BCSD trains LLM agents to actually consult their skill libraries

Handing an LLM agent a library of skills is the easy part; getting it to actually use them well is where things break down. That is the problem targeted by a new arXiv paper from Tianjun Pan and a group of eleven authors. Their framing is that in agent RL, "task-level rewards offer limited supervision and struggle to capture subtle differences in how effectively the policy uses the provided skills."

Their proposed fix, called BCSD, evaluates each trajectory through two complementary lenses. An augmented view adds higher-level Meta-Skill guidance on top of the standard prompt; a reduced view prunes general guidance and leaves the task-specific skills bare. The gap between what the policy does in each condition becomes a token-level signal that rescales the RL advantage function. In short, the model gets rewarded not just for finishing the task but for behaving differently when the guidance changes, which serves as a proxy for actually reading the skills it was given.

The reported result is "strongest overall performance across model scales" on ALFWorld and WebShop, two long-running text-based agent benchmarks. That framing is thin on numbers: the abstract as retrieved does not publish specific gains against baselines, name the base models used, or describe how BCSD compares against the several other self-distillation approaches circulating in the same window. The paper is nine pages with six figures, so the ablation detail lives inside the PDF rather than in the abstract. Two researchers in our tracked directory have already shared the link, a mild signal that the specifics are worth digging into.

For teams building on open agent frameworks, the interesting bit is that this is training-side rather than prompt-side. If dual-context self-distillation holds up under scrutiny, it becomes another lever for making smaller open models behave more like disciplined tool users, which is the practical bottleneck for a lot of agent stacks today.