PaperlayerView exact revision on arXiv

Source-linked AI summary

Skill$^{\infty}$: Recursive Self-Improvement for Agent Skill Optimization

Wangcheng Tao, Zhanglu Yan, Han Wu, Weng-Fai Wong

arXiv:2610.35434v1cs.AIcs.CL

TL;DR

Existing skill optimizers rely on hand-designed procedures, and meta-skill approaches keep fixed pipelines and relearn for each task. Skill∞ refines a reusable meta-skill through the same agent-led procedure it uses to optimize task skills. On SkillsBench, it reaches a 60.7% pass rate, above human-curated skills and SkillEvolver.

  • Problem

    Existing skill optimizers follow hand-designed procedures, while meta-skill approaches retain fixed pipelines and bind or relearn meta-skills for individual tasks.

    Source: §1 · pp. 1–2

  • Method

    Skill∞ uses an agent-led meta-skill to optimize task skills and applies the same procedure to refine that meta-skill on skill optimization tasks for reuse across task skills.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    60.7% pass rate on 83 SkillsBench tasks, compared with 56.4% for human-curated skills and 54.5% for SkillEvolver.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Skill∞ makes recursive self-improvement direct by encoding the optimizer’s procedure as a reusable skill, loaded through native skill discovery at both optimization levels.

    Source: §5 · p. 8

    Skill∞ applies to domains where task rewards can be computed reliably because both optimization levels rely on automatic verifiers.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Recursive self-improvement (RSI), in which a system also improves the way it improves, is a long-standing aspiration on the path to AGI. Language agents increasingly store procedural knowledge in reusable skills. Optimizing skills offers a direct way to achieve RSI, because the optimization procedure itself can be encoded as a skill. Existing skill optimizers, however, follow a hand-designed procedure. Even methods that evolve a meta-skill confine it to a fixed pipeline and relearn it for each task. We present Skill$^{\infty}$, which makes skill optimization recursively self-improving: a meta agent follows a meta-skill to optimize task skills and optimizes the meta-skill itself using the same procedure. The optimization is agent-led. The meta agent diagnoses execution traces, proposes candidates, requests independent reviews, and allocates the search budget. The meta-skill defines these steps, allowing self-improvement to rewrite them. This meta-skill is also decoupled from task skills, refined on multiple skill optimization tasks and then reused to optimize task skills across diverse domains. As in deployment, skills at both levels are standard packages loaded through native skill discovery. On 83 SkillsBench tasks, Skill$^{\infty}$ reaches a pass rate of 60.7%, compared with 56.4% for human-curated skills and 54.5% for SkillEvolver.

1 Introduction

Skill∞ makes skill optimization recursively self-improving by refining a reusable meta-skill that governs task-skill optimization. On SkillsBench, it outperforms human-curated skills and SkillEvolver.

1 Introduction

  • Skills can encode both the target of improvement and the procedure that improves it, making skill optimization a direct and interpretable route to recursive self-improvement.The shared skill specification also supports loading skills through multiple agent harnesses.Source: §1 · p. 1
  • Earlier optimizers use hand-designed procedures, while methods that evolve a meta-skill retain fixed pipeline roles and bind it to the task skill being optimized.Some methods also evaluate skills in settings that prompt or inject them rather than testing whether agents load them on their own.Source: §1 · p. 1
  • Figure 1 depicts a self-improvement loop that refines the meta-skill and task-skill loops that reuse it.The meta-skill is a standard package loaded through native skill discovery, and its instructions and scripts are within the scope of self-improvement.Source: Figure 1 · p. 2
  • Skill∞ applies the same agent-led procedure to task skills and the meta-skill, which is refined on optimization tasks and reused across diverse task skills.The meta agent diagnoses execution traces, proposes and evaluates candidates within a search budget, requests independent reviews, and selects the next parent.Source: §1 · pp. 1–2
  • 60.7% pass rate on SkillsBench, compared with 56.4% for human-curated skills and 54.5% for SkillEvolver.The evaluation covers 83 agentic tasks across eight domains.Source: §1 · p. 2
  • Without meta-skill improvement, the pass rate falls to 55.9%.Skill∞ still outperforms human-curated skills when the backbone model or harness is replaced.Source: §1 · p. 2

2 Related Work

Related work spans recursive self-improvement, agent skills and skill optimization, and prompt and context optimization. Skill∞ differs by encoding an editable optimizer procedure as a reusable skill.

