A group of eleven researchers posted a paper to arXiv on 2 September describing DisCo, a method that reads a GitHub repository and distills its operational knowledge, the setup steps, the gotchas, the working configