Source-linked AI summary
OpenSkill: Open-World Self-Evolution for LLM Agents
Zhiling Yan, Dingjie Song, Hanrong Zhang, Wei Liang, Yuxuan Zhang, Yutong Dai, Lifang He, Philip S. Yu, Ran Xu, Xiang Li, Lichao Sun
TL;DR
Open-world deployments may provide only a task prompt, leaving agents without skills or verification signals for self-evolution. OpenSkill builds both from open-world resources and achieves the best automated pass rate across three benchmarks and two target agents without target-task supervision during learning.
Problem
Existing self-evolving agents assume curated skills, successful traces, or task feedback, whereas open-world deployments may provide only a seed task prompt.
Method
OpenSkill acquires grounded knowledge and verification anchors, synthesizes transferable skills, and refines them against self-built virtual tasks without target-task supervision.
Results
Across three benchmarks and two target agents, OpenSkill achieves the best automated pass rate; it records 43.6% on Opus 4.6.
Takeaways & Limitations
Its skills transfer across models, and its self-built verifier aligns with ground-truth outcomes it never observes.
Takeaways & Limitations
Virtual tasks may not capture the full difficulty of target tasks, potentially overestimating skill quality.
Abstract
from arXiv · showhide
Self-evolving agents requires adaptation after deployment, but existing approaches assume a usable learning loop, such as curated skills, successful trajectories, or verifier signals. Real open-world deployments may provide none of these, offering only a task prompt. In this work, we study open-world self-evolution, where an agent must build both its skills and its own verification signals from scratch, using open-world resources but no target-task supervision. We propose OpenSkill, a framework that bootstraps this loop: it acquires grounded knowledge and verification anchors from documentation, repositories, and the web, synthesizes them into transferable skills, and refines those skills against self-built virtual tasks grounded in the anchors rather than in target answers. The open world thus supplies both the knowledge to be learned and a supervision-independent practice environment, with target-task supervision reserved for final evaluation. Across three benchmarks and two target agents, OpenSkill attains the best automated pass rate while satisfying the no-supervision constraint. Analysis shows its skills transfer across models without model-specific adaptation, and its self-built verifier aligns with ground-truth outcomes despite never accessing them.
1 Introduction
Open-world self-evolution requires agents to build reusable skills and verification signals from only a task prompt and open-world resources, without target-task supervision. OpenSkill bootstraps this learning loop through grounded knowledge acquisition, leakage-free skill refinement, and transferable practice environments.
- Limitations: Existing self-evolving agents often rely on curated skills, successful traces, or task feedback, which may be unavailable in real open-world deployments.The open-world setting may expose only a seed task prompt, with no initial skills or verifier for judging improvement.
- Problem definition: Open-world self-evolution starts from only a seed task prompt and open-world resources, requiring agents to build skills and verification signals from scratch.Permitted resources include documentation, repositories, papers, tutorials, and web pages, while hidden target answers, rewards, verifier outputs, and solution traces are excluded.
- OpenSkill: OpenSkill bootstraps a learning loop by acquiring grounded knowledge and verification anchors, then refining skills against self-built virtual tasks rather than target answers.Its three stages are open-world knowledge acquisition, leakage-free skill evolution, and zero-shot target evaluation.
- OpenSkill: OpenSkill defines a practice environment for skill refinement and produces skills that transfer across models without target-task supervision during learning.The framework builds both the skill content and the verification signals needed for self-evolution.
- Results: OpenSkill achieves the best automated pass rate across three benchmarks and two model families while satisfying the no-target-task-supervision constraint.The reported result covers learning without target-task supervision and cross-model transfer.
2 Open-World Self-Evolution
Open-world self-evolution requires an agent to improve from only a task prompt and open-world resources, without initial skills, demonstrations, rewards, or verifiers. OpenSkill bootstraps this process by acquiring knowledge and verification anchors, refining skills on self-generated virtual tasks, and deploying the resulting artifact zero-shot without target-task supervision.
- Open-world setting: Open-world self-evolution starts from only a task prompt and open-world resources, with no initial skills, demonstrations, rewards, or verifiers.The framework formalizes this supervision constraint before presenting its three-stage pipeline.
