OpenSkill for Open-World Self-Evolution in LLM Agents

Mode

Step

task and deployment boundary open-world evidence and knowledge durable skill artifact proxy rehearsal and diagnosis

View

low / weak middle / ambiguous high / hot

What To Watch

Proxy tests are useful before they are faithful

High intent coverage plus much lower full agreement means the verifier often sees the right target, yet still misses deeper semantics or anti-cheat requirements.

Strong

API calls, argument schemas, explicit invariants, and reference values grounded in public artifacts.

Weak

Semantic quality, hidden edge cases, and the parts of the task that look obvious only after you already know the answer.

Operational Burden

Leakage control is filters, blocklists, audits, and retrieval policy, not a one-line promise.

The episode flags missing operational details: search provider, query templates, dedup policy, blocklists, and stopping criteria still matter if the benchmark is public.

View

OpenSkill baseline or transfer source verifier or component pressure

Interpretation

Cold-start gains, bounded confidence

The strongest evidence is the SkillsBench jump and the claim of best automated results on SocialMaze and ScienceWorld. The weakest point is still verifier independence, not the idea of writing skills down.

Reported

SkillsBench pass rates and verifier metrics use the exact numbers quoted in the episode.

Derived

Strongest automated baselines are back-computed from the stated deltas: 34.7 and 33.3.

Illustrative

Transfer and ablation views are labeled mock data because the transcript gives qualitative claims, not full tables.

Read the page as a map of what the paper is trying to prove: not “agents can act,” but “agents can keep a reusable improvement artifact after acting.”

View

verbal critique / judge loops environment or trajectory-heavy skill learning memory and skill-structure systems open-world grounded skill evolution

References

Paper Graph

Core papers and evaluation threads cited in the episode. arXiv links point directly to the preprints when available; later 2026 systems without public arXiv IDs use the provided Scholar lookup URLs.