Recursive Self-Improvement

  • Earlier RSI systems modify programs, agent logic, or policies; Skill∞ instead encodes the optimizer’s procedure as a skill, which the paper describes as direct, interpretable, and harness-agnostic.The passage contrasts Skill∞ with systems that implement the optimizer as a program or trained policy.Source: §2 · p. 2

Agent Skills and Skill Optimization

  • Skill optimizers learn from execution feedback, but prior meta-skill approaches retain a hand-designed procedure; MetaSkill-Evolve fixes pipeline roles and binds each meta-skill to its task skill.Skill∞ refines its agent-led procedure across several skill optimization tasks, then reuses the decoupled meta-skill on diverse task skills.Source: §2 · p. 3

Prompt and Context Optimization

  • Prompt and context optimization edits text inserted into model context, whereas skills can bundle scripts and reference files, load on demand, and package the optimizer itself.The paper distinguishes skill packages from text optimized by the surrounding system for insertion into context.Source: §2 · p. 3

3 Method

Skill∞ uses an editable meta-skill to optimize task skills and itself through the same procedure. It separates the reusable optimization procedure from task-specific skills and loads both as standard skills.

3 Method

  • Skill∞ formalizes native skill discovery and makes the meta-skill an optimization target alongside task skills.The same agent-led procedure optimizes both levels.Source: §3 · p. 3

3.1 Preliminaries

  • A skill package bundles instructions and optional resources, which agents load on demand through native skill discovery.Agents initially see a skill’s name and description, then load its full instructions and resources when useful.Source: §3.1 · p. 3

3.2 Recursively Self-Improving Skill Optimization

  • The meta-skill is evaluated on skill optimization tasks by the gain its guided task-skill optimization achieves over the seed skill.The meta score averages gains across optimization tasks, so the same procedure can be evaluated and improved.Source: §3.2 · p. 4

3.3 Agent-Led Optimization

  • The meta agent diagnoses rollouts, proposes and evaluates candidates, and uses independent review and an adaptive search budget to choose how optimization proceeds.It can interpret scores alongside diagnoses and reviews, retain tied candidates that address failures, backtrack when revisions stall, and stop when little remains to improve. The meta-skill encodes these steps and can itself be revised; scoring remains outside self-improvement.Source: §3.3 · pp. 4–5

3.4 Decoupling Meta-Skill from Task Skills

  • The meta-skill is refined on optimization tasks from different domains, then reused to optimize task skills from their seed versions.Averaging gains across domains favors revisions that help broadly, and reusing the refined package amortizes its self-improvement cost.Source: §3.4 · pp. 5–6

3.5 Native Skill Discovery at Both Levels

  • Both skill levels are installed and loaded through native skill discovery as in deployment, without prompt injection or instructions to load them.Activation depends on each skill’s description, which is within the optimization scope; the meta-skill is portable across harnesses supporting native skill discovery.Source: §3.5 · p. 6

4 Experiments

The experiments evaluate Skill∞ on agentic tasks and kernel optimization, then examine its self-improvement, portability, task coverage, search-budget scaling, and cost.

4.1 Experimental Setup

  • Experiments cover 83 SkillsBench tasks across eight domains and three KernelBench problems, using avg@5 pass rate and correctness-weighted speedup.The main setup uses Codex CLI with GPT-6 Luna; the KernelBench skill is shared across a deep narrow MLP, ShuffleNet, and GRU.Source: §4.1 · p. 6

4.2 Main Results

  • 60.7% overall pass rate for Skill∞ exceeds human-curated skills (56.4%) and SkillEvolver (54.5%) on SkillsBench.It leads in software engineering, media, and mathematics.Source: §4.2 · p. 6, Table 1 · p. 7
  • 1.62 mean speedup on KernelBench for Skill∞ compares with 1.18 for SkillEvolver and 1.02 without a skill.On the deep narrow MLP and ShuffleNet, Skill∞ reaches 1.87 and 1.97, respectively.Source: Figure 3 · p. 7, §4.2 · p. 7

4.3 Ablation of Self-Improvement

  • 55.9% pass rate with the seed meta-skill rises to 60.7% with the refined meta-skill.The largest reported gains are 24.0, 22.2, and 11.4 points in mathematics, media, and software engineering; the authors attribute them to stricter rollout evidence for confirming fixes.Source: §4.3 · p. 7

4.4 Generality across Harnesses and Models

  • 66.5% pass rate with GPT-5.6 Luna and 56.9% with OpenCode exceed human-curated skills at 61.2% and 53.3%, respectively.The evaluation changes either the backbone model or the harness.Source: Figure 4 · p. 7, §4.4 · pp. 7–8

4.5 Where Skill∞Helps

  • 30.5% in category D and 29.2% in C1 compare with SkillEvolver’s 22.1% and 15.4%, respectively.D tasks are unsolved by the agent and human-curated skills; C1 tasks are solved only by human-curated skills at a pass rate of at least 40%.Source: Table 2 · p. 8, §4.5 · p. 8