- Knowledge acquisition: OpenSkill retrieves task-relevant open-world knowledge and synthesizes a structured skill plan specifying architecture, procedures, and domain rules.Retrieved sources include public documentation, code repositories, papers, and tutorials.
- Virtual-task refinement: The pipeline retrieves independently verifiable anchors, then generates deterministic virtual tests grounded in facts such as dataset row counts, metric ranges, and documented output formats.These tests proxy hidden tests without guessing their contents or observing target-task supervision.
- Virtual-task refinement: OpenSkill iteratively executes and evaluates skills against virtual tests, using failure diagnostics to revise implementations or retrieve knowledge for identified gaps.Refinement stops when the virtual pass rate reaches 1 or after J=3 rounds.
- Zero-shot deployment: The final skill is deployed zero-shot as a portable artifact to a target agent, which may differ from the construction agent, while hidden tests enter only at final evaluation.Because skills are explicit artifacts rather than model weights, deployment requires no retraining.
3 Experiment
OpenSkill is evaluated under an open-world, no-target-supervision protocol across three benchmarks and two target agents. It achieves the best automated pass rates across agents and benchmarks, with broad domain gains but a shared failure on Manufacturing.
- Benchmarks: Evaluation covers SkillsBench’s 11 domains plus SocialMaze and ScienceWorld, with hidden ground-truth tests used only for final evaluation.The benchmarks run under the open-world protocol, and ground-truth tests remain hidden during construction.
- SkillsBench results: 43.6% on Opus 4.6 and 42.1% on GPT 5.2 are OpenSkill’s best overall SkillsBench pass rates, beating the strongest baselines by +8.9 and +8.8 points.The strongest baselines are Skill-Creator on Opus 4.6 and CoT on GPT 5.2; results land within 1–3 points of the Human upper bound.
- SkillsBench results: OpenSkill is best or tied-best in 8 of 11 Opus 4.6 domains and 7 of 11 GPT 5.2 domains, with largest gains in knowledge-intensive domains.Highlighted scores include Opus Health 69.6% and Software 59.9%, and GPT Energy 80.0% and Cybersecurity 52.5%, both above Human.
- Limitations: Manufacturing reaches 0.0% for every automated method, showing that open-world acquisition alone does not resolve this domain.This is the main shared per-domain failure despite OpenSkill’s broad gains elsewhere.
- Cross-benchmark results: OpenSkill is the best automated method in all four SocialMaze and ScienceWorld columns: 82.7%/70.7% on SocialMaze and 90.0%/85.3% on ScienceWorld.Scores are reported for Opus 4.6/GPT 5.2 and improve over the strongest baseline by +0.9 to +2.2 points.
4 Analysis
OpenSkill’s skills transfer across weaker models without model-specific adaptation, while its ground-truth-free virtual verifier provides a meaningful quality signal and covers most human-authored test intents. Ablations show that both open-world querying and virtual verification improve performance substantially, but excessive refinement causes overfitting.
- Transferability: OpenSkill-generated skills achieve the highest reward across four weaker models, improving over the no-skill baseline by 5.5%–14.8% points without model-specific adaptation.The same skill files produced by Opus 4.6 are used as-is on Haiku 4.5, Qwen 3 Coder, DeepSeek V3, and Mistral Large 3.3.
- Verifier quality: The virtual verifier achieves 56.9% precision, 80.5% recall, and 60.7% overall agreement with ground-truth evaluation outcomes.Its decisions associate significantly with ground-truth reward despite operating without ground-truth tests.
- Verifier quality: Across 15 sampled tasks, virtual-verifier tests cover 88.9% of ground-truth test intents, or 120 out of 135.The uncovered intents cluster in benchmark-specific anti-cheat checks and deep semantic properties requiring domain expertise beyond task specifications.
- Component contribution: Performance peaks at 3 refinement iterations (82.7%) and declines to 79.9% at 5 and 78.0% at 10 iterations.The decline suggests that excessive refinement introduces overfitting to virtual test feedback.
- Component contribution: Removing both components yields 74.5%, while open-world querying alone reaches 80.6% and virtual verification alone reaches 80.8%.Combining both components achieves 82.7%, a further +2.1 percentage points over the better single-component configuration.
