strongest baselines while enabling more efficient skill evolution.

Critical strengths

This paper identifies critical limitations in existing literature and proposes an effective solution. It presents a novel idea backed by solid experiments. The idea of profiling models based on their online behaviors can not only be applied to targeted skill refinement, but also inspire other model-dependent frameworks such as harness RSI. Therefore, I believe this work has clear contributions and merits, and meets the acceptance criteria of the conference.

Critical weaknesses

None

Questions for authors
  1. Figure 1 does not clearly distinguish between the trained skill curators and executors.
  2. Why does a general curator yield better performance than dedicated curators trained for specific executors?
2 comments

We greatly thank the reviewer bRrz for acknowledging our “novel idea of profiling models based on their online behaviors,” and we strongly agree with the reviewer’s insight that this idea “can also inspire other model-dependent frameworks such as harness recursive self-improvements”. We’d like to continue working in this direction for our future work. Also, to further strengthen our manuscript, we will answer and address each one of the comments that the reviewer kindly provided as follows:

Q1: Can authors try to distinguish the trained skill curators and executors in Figure 1?

We thank the reviewer for pointing this out. We acknowledge that because the same model names appear on both axes, it is not very intuitive to read Figure 1. We would like to first clarify the meaning of Figure 1 and how we are going to improve its formulation for better readability.

In Figure 1’s settings, each curator is optimized to best serve a single executor. Each row shares a single optimized curator, and each column represents a single executor. Therefore, all configurations on the diagonal line are in-domain evaluation (trained to serve X, test on X), while others are out-of-domain evaluation (trained to serve X, test on Y).

To make Figure 1 and its semantic meaning more accessible for readers, we use more explicit labels on the axes as well as in the caption to make it easier to understand. We deeply appreciate the reviewer for this constructive suggestion.

Q2: Why does a general curator perform even better than previously dedicated curators optimized for a single executor?

We appreciate the reviewer for this insightful question. Intuitively, a dedicated curator learns a fixed policy to serve a single executor, whereas EASE learns from diverse executors how to adaptively respond to different behaviors. It therefore benefits from richer experience while tailoring skills to each observed behavior, ultimately making it more adaptive and competent than a curator dedicated to it. Additionally, equipped with temporal-attribution gates and behavior conditioning, our EASE’s training approach is superior to those dedicated curators. To further analyze the contribution of our multi-executor training recipe, we provide an additional set of budget-controlled experiments with varying numbers of executors during training to examine the learned policy for in-domain (ID) and out-of-domain (OOD) performance:

MethodID (Qwen3-8B)OOD (Kimi K2.6)
SkillOS5068.9
EASE-single53.271.3
EASE-double54.675.8
EASE (full)5581.1

Compared with vanilla SkillOS, EASE-single already outperforms it by 3.2 pp for in-domain executor evaluation, confirming that behavioral conditioning and temporal-attribution mechanisms provide benefits. By increasing the number of executors (adding Qwen3-32B for -double and GPT-OSS-120B for full) during the training stage, we observe performance increases in both in-domain and out-of-domain scenarios. Especially, for the OOD case, EASE achieves a 17.7% increase compared with SkillOS.

Given these intuitive reasons and empirical results, we believe that EASE presents a feasible way towards stronger (in-domain) and transferable (out-of-domain) skill curators. We have included this discussion in our updated manuscript for clearer analysis and to strengthen the empirical evidence of our work.

In summary, we thank the reviewer for all the constructive questions and suggestions. We hope the revised manuscript can be more beneficial to the open academic community and better explain our thoughts and study in this work.

Thanks for your explanations.

Reviewer_7mzPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces EASE, a framework that learns a shared skill curator across multiple frozen executors, conditioning skill addition, modification, and deletion on an online behavioral profile.

Critical strengths
  • This paper addresses the important problem of building a shared skill curator that can serve different executors without retraining for each one.

  • The evaluation covers multiple agent benchmarks and both seen and unseen executors, providing evidence for the effectiveness and transferability of the proposed approach.

Critical weaknesses
  • The benefit of multi-executor training is not clearly isolated. EASE is compared with executor-specific SkillOS curators, but the methods also differ in behavioral conditioning and attribution gates. It remains unclear whether training EASE with a single executor would already provide comparable transfer.

  • The value of training a separate curator is insufficiently established. Although training-free memory baselines are included, there is no direct comparison with letting the executor itself manage skills using the same behavioral profile and editing tools, potentially with context or prompt optimization [1, 2].

  • The hand-designed behavioral profile needs more analysis. The window-size and whole-profile ablations are useful, but do not identify which features drive the gains or establish robustness to missing and noisy behavioral measurements.