4.6 Budget Scaling and Amortized Cost

  • 47.5% pass rate with one candidate rises to 82.5% with eight, while the agent without a skill reaches 12.5%.These budget results come from a subset where the default budget substantially improves the seed task skill; the meta agent allocates its budget without retuning the procedure.Source: Figure 5 · p. 8, §4.6 · p. 8
  • 8.7M input and 66k output tokens per task-skill optimization with the refined meta-skill are close to SkillEvolver’s 8.1M and 76k.Self-improvement costs about as many tokens as eleven task-skill optimizations; spread over 83 tasks, it adds less than 15% per optimization.Source: §4.6 · p. 8

5 Conclusion

Skill∞ makes skill optimization recursively self-improving by using a meta-skill both to optimize task skills and to improve the optimization procedure. The paper reports stronger results than human-curated skills and SkillEvolver on SkillsBench, and improved kernel speedups on KernelBench.

5 Conclusion

  • Skill∞ addresses skill optimizers’ hand-designed procedures and fixed pipelines by applying the same agent-led procedure to task skills and the meta-skill.The meta-skill is refined on skill-optimization tasks, decoupled from task skills, and reused across domains.Source: §5 · p. 8
  • The framework uses standard skill packages loaded through native skill discovery at both the task-skill and meta-skill levels.The paper characterizes this encoding of the optimizer’s procedure as direct, interpretable, and harness-agnostic.Source: §5 · p. 8
  • On SkillsBench, Skill∞ outperforms human-curated skills and SkillEvolver; on KernelBench, it improves kernel speedups.The conclusion presents recursive self-improvement as a view that may invite further work on agents improving how they improve.Source: §5 · p. 8

Limitations

Skill∞ has three author-identified boundaries: gains are limited by skill-package expressiveness, rewards must be reliably computable, and agent adherence to long instructions is not enforced by tools.

Limitations

  • Gains are bounded by skill-package expressiveness because the backbone model stays fixed.The method improves skills recursively without changing the backbone model.Source: Limitations · p. 9
  • Skill∞ applies only where task rewards can be computed reliably, since both optimization levels rely on automatic verifiers.The meta score is computed from task rewards.Source: Limitations · p. 9
  • Agent-led flexibility leaves adherence to long instructions dependent on the backbone model, meta-skill, and independent review rather than tool enforcement.The meta agent follows its procedure instead of executing it as code.Source: Limitations · p. 9

Ethical Considerations

The paper recommends reviewing optimized skills before deployment because they can run scripts with tool access and may activate without user approval; it also describes the benchmark’s task utility categories.

Ethical Considerations

  • Optimized skills can bundle executable scripts and may activate without user approval, so the authors recommend reviewing their behavior and activation conditions before deployment.The paper notes that some SkillsBench cybersecurity tasks are used only in the form provided, and reports Apache 2.0 and MIT licenses for the named artifacts.Source: Ethical Considerations · p. 9
  • The paper uses these artifacts for research evaluation consistent with their intended use and identifies their respective licenses.SkillsBench is Apache 2.0; KernelBench and the SkillEvolver dataset are MIT-licensed.Source: Ethical Considerations · p. 9

A Skill-Utility Categories

  • Category A tasks have No skill pass rates of at least 80%, while category D tasks are passed by neither No skill nor Human-curated.The intermediate categories distinguish No skill pass rates and whether Human-curated performs better, equally, or worse; C1 and C2 distinguish Human-curated pass rates at or above 40% versus between 0% and 40%.Source: A Skill-Utility Categories · p. 11

B Implementation Details

Implementation gives the meta agent a bounded candidate-search budget and a repeated workspace request to optimize the task skill and record its decision.

B Implementation Details

  • The inner task-skill optimization budget is two candidates per task, and self-improvement may create at most four meta-skill candidates.Source: B Implementation Details · p. 11
  • At each iteration, the meta agent is asked to optimize the task skill in the workspace and complete the iteration.The request names the workspace and iteration id, asks the agent to write its decision, and then return.Source: B Implementation Details · pp. 11–12
  • The implementation presents this instruction as the request to the meta agent.Source: B Implementation Details · p. 12

The Meta-Skill Rewrites Its Selection Rule

The meta agent revised candidate selection to require evidence of progress before accepting tied or worse candidates, then applied that rule to a later tied candidate that fixed the targeted task failure.

