SkillLift cuts agent skill-evolution token cost 40-70% by ranking, not rollouts
dair_ai · x · 2026-09-22
A paper introduced by dair-ai shows agents can improve their skill prompts by ranking candidates with a learned rubric instead of scoring every revision with a full rollout — the rollout cost is why skill self-evolution usually only patches observed failures. SkillLift trains a rubric to agree with real outcomes on pairwise skill rankings; an inner loop revises skills against the frozen rubric at no rollout cost, while an outer loop spends a few real rollouts to re-align the rubric by rank correlation. On SkillsBench and WildClawBench (147 tasks) across three models, it beats SkillOpt and CoEvoSkills in all six combinations, with 40–70% less token cost than frontier evolving methods.
More from coding & agent
- MechFaber: Claude Code designs a 99-part quadruped with firmware co-simulated in Renode and MuJoCo — SpeedyBrowser45 · 2026-09-22
- Exa MCP hits 5,000 GitHub stars as AI agents flock to its search integration — TheIshanGoswami · 2026-09-22
- 670,000 agent skills, no trust layer: bot scan finds 69% never reliably fire — markjeffrey · 2026-09-22
- Eight disruptive use cases for Jev: from millisecond evals to AI guardrails — nkmrao · 2026-09-22
- Training on production traces: single-trajectory RL may unlock continual learning — rhythmrg · 2026-09-22
- Anthropic's Swiss cheese model explains why passing evals isn't enough for agents — hugobowne · 2026-09-22