5 Related Work
Prior work advances agent reasoning, tool use, planning, self-evolution, retrieval, and output verification through specialized mechanisms. OpenSkill is situated among these approaches by combining open-world retrieval with skill refinement and supervision-free verification.
- Agent reasoning and self-evolution: LLM agents interleave reasoning and actions, learn to call tools, and improve planning through structured deliberation.Self-evolving agents also accumulate reusable knowledge through reflection, exploration, distilled subagents, and cumulative skills.
- Capability comparison: Capability comparisons distinguish open-world retrieval, iterative skill refinement, supervision-free verification, and explicit model-transferable artifacts.These dimensions frame the capabilities targeted by automated methods in the related-work comparison.
- Open-world retrieval: Retrieval-augmented and browser-assisted agents use external evidence, web search, repositories, and tools for knowledge-intensive or long-horizon tasks.These methods generally retrieve knowledge to answer a query or complete a single task, whereas OpenSkill uses open-world retrieval as part of self-evolution.
- Supervision-free verification: Without target-task supervision, prior methods judge outputs by aggregating reasoning paths, iterating on self-feedback, using LLM judges, or executing self-generated tests.These verification strategies span general reasoning and code domains, where tests can filter or repair solutions through execution feedback.
6 Conclusion
OpenSkill addresses open-world self-evolution by building both transferable skills and verification signals from open-world resources without target-task supervision during learning. Across three benchmarks and two target agents, it achieves the best automated pass rate while honoring this constraint.
- 6 Conclusion: Open-world self-evolution begins with only a task prompt and requires the agent to build its skills and verification signals from open-world resources.The learning process operates without target-task supervision.
- 6 Conclusion: OpenSkill acquires grounded knowledge and verification anchors, synthesizes them into transferable skills, and refines them through self-built virtual tasks rather than target answers.This bootstraps a learning loop from open-world resources without target-task supervision.
- 6 Conclusion: Across three benchmarks and two target agents, OpenSkill attains the best automated pass rate while honoring the no-supervision constraint.The conclusion reports this result across both the evaluated benchmarks and target agents.
Limitations · A Experimental Details · A.1 Model Details
OpenSkill faces open-world risks from unreliable sources and imperfect virtual tasks, while its pipeline combines Claude, GPT, and Gemini models for language-model roles and retrieval. These design choices motivate provenance, validation, and careful separation from target-task supervision.
- Limitations: Web and repository sources may be noisy, outdated, or contradictory, requiring provenance tracking and source validation.
- Limitations: Virtual tasks may not capture the full difficulty of real target tasks, limiting how well they assess skill quality.
- Limitations: Overly easy virtual tasks may overestimate skill quality.
- Limitations: Virtual tasks derived from hidden answers or verifier behavior could reintroduce target-task supervision.
- A Experimental Details: The pipeline uses models from three families for its language-model roles and retrieval components.
- A.1 Model Details: All LLM roles use Anthropic Claude claude-opus-4-6 or OpenAI GPT gpt-5.2.
- A.1 Model Details: Google Gemini handles both open-world retrieval roles: Gemini Deep Research performs main knowledge acquisition, while gemini-3.1-flash-lite supports verification retrieval and targeted retrieval.
A.2 Dataset Details … A.5 OPENSKILL Hyperparameters
The evaluation spans three agentic benchmarks with hidden-test protocols, compares OpenSkill against automated and human-curated skill conditions, and reports benchmark-specific rewards. The appendix also specifies the concrete OpenSkill experimental configuration.
- A.2 Dataset Details: SkillsBench is the primary benchmark, covering 11 domains and isolating skill quality as the task-success bottleneck.The domains are Software, Office, Science, Media, Cybersecurity, Finance, Robotics, Energy, Manufacturing, Health, and Math.
- A.2 Dataset Details: SocialMaze evaluates six social-reasoning subtasks: FTS, HRD, REFT, RDP, SGA, and UPI.The ground-truth test suite is hidden during skill construction and used only for final evaluation.
- A.3 Baselines: OpenSkill is compared with seven automated conditions and one human-curated reference upper bound under shared target agents and hidden-test evaluation.Human Curated skills are excluded from comparisons among the best automated methods.
