SkillDisCo paper: distilling agent traces into skills

SkillDisCo is a distillation-and-compilation framework, introduced in an arXiv paper submitted 25 Jun 2026, that extracts reusable procedural skills from successful agent traces and compiles them into callable, executable, and verifiable modules. The paper, arXiv:2606.26669, is authored by Zhongxin Guo, Danrui Qi, Hanwen Gu, Peng Cheng and Yongqiang Xiong.

What is SkillDisCo?

SkillDisCo distills reusable PFSM subgraphs from successful traces and compiles those subgraphs into procedural skills that agents can call and execute. The authors frame task settings as FSM-defined scenarios, view successful traces as paths in an unknown transition graph, and define procedural skills as parameterized control-flow subgraphs (PFSM subgraphs) that can be reused across instances.

The framework performs two linked steps: distillation, which identifies shared PFSM subgraphs across successful agent traces, and compilation, which turns those subgraphs into callable, executable, and verifiable procedural skills. That pipeline is intended to capture shared execution structure so agents need not re-solve similar task instances from scratch.

How does SkillDisCo perform on benchmarks?

Experiments on ALFWorld and WebArena show SkillDisCo improves success rates and reduces agent turns across benchmarks and model scales. The paper reports empirical gains on those two evaluation suites, demonstrating that representing shared experience as reusable execution structures provides measurable benefits in practice.

The evaluation context is FSM-defined scenarios where successful traces correspond to paths in an unknown transition graph; SkillDisCo seeks PFSM subgraphs that recur across those paths and compiles them into skills the agent can invoke. The authors present results across benchmark tasks and different model scales to support the claim that distilled procedural skills cut down redundant reasoning and shorten execution traces.

Why it matters

Agents frequently re-solve similar task instances from scratch, which the paper identifies as causing unnecessary reasoning cost and long execution traces. By turning repeated subsequences of successful behavior into callable procedural skills, SkillDisCo targets that waste directly: reusable skills encode shared control-flow so agents can reuse prior solutions instead of re-planning every step.

That matters for agents operating in structured environments: if common subroutines can be reliably detected and executed, planners and learned policies can focus compute on novel decisions rather than repeating solved routines. The paper’s contribution is both conceptual—defining skills as PFSM subgraphs in FSM-based tasks—and practical, with experiments on ALFWorld and WebArena showing improvements in success rates and fewer agent turns.

What to watch

Check the arXiv entry (arXiv:2606.26669) for code, demos, or data links; the paper’s page includes toggles for demos and links to code and data resources. Watch for follow-up work that applies SkillDisCo beyond FSM-defined scenarios or that measures gains in larger, more open-ended environments.

The submission date and bibliographic record are concrete early signals: the paper was submitted on 25 Jun 2026, and the arXiv DOI is https://doi.org/10.48550/arXiv.2606.26669. Those entries will likely list artifacts and further experiments as the authors publish source code or additional evaluations.