arxiv arXiv cs.AI · just now · research

DisCo distills GitHub repos into 5,000+ AI research skills

from English

Researchers present DisCo, a skill-powered autonomous agent designed to bridge the gap in operational knowledge for machine learning research by distilling expertise from code repositories. The system condenses widely used open-source projects into reusable, verified skills through two complementary forms: task-agnostic distillation of general ecosystems and task-oriented generation for specific needs.

DisCo creates the AREX-Skill Library, containing over 5,000 verified skills distilled from 1,000 ML repositories organized into 20 areas and 178 capability families.
When equipped with these skills and using a GPT-5.5 backbone, the agent achieves a 134.3% score increase on MLE-bench compared to the same agent without skills.
The skill-equipped agent also improves performance by 34.4% on PaperBench, 9.2% on FrontierCS, and 14.0% on PassNet under fixed research harness conditions.

These results demonstrate that adding distilled operating context allows autonomous agents to reuse domain-specific know-how rather than rediscovering it during each run.

Importance 2/3 New harness with differentiators arXiv cs.AI OpenAI AI agents Research paper

Related articles

arxiv arXiv cs.CL · just now Live

DisCo distills GitHub repositories into AI4AI skills to boost research agent performance

The authors present DisCo, a skill-powered research agent designed to bridge the gap in operational knowledge by distilling domain-specific know-how from external sources. The system condenses widely used machine learning repositories into reusable, verified skills through two complementary forms: task-agnostic distillation for broad utility and task-oriented distillation for specific needs.

arxiv arXiv cs.AI · 23d ago · 7 views

Open-Ended Optimization lets GPT-5.5 compose its own improvement process

The article introduces Open-Ended Optimization (OEO), a framework that allows a frontier model optimizer to dynamically compose its improvement process online, rather than relying on prescribed optimization pipelines. The authors compare OEO against two staged approaches, SkillOpt and GEPA, across 14 head-to-head comparisons over eight benchmark-target-model settings.

lab OpenAI News · 33d ago · 25 views

Astra model solves ten open problems in mathematics and theoretical computer science

OpenAI's internal Astra model has generated solutions to ten long-standing open problems across mathematics and theoretical computer science, with the arguments formalized in Lean. These results span high-dimensional geometry, coding theory, group theory, and quantum complexity, addressing questions that had seen no progress for at least a decade.

lab OpenAI News · 70d ago · 15 views

OpenAI Research Shows AI Agents Transforming Work

A new research paper from OpenAI demonstrates how artificial intelligence agents are fundamentally changing the nature of work. The study highlights the capability of these agents to execute longer and more complex tasks than previously possible. This technological advancement is credited with expanding productivity across a wide variety of professional roles. The findings suggest a significant shift in how labor is organized and performed through automation. By handling intricate workflows, AI agents are enabling users to achieve greater efficiency. The paper serves as evidence of the growing impact of autonomous systems on modern employment.

arxiv arXiv cs.CL · 72d ago · 7 views

OpenBioRQ: Benchmark for Agentic Biomedical Research Faithfulness

OpenBioRQ introduces a benchmark of 12,553 unsolved biomedical research questions across 12 domains, designed to test agentic models' faithfulness and abstention. It evaluates models in a tool-using setting without answer keys, using real follow-up evidence rather than parametric knowledge, and reveals significant agentic collapse on the hardest questions where tools are no longer used despite being critical.