acceptodds
Under review as a conference paper at ICLR 2027

SkillZip Pro: Execution-Aware Dynamic Compression of Progressively Loaded Skills for Self-Evolving Agents

Abstract

Self-evolving agents accumulate useful procedures, but also repeat the same rules across references and subskills. Compressing these skills as flat prompts misses a crucial distinction: the agent stores a whole directory but loads only the files needed for a task. A smaller directory can therefore increase execution context or leave a required subskill inaccessible. We introduce \method, an evaluation-free compiler that compresses content and its loading structure together. It shares repeated instructions within the branches that use them, removes obligations already supplied by declared context, and defers explicitly guarded sections. Typed removal evidence and an independent audit of the emitted files preserve routing, interfaces, and standalone entries. The output uses ordinary files without changing the agent harness. One-Shot and Continual schedules support both rewritten bundles and source-preserving execution views. Across three task suites, \method saves 19.7% of shipped and 14.3% of mean-run tokens while preserving every declared route and meeting a five-point quality non-inferiority criterion. On a production moderation service, \method shows 38.1% skill token and 10.4% end-to-end runtime token savings, with nearly no performance drop. blueCode available at: https://anonymous.4open.science/r/skillzip-pro-ICLR

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Self-Evolving Agents, Skill Compression

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_AQH2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

SkillZip Pro studies how to compress the progressively loaded skill files to improve efficiency of applying the skills. It finds that shrinking the menu does not necessarily mean a lower cost, and overly deduplication may cause damage to sub-skill utilization. The method optimizes the context, shared location and loading structures at the same time. In the experiments, the method achieves SOTA performance in balancing the skill compression rate and the effectiveness maintained.

Critical strengths
  1. The targeted problem is practical and meaningful. As the skill can be updated often, reducing its size to improve efficiency and clarity is increasingly important and widely applicable.
  2. The proposed method successfully achieves a good balance between the compression ratio and the effectiveness.
  3. The experiments are thorough and cover multiple aspects of possible application scenarios and potential concerns about the method.
Critical weaknesses
  1. The experiments are too dense to understand fully. There are many experimental settings for each figure or table, making the reader hard to follow.
  2. The method is straightforward, making the contribution more centered on the engineering part. The proposed methods are more likely heuristic-based. Since this is a pro compared to SkillZip, providing a certain level of first-principles rules for compressing skills would be a major advance.
Questions for authors

Please refer to the weakness

1 comment

I thank the authors for the detailed rebuttal and follow-up. Organizing the experiments around a few explicit questions makes the evaluation easier to follow, and the stated decision rules, together with the worked trace, clarify why each compression move is taken. I encourage including that worked example in the final version. However, these rules are local cost-benefit conditions derived from the paper's own cost model, and the audits mainly show that the compiler follows them. They therefore do not yet amount to general principles for compressing skills. Moreover, part of the deletion safety rests on a model-based entailment check, which the paper itself treats as auditable evidence rather than a formal guarantee, so the soundness claim in the response should be stated more cautiously. I will keep my current positive score.

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

SkillZip Pro compresses progressively loaded skill directories while accounting for where shared content is placed and whether individual entry points remain usable. It represents resources, guards, and obligations in a graph; applies scoped sharing, conditional loading, and removal of content supplied by declared host contracts; and audits the emitted files with fallback on failure. The evaluation includes constructed growth workloads, evolving skill libraries, cross-executor tests, and two sessions in a production system. The central claim is reduced storage and execution context with limited task-quality loss.

Critical strengths
  1. The formulation captures a practical execution constraint. Shorter stored files can increase loaded context or break directly invoked subskills. Treating placement and entry-point validity as optimization targets is useful. Locked resources, guarded obligations, final-file auditing, and fallback make the approach operational, with preservation claims explicitly conditional on extraction and auditor assumptions.

  2. The evaluation includes evidence of real execution savings. Table 1 reports 19.7% shipped-token and 14.3% mean-run reductions, and the production sessions report a pooled 10.4% token reduction. The production evidence complements structural tests. Appendix H.4 candidly identifies unresolved provenance and statistical limitations, allowing a more careful assessment of the claimed benefit.

Critical weaknesses
  1. Quality preservation depends on an insufficiently justified tolerance. Table 7's paired 95% interval [−0.029, +0.059] supports non-inferiority at a five-point margin, but not the one-point margin in Appendix G.8. Appendix H.4 acknowledges this distinction and the expansion of the Spreadsheet sample after observing ranking sensitivity. The margin's timing and practical justification are therefore consequential. Matched production accuracy is 89% versus 92% and 89% versus 91%, supporting a cost-quality trade-off rather than negligible loss independent of application tolerance. Failure at the stricter margin does not establish significant inferiority.

  2. The practical magnitude of savings is difficult to assess across workload and cost definitions. The main grown track deliberately appends repeated content; the additional libraries retain all rounds as files and have approximately 79–84% redundancy. These are useful mechanism tests but leave dependence on the growth construction unresolved. Appendix H.4 also retains unreconciled main versus ablation figures of 19.7/14.3 and 21.1/15.8, along with different tail and cumulative-cost measures. Native-directory results and a unified accounting of loaded context, billed full-run tokens, and compression/audit overhead are needed to substantiate the practical advantage.

Questions for authors
  1. When and why was the five-point non-inferiority margin chosen, and was the additional Spreadsheet sample an independent held-out set? Please distinguish prespecified and exploratory analyses and justify the practical quality tolerance.
  2. Can you isolate results on the native skill directories before deterministic growth augmentation, using the existing tasks and entry constraints? This would directly test dependence on artificially introduced redundancy without requiring a new benchmark suite.
  3. Could you reconcile the main and ablation cost figures using their source runs and specify tokenizer, workload weighting, included calls, and audit overhead? For production quality, please provide paired outcomes and identify repeated tasks across sessions.
Reviewer_1JaBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Many skills are evolved during agent execution. This paper introduces SkillZip Pro, an evaluation-free compiler that compresses content and its loading structure together. It shares repeated instructions within the branches that use them, removes obligations already supplied by declared context, and defers explicitly guarded sections.

SkillZip Pro combines two mechanisms. Activation-aware cross-file compression finds reuse within the scope that needs it. Cross-file routing preservation checks that the result remains accessible.

SkillZip Pro saves 19.7% of shipped and 14.3% of mean-run tokens while preserving every declared route and meeting a five-point quality non-inferiority criterion.

Critical strengths
  1. The proposed skills compression method is practical and the topic is important. It uses resource graph, root activation cost, and expected execution cost rather than just shorten the prompt.

  2. It support both offline compression of prepared skills and online compression during agent execution.

  3. Experiments are provided with real-world deployment. It saves 19.7% of shipped and 14.3% of mean-run tokens.

Critical weaknesses
  1. The improvement over SkillZIP can be limited. Add a direct comparison with SkillZIP in both algorithm and experiments can be helpful.

  2. The online evaluation can be improved. It needs show how an agent continuously solving unseen new tasks, learning from feedback, and improving its skills, which is unclear in section 5.2.

  3. Overall experiments can be improved. As in figure 1, the accuracy and cost saving may not be sufficient. Production accuracy drop: "Accuracy is 89% in both sessions, against paired raw baselines of 92% and 91%(Table 6).". Also in figure 6, Kimi K2.6 performs worse with compressed skills.

Questions for authors
  1. The code is incomplete. It lacks the SkillOpt evolution pipeline before the compression and BFCL evaluation scripts, making the results difficult to reproduce.

  2. Refine figure 1 and figure 2 to help reviewers understand the entire process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.