- A.3 Baselines: The baselines span no-skill, single-pass parametric generation, structured chain-of-thought generation, iterative skill refinement, trace-based skill reuse, and memory-based reinforcement learning.These conditions include No Skill, Self-Gen, CoT, Skill Creator, AutoSkill, and Memento, while SkillNet is included only on ScienceWorld.
- A.3 Baselines: SkillNet is evaluated only on ScienceWorld, whereas Human Curated skills provide expert-authored reference skills for each benchmark.SkillNet’s reported evaluation setting leaves corresponding cells empty on the other benchmarks.
- A.4 Evaluation Metrics: SkillsBench reports average reward as hidden-test pass rate over five independent zero-shot evaluation runs, with per-domain scores averaged.Virtual-verifier analysis additionally reports precision, recall, agreement, and association tests.
- A.4 Evaluation Metrics: SocialMaze uses task-specific accuracy (%) macro-averaged across six tasks, while ScienceWorld uses mean simulator completion score on a 0–100 scale.The paper refers to all metrics uniformly as reward.
- A.5 OPENSKILL Hyperparameters: Table 5 provides the concrete OpenSkill configuration used in the experiments.The table abbreviates model names and refers readers to Appendix B for exact identifiers and component roles.
B Pipeline Implementation Details … B.6 Final Skill Selection
OpenSkill’s implementation separates host orchestration from containerized skill creation and verification, combining leakage-filtered retrieval, independent virtual tests, iterative refinement, diagnosis-driven retrieval, and in-place final skill deployment. The pipeline avoids target-test supervision during construction by isolating the verifier from hidden tests and using externally grounded anchors instead.
- B Pipeline Implementation Details: The pipeline is host-orchestrated, while skill creation and verification execute inside a per-task Docker container.This organization operationalizes the abstract retrieval, test-generation, refinement, and diagnosis functions for reproducibility.
- B.1 Open-World Retrieval D: D synthesizes task-aware research queries, removes benchmark identifiers before retrieval, and uses deep research followed by decomposition into 1–4 skills.Queries use task instructions, metadata, and truncated environment previews while requesting APIs, signatures, defaults, examples, and pitfalls without the solution app.
- B.2 Verification-Knowledge Retrieval Dv: Dv independently retrieves four classes of checkable verification anchors through a single search-grounded generation call, while excluding material already present in D’s background.The anchors include hand-computable reference values, dataset invariants, and other independently checkable information; the first 4000 background characters are marked “Already Known.”
- B.3 Virtual-Test Generation g: An isolated Independent Verifier generates deterministic pytest equality assertions from environment rules or verification knowledge without accessing hidden-test files or ground-truth paths.The verifier shares the container’s outputs but not the creator’s conversation, reasoning, or code, reducing confirmation bias.
- B.3 Virtual-Test Generation g: Virtual pass rate is computed as passed/total with skipped tests excluded, while inherited scripts and failure lists are repaired across rounds under a 60-test cap.The resulting assertions provide an exact, reproducible quality signal for refinement.
- B.4 Iterative Refinement and Termination: Refinement returns structured failure reports and requires the creator to repair skill logic and regenerate outputs, terminating only at ˜r = 1.0 with structural checks satisfied.The implementation caps surrogate refinement at 3 rounds.
- B.5 Gap-vs-Bug Diagnosis and Targeted Retrieval: A classifier distinguishes SELF-FIXABLE implementation bugs from NEEDS-DR knowledge gaps, triggering unaided fixes or targeted retrieval respectively, with targeted searches capped at 3 per task.If classification fails, the system conservatively defaults to SELF-FIXABLE and performs no retrieval.
- B.6 Final Skill Selection: The deployed skill set is the most recently refined in-place state at loop termination, exported from the container and evaluated in 5 independent zero-shot runs scored by T GT.It is not selected as a best-of-N snapshot by virtual pass rate.
C Failure Modes of Virtual Verifier
The virtual verifier disagreed with ground-truth evaluation in 33 cases: 25 false positives and 8 false negatives. False positives mainly reflected strict evaluation or incomplete semantics, while false negatives arose from overly strict surrogate tests or verifier infrastructure failures.
- Disagreement Overview: 33 disagreement cases comprised 25 false positives and 8 false negatives between the virtual verifier and ground-truth evaluation.These cases were analyzed to characterize the verifier’s failure modes.
