An 8B Model Beats 4 Frontier LLMs by 25%+ — By Mining Its Own Skill Bank

Notes on Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Tasks (arXiv:2604.20987) — Xiyang Wu, Zongxia Li, Guangyao Shi, Alexander Duffy, Tyler Marques, Matthew Lyle Olson, Tianyi Zhou, Dinesh Manocha · April 2026

Note published July 24, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

COSPLAY runs two roles inside one system: a decision agent that plays a game by drawing on a skill bank, and a skill-pipeline agent that mines the decision agent's own unlabeled rollouts, extracting new skills and refining the ones already banked — each one shipped with a skill contract instead of a bare snippet. The two roles co-evolve: a better bank improves play, and better play produces better rollouts to mine next. The proving ground is long-horizon tasks — game environments where success depends on chaining skills across many timesteps under delayed reward and partial observability. Six such environments were tested, split between single-player and multiplayer social-reasoning games, with an 8B base model measured against four frontier LLM baselines.

The payoff splits down the middle. On single-player benchmarks, the 8B COSPLAY agent beats all four frontier baselines by an average of more than 25.1% in reward, with no increase in model scale — the gain traces to the retrieval-and-refinement loop, not bigger weights. On multiplayer social-reasoning games, the same setup only holds even with the baselines rather than surpassing them. The pattern suggests the self-mined skill bank pays off most when a task decomposes into reusable procedural sub-skills, and pays off least when success hinges on modeling another agent's behavior.