  • The attribution gates are only indirectly validated. Retrieval and behavioral deviation do not establish that an edit improves performance. The rollback analysis reports average benefits, but does not show whether the gates reliably select more useful edits.

[1] Zhang et al. Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models. ICLR, 2026.

[2] Agrawal et al. GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning. ICLR, 2026.

Questions for authors
  • How does transfer to unseen executors change when EASE is trained with one, two, or three executors under the same total rollout budget? Can single-executor training already achieve comparable generalization?

  • How does EASE compare with letting the executor itself curate skills through prompting, without training a separate curator? Please consider the same behavioral profile, editing tools, and optimized prompts under comparable compute budgets.

  • Which behavioral features are most important? Feature ablations and tests with shuffled, missing, or noisy profiles would help establish whether the curator uses reliable behavioral information.

  • Do edits admitted by the gates have larger rollback benefits than excluded edits? How well does behavioral deviation predict actual edit utility, and have the authors considered using estimated performance effects directly for credit assignment?

  • Could the authors provide concrete before-and-after skill examples and separate rollback results for additions, modifications, and deletions? In particular, do deletions remove harmful skills, or mainly reduce repository size?

6 comments

We greatly thank the reviewer 7mzP for recognizing the important problem that we worked on in this paper to “build a shared skill curator that can serve different executors without retaining”. We also provided multiple empirical results and ablation studies to “provide evidence for the effectiveness and transferability” of our proposed method. To further strengthen our manuscript with the help of the reviewer’s insightful comments, we now carefully address and answer each one of them as follows:

W1 & Q1: To better separate the contribution of multi-executor training from behavioral conditioning and attribution gates, the authors should consider adding comparison experiments with EASE being trained with a single, two, and three executors.

We thank the reviewer for this insightful suggestion. The purpose of multi-executor RL training aims at incentivizing the policy to learn cross-executor transferable skill curation rather than specifically optimizing for a single executor (as the motivation suggested in Figure 1 and Section 1). Additionally, EASE incorporates behavioral conditioning for explicit executor modeling and attribution gates for more efficient credit assignment during RL. To clarify each component’s contribution, current ablation studies already isolated behavioral conditioning (Figure 4(a)) and temporal attribution (Figure 4(b)) in Section 4.5, showing their individual benefit.

To further analyze the contribution of multi-executor training, we carry out an additional set of budget-controlled experiments with varying numbers of executors during training to examine the learned policy for in-domain (ID) and out-of-domain (OOD) performance:

MethodID (Qwen3-8B)OOD (Kimi K2.6)
SkillOS5068.9
EASE-single (SkillOS+)53.271.3
EASE-double54.675.8
EASE (full)5581.1

The EASE-single is an augmentation of the Qwen3-8B-specific SkillOS, denoted as SkillOS+, with EASE’s temporal-attribution gates and behavioral conditioning techniques. Compared with vanilla SkillOS, EASE-single already outperforms it by 3.2 pp for in-domain executor evaluation. This result confirms that EASE's behavioral conditioning and temporal-attribution mechanisms provide benefits even under single-executor training. By increasing the number of executors (adding Qwen3-32B for -double and GPT-OSS-120B for full) during the train stage, we observe performance increases in both in-domain and out-of-domain scenarios. Especially, for the OOD case, EASE (full) achieves a 17.7% increase w.r.t the SkillOS baseline and 13.7% w.r.t the SkillOS+. With all the evidence, we believe that the multi-executor training is necessary and presents a feasible way towards transferable skill curation for unseen executors at test time.

We agree that these additional experiments can help build up a clearer picture of the contribution of each component we proposed, and we have included this in our updated manuscript for clearer analysis and to strengthen the empirical evidence of our work.

W2 & Q2: The motivation for training a separate curator can be strengthened. How does EASE compare with letting the executor itself curate skills through prompting, without training a separate curator, and potentially with context or prompt optimization [1, 2]?

We thank the reviewer for mentioning these two related works for comparison. We’d like to first clarify our understanding of these methods, explain why EASE is different from and potentially superior to them, and present the empirical results for the comparison.

GEPA [2] employs natural-language reflection and evolutionary prompt optimization, providing an effective alternative to reinforcement learning for improving LLM systems without parameter updates. ACE [1] develops evolving contexts through generation, reflection, and incremental curation, demonstrating the effectiveness of structured, training-free context evolution. These context optimization methods, although training-free, do require a carefully designed agentic system loop that enables text-based optimization to happen reliably and largely depend on the strong executor’s capability to self-reflect and self-improve, which may not necessarily be budget-friendly due to long-context inference and API costs.

