What they did

The authors developed DisCo, a research agent that creates and applies reusable skills encoding operational knowledge—the practical steps required to make methods work. Its pipeline supports both task-agnostic distillation from widely used repositories and task-oriented distillation for a specific research problem.

They used the task-agnostic pipeline to build the AREX-Skill Library: more than 5,000 verified skills derived from 1,000 machine-learning repositories and organized into 20 areas and 178 capability families. They compared skill-equipped and skill-free versions of the agent while holding the GPT-5.5 backbone, research harness, and downstream execution budget fixed.

Key findings

• Adding distilled skills increased the agent’s MLE-bench score by 134.3% relative to the no-skill version.
• Scores also rose by 34.4% on PaperBench, 9.2% on FrontierCS, and 14.0% on PassNet.
• The resulting library contains more than 5,000 verified skills spanning 178 capability families across 20 machine-learning areas.
• Because the backbone, harness, and execution budget were fixed, the reported differences isolate the contribution of the added operating context within this setup.

Why it matters

The results suggest that an agent’s capabilities depend not only on its model and orchestration harness, but also on whether practical domain knowledge is packaged into compact, retrievable instructions. Repository-to-skill distillation could let research agents reuse implementation experience rather than repeatedly reconstructing it from large codebases and papers.

Caveats

The abstract reports relative improvements but not absolute scores, uncertainty estimates, distillation costs, or detailed verification procedures. The evaluation also uses one stated backbone and focuses on machine-learning research benchmarks, leaving open how well the approach transfers to other models, domains, and genuinely novel tasks.