- False Positives: 12 false positives were high-accuracy near-misses with mean accuracy 0.81, passing most ground-truth tests but receiving zero reward under all-or-nothing evaluation.The verifier recognized largely correct outputs but missed remaining edge-case failures caught by the ground-truth suite.
- False Positives: 11 false positives reflected partial correctness, with mean accuracy 0.52, where structurally valid surrogate-test outputs remained semantically incomplete.The passage gives well-formed but incomplete JSON as an example.
- False Negatives: 3 false negatives were near-passes with 89–97% surrogate pass rates, failing one or two overly strict assertions despite succeeding on ground-truth evaluation.Their unmet surrogate tests targeted edge cases absent from the ground-truth suite.
- False Negatives: 5 false negatives resulted from verifier infrastructure failure, producing 0% surrogate pass rates even though agents generated correct solutions.These cases were predominantly non-standard tasks such as CVE patches, build-system fixes, and proof assistants, which challenged the pytest-based pipeline.
D SocialMaze Per-Subtask Results
Table 6 reports substantial variation across the six SocialMaze subtasks, with OpenSkill achieving the best overall average for both target agents. Its gains are driven mainly by harder reasoning subtasks rather than uniform improvements.
- D SocialMaze Per-Subtask Results: OpenSkill achieves the best overall SocialMaze average on both target agents: 82.7% on Opus and 70.7% on GPT.The results aggregate FTS, HRD, REFT, RDP, SGA, and UPI.
- D SocialMaze Per-Subtask Results: Per-subtask scores vary widely across methods on the six SocialMaze subtasks.The subtasks are FTS, HRD, REFT, RDP, SGA, and UPI.
- D SocialMaze Per-Subtask Results: OpenSkill’s average is driven largely by REFT and UPI on Opus, and by REFT and RDP on GPT.These harder reasoning subtasks account for much of the advantage rather than uniform gains across all subtasks.
E Computational Cost · F Information Isolation Audit · G Baseline Prompts
OpenSkill’s skill creation is costly but reusable, while its evaluation cost is comparable to the No-Skill baseline. A four-layer audit enforces information isolation, and self-generated skill baselines remain weak without external verification.
- E Computational Cost: Skill creation consumes 749K of 1.14M tokens (66%) but only 39 of 131 minutes (30%), because verification is sequential while GT evaluation uses five independent runs.Across 84 tasks, skill creation totals 31.4 API-hours and ∼47M tokens.
- E Computational Cost: The full end-to-end pipeline totals 140 hours and ∼97M tokens, costing an estimated ∼$1,800 at Opus 4.6 list pricing.The estimate uses $15/M input and $75/M output pricing.
- E Computational Cost: Generated skills incur zero additional creation cost when reused across models and runs, with cross-model evaluation taking 16–27 minutes per task.Reported times are Haiku: 17.7 min, DeepSeek: 18.1 min, Qwen 3 Coder: 15.7 min, and Mistral: 27.0 min.
- E Computational Cost: OpenSkill requires a median of 368.2 s per evaluation run, comparable to the No-Skill baseline’s 465.0 s average.The eval-only mean is 845.4 s because a small number of tasks run for unusually long durations.
- F Information Isolation Audit: The Virtual Verifier enforces strict isolation from GT tests, solutions, and the skill creator through code-level, container-level, oracle-bypass, and log-level controls.The base agent receives only the task instruction and environment files; the verifier uses a separate LLM session, and GT artifacts are excluded from containers and trajectories.
- F Information Isolation Audit: During skill creation, the GT oracle is never invoked; only a single GT-derived pass/fail bit may guide verifier reinvocation, without exposing test content or expected values.Post-hoc GT fields are recorded after agent exit and are never fed back to agents.
- G Baseline Prompts: The Self-Generated Skills prompt requires in-session analysis, creation of 1–5 markdown skill files under environment/skills/, and task solving using those skills without external verification.This replicates SkillsBench’s self-generation condition.
- G Baseline Prompts: The CoT-Guided Self-Generation prompt adds a structured five-step workflow but still lacks external verification, achieving only 30.7% pass rate.Its workflow covers task analysis, skill architecture, progressive-disclosure skill writing, self-verification, and execution.