The Meta-Skill Rewrites Its Selection Rule

  • The refined selection rule required checking tied or worse candidates’ rollouts for the targeted failure, allowing a tie only when the rollout showed progress.The first revision followed an optimization that accepted two candidates despite the same link-geometry failure, including one that only tied its parent.Source: The Meta-Skill Rewrites Its Selection… · p. 12
  • The later tied candidate was selected after its rollout addressed the targeted failure and brought the 3D scene task to a perfect score.Source: The Meta-Skill Rewrites Its Selection… · p. 12
  • Across 83 task-skill optimizations, the refined meta-skill selected 54% of tied candidates, versus 83% for the seed meta-skill.Self-improvement stopped after two further candidates and retained the refined meta-skill.Source: The Meta-Skill Rewrites Its Selection… · p. 12

A Task Skill Gains a Script

The seed skill passed individual mesh exports but failed the merged link geometry check. The meta agent diagnosed the mismatch, added an exporter script and explicit ownership and transform rules, and the optimized skill passed all five test trials.

A Task Skill Gains a Script

  • The meta agent traced the failed link OBJ to incorrect group ownership and coordinate transforms, then added a reusable exporter and explicit guidance for both.The rollout selected direct children and kept group-local transforms, while the ground-truth generator used nearest named ancestors and each mesh’s world matrix.Source: A Task Skill Gains a Script · pp. 12–13
  • The optimized skill passed all five test trials after the reviewer confirmed the exporter script existed and matched its documented interface.The seed skill had failed all test trials.Source: A Task Skill Gains a Script · p. 12
  • The exporter assigns each mesh to its nearest named ancestor group, applies its full world matrix, and creates individual mesh files and combined group link files.The skill also specifies preserving required names and directory structure and detecting duplicate output names.Source: A Task Skill Gains a Script · p. 12
  • The optimized skill directs agents to inspect the exporter output and validate filenames, OBJ contents, and link geometry against supplied tests or the verifier.It also instructs agents to address verifier failures before reporting completion.Source: A Task Skill Gains a Script · pp. 12–13

D Results with Another Model or Harness

The section reports per-domain results with another backbone model or harness and lists optimization token costs, excluding methods that do not optimize task skills in the same way.

D Results with Another Model or Harness

E Optimization Cost

  • Table 4 lists tokens consumed during task-skill optimization and self-improvement, but omits Self-generated and Human-curated skills.Self-generated skills are written within test trials, while Human-curated skills require no optimization.Source: E Optimization Cost · p. 13

F Seed Task Skills

The seed skills establish the task-agnostic guidance and procedures for iteratively optimizing task skills or the meta-skill itself. The meta-skill uses rollout and evaluation commands to assess skill nodes and candidates, keep submitted nodes unchanged, and guide whether the search continues or stops.

F Seed Task Skills

  • The seed task skills share a template with task descriptions but only generic inspection, execution, and validation guidance; the KernelBench seed adds generic profiling, correctness, and timing advice.Source: F Seed Task Skills · p. 13

G Seed Meta-Skill

  • The seed meta-skill package includes SKILL.md, three reference files, and a Python script that defines the optimization iteration, diagnostic procedures, scores, interfaces, and candidate initialization.Source: G Seed Meta-Skill · p. 13
  • Each iteration inspects the archive, diagnoses parent rollouts, forms falsifiable hypotheses, and creates one or two budgeted candidates with evidence-based proposals.Source: G Seed Meta-Skill · pp. 13–14
  • Candidates are evaluated in parallel, independently reviewed by fresh reviewer subagents, and selected against the parent using validation scores, reviews, history, and remaining budget.A tied candidate may be selected when it addresses a diagnosed failure or requirement.Source: G Seed Meta-Skill · p. 14
  • For task skills, diagnosis compares agent actions and outputs with task requirements and verifier checks, then targets missing or misleading guidance with reusable scripts or reference files.Source: G Seed Meta-Skill · pp. 14–15
  • For meta-skills, diagnosis compares task-skill score changes and traces of agent and reviewer behavior to identify improvements to diagnosis, candidate generation, review, and selection guidance.Source: G Seed Meta-Skill · p. 15
  • Continue by making the selected node the next iteration's parent and specifying a new optimization direction; stop when the run ends, with a reason grounded in evidence.A later increase in the candidate budget can resume from the final skill; submitted nodes remain unchanged, and further edits belong in a new candidate.Source: G Seed Meta-Skill · pp. 16–17
  • Rollout runs a skill node on the training set and records its score, activation, usage, elapsed time, and supporting trajectory materials.The response includes the optimization target, node ID, and rollout path.Source: G Seed Meta-Skill · p. 17
  • Candidate evaluation runs on the validation set and returns the node ID, validation score, and skill-activation status.The result is saved in the candidate directory's evaluation.json file.Source: G Seed Meta-Skill · p. 17
  • Skill activation is true when confirmed across evaluated executions, false if any execution lacks activation, and null when unknown.Source: G Seed Meta-Skill · p. 17

Related papers

Loading 2610.35434v1…