DisCo reports reusable skills raise MLE-bench from 31.11% to 72.89%

• DisCo reports raising MLE-bench's Any Medal rate from 31.11% to 72.89% with reusable repository-derived skills, according to rohanpaul_ai's DisCo summary.
• The reported comparison holds the GPT-5.5 backbone, research harness, and execution budget constant in the linked Repo-To-Skill paper.
• Each skill packages operating instructions with references, scripts, validation, and recovery guidance, as TheTuringPost's DisCo thread details.
• SkillGLoW takes a separate approach to long-running agents, reporting a 17.2-point hard-task gain while keeping its skill library 3.6 times smaller, per rohanpaul_ai's SkillGLoW overview.

The Repo-To-Skill paper calls the missing layer in research agents “operational knowledge”: the setup and repair know-how trapped in repositories and papers. The AREX-Skill repository includes a one-minute demo of a router narrowing thousands of skills to a task-specific branch.

DisCo separates that knowledge from both the model and its harness. It distills it in two modes:

1. Task-agnostic distillation turns widely used ML repositories into a shared library.
2. Task-oriented distillation generates skills for the research task currently in front of the agent.

The resulting AREX-Skill library is organized into 20 areas and 178 capability families. The collection stood at 5,353 skills in rohanpaul_ai's paper summary, while the project headlines the library as 5,000-plus verified skills from 1,000 repositories.

A repository with several capabilities becomes a skill graph rather than one giant context file. A router resolves a request to a relevant area, family, repository, workflow, and finally the branch the agent needs.

Each branch is centered on SKILL.md and can carry:

• focused reference material
• executable scripts
• validation checks
• failure-recovery guidance

The agent loads the relevant branch and follows linked skills as needed. The project says the format can be exported to Codex, Claude Code, and other compatible agents.

In the Repo-To-Skill paper, the authors compare the same GPT-5.5 agent, harness, and downstream budget with and without the distilled skill layer. The paper reports:

• MLE-bench Any Medal, according to rohanpaul_ai's paper summary: 31.11% to 72.89%, +41.78 percentage points and +134.3% relative.
• PaperBench: +34.4%.
• FrontierCS: +9.2%.
• PassNet: +14.0%.

Those are component-comparison results: the experimental variable is the added operating context, rather than a different model or larger execution allocation.

SkillGLoW addresses a different failure mode: a self-improving agent accumulating either one over-general document or a sprawling pool of task memories. It keeps a generalized global prior for each procedural family, while reconstructing task-specific details locally.

Its flow is compact:

1. Generate a local skill while solving a task.
2. Group related procedures into a family.
3. Compress the family into a de-instantiated global prior.
4. Regenerate the instance details for the next task.

On hard tasks, rohanpaul_ai's SkillGLoW notes reports a 17.2-point average improvement and a library roughly 3.6 times more compact than a per-task skill pool.

SkillGLoW adds a commit gate: a candidate prior is adopted only after real execution shows it does not degrade performance. The paper reports transfer on unseen ALFWorld tasks from 73.9% to 83.9%.

HarnessDev evaluates a related but distinct ability, agents building and evolving their own execution harnesses. Its authors found generated harnesses varied widely in capability and token cost, lagged mature human-engineered references on code and search or research, and transferred only partially across models and held-out tasks.