Our proposed EASE framework targets a different optimization problem: how to curate executor-adaptive skills that facilitate any agent’s performance. This path offers a plug-and-play solution for any agentic system that supports agent skills. Additionally, since we focus on executor-adaptive skill curation policy learning for any executor, from weak to strong, even lightweight open-source models (e.g., Qwen3-8B) benefit from our solution (Table 1). Also, we believe plugin-like skills provide more controllability and flexibility for humans to understand, monitor, and manipulate.

As the reviewer recommended, we provide a training-free EASE using the same behavioral profile and editing tools as we proposed in the paper for comparison. We also implemented GEPA[2] and ACE[1] as representatives of context optimization methods.

MethodID (Qwen3-8B)OOD (Kimi K2.6)
No Memory39.672.1
MemP42.776.3
Memento44.974.9
EASE-free44.570.6
GEPA43.775.6
ACE45.676.8
SkillOS50.075.3
EASE5581.1

According to the results, the training-free EASE variant (EASE-free) is significantly worse compared with our RL-finetuned version, especially for out-of-domain scenarios, suggesting the necessity of RL optimization in the EASE framework to learn better executor-transferable skill curation; GEPA and ACE’s performance largely depends on the executor’s capability for reflection and prompt optimization retrospectively. Empirically, working with Kimi K2.6, GEPA and ACE outperform SkillOS but are still below EASE’s record. During experiments, we noticed that both GEPA and ACE are very sensitive to the seed prompt, and we tried lots of heuristic prompt-tuning in the first place.

In summary, we hope these in-depth discussions and comparisons against the training-free executor with context / prompt optimization, as well as empirical evidence, can strengthen our motivation for this work. We greatly appreciate the reviewer’s suggestion on this, and we have updated our manuscript accordingly.

[1] Zhang et al. Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models. ICLR, 2026.

[2] Agrawal et al. GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning. ICLR, 2026.

W3 & Q3: The design of the behavioral profile needs more discussion. It can be helpful to provide more ablation studies to understand the contribution of each feature and their robustness more clearly.

We thank the reviewer for paying attention to one of our key components, behavioral profiles, in this paper. The design of the behavioral profile starts with a simple intuition: adaptive skill curation should learn from the executor’s behaviors. To capture these behavioral patterns, we construct a behavioral profile using six complementary feature families (Section 3.2 and Appendix B, Table 4), without needing additional model inference or access to future tasks. To our knowledge, our idea of a behavioral profile is a first attempt for executor-conditioning and transferability, which has been shown to be effective and highly interpretable and traceable (formulated with natural language).

Nonetheless, we have not taken the given design of the behavioral profile for granted and have provided further analysis and justifications. As the reviewer has recognized, we have shown an ablation study in Figure 4(a) and Appendix G to support the effectiveness of the profile design. Table 6 and Appendix H also provide analysis for the profile’s window size. Additionally, we conduct an additional set of experiments where some features are randomly masked out intentionally for ablation.

Profile Features100% Mask50% Mask0% Mask (EASE)
Action Validity66.968.569.7
Failure Recovery67.168.969.7
Skill Utilization68.469.269.7
State Tracking67.569.169.7
Efficiency69.669.769.7
Error (Patterns) Types69.269.669.7

According to the results, with certain features being completely masked out, we observe performance decreases compared with our full EASE framework. Specifically, the "Action Validity" feature contributes the greatest degradation (-2.8), while "Efficiency" is the mildest (-0.1). Additionally, when we randomly masked out 50% of a given feature, the performance recovered to different degrees (for more significant features like “Action Validity” and “Failure Recovery”, only 50% observations bring 60%+ performance recovery on average).

In conclusion, we believe these additional discussions and ablation studies can foster a better understanding of each component of our profile and also examine the robustness quantifiably even if certain features are missing. We’ve updated our manuscript to reflect these new findings as the reviewer suggested.

W4 & Q4: The attribution gates can be further validated: do edits admitted by the gates have larger rollback benefits than excluded edits? How do these gates work to shape a better credit assignment strategy?

We thank the reviewer for this insightful comment. It is true that these two gates are not intended to estimate edit utility or distinguish beneficial edits from harmful ones. Rather, their purpose is to refine GRPO’s coarse and universal credit assignment per rollout by filtering out unrelated actions that lack temporal evidence of influencing subsequent consequences. During RL, the relative utility or value given an action is still estimated by the computation of GRPO’s group-relative advantage.

Our controlled ablations provide further evidence that this proposed attribution mechanism does improve learning (Figure 4(b)). Removing both gates reduces ALFWorld success rates by an average of 5.7 percentage points, while keeping only the retrieval gate still underperforms full EASE by 1.7 points. Moreover, a random mask with the same admission rate performs similarly to ungated training, suggesting that the improvement comes from selecting curation steps based on downstream relevance rather than merely reducing the number of policy-gradient updates.

We have further clarified this distinction and the approximate nature of our attribution mechanism in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.