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

AI-assisted notes · reviewed by SkillFed Skill evolution

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.

Key numbers

Average reward improvement (single-player, 8B model)	25.1%+
Game environments tested	6
Frontier LLM baselines compared against	4
Base model size	8B

What this means for skill authors

The detail worth sitting with is the contract. COSPLAY doesn't just cache a snippet of successful behavior — it attaches conditions for when and how that snippet applies, and keeps refining both the snippet and its conditions as new rollouts arrive. That's the same distinction SkillFed's vetted catalog draws between a skill and a trick: a skill needs metadata about its own boundaries to be reusable by an agent that didn't write it. Here that metadata is machine-generated rather than author-written — a preview of where automated skill mining is headed: self-authored contracts, no human curator required by default.

The more useful finding for anyone weighing a bigger model against a better toolset: an 8B model with a self-refining skill bank outperformed four frontier baselines on single-player long-horizon reward by upwards of a quarter. That's an argument for architecture over scale, on exactly the class of task skills are supposed to help with: multi-step, chainable work. But a skill bank mined from one agent's rollouts is only as good as the tasks that agent actually attempted — nothing here claims it transfers to environments the agent never played.

The multiplayer caveat is the honest part of the result: take 'competitive' at face value, not 'improved.' Skill-chaining architectures earn their keep on decomposable, mechanic-driven tasks; they don't obviously help with reading another mind. Authors building skills for negotiation or multi-agent coordination shouldn't assume a rollout-mining pipeline like this one will surface the sub-skills that matter — it was built and validated on tasks where the environment, not another agent, is the main source of complexity.

Skills related to this research

seedance-recipes Seedance-recipes provides production-ready recipe patterns for video content across genres: product, lifestyle, drama, music video, landscape, commercial, animation, and more. Each recipe preserves core creative constraints while inviting customization of subject, camera, lighting, and sound. Use recipes as proven starting shapes, not rigid templates.★ 5,445 Human Writing Human Writing identifies and removes telltale signs of AI-generated text—promotional language, superficial analyses, vague attributions, and overused vocabulary—so your writing sounds like it came from a real person. Beyond just stripping AI patterns, it guides you to add genuine voice: opinions, varied rhythm, specificity, and personality that make prose feel alive.★ 97 Inno Humanizer Inno Humanizer identifies and eliminates signs of AI-generated writing—including promotional language, vague attributions, excessive em dashes, and superficial analyses—to restore a natural voice to your text. Beyond pattern removal, it helps inject genuine personality and perspective into writing that might otherwise feel sterile or algorithmic.★ 1,038 humanizer Humanizer scans text for telltale AI patterns—vague attributions, excessive em dashes, promotional phrasing, and passive constructions—then rewrites them to sound genuinely human. It can match your own writing voice if you provide a sample, and goes beyond pattern removal to inject personality and opinion where the writing feels sterile.★ 112

Related notes

References

Wu, Li, Shi, Duffy, Marques, Olson, Zhou, Manocha. "Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Tasks." arXiv:2604.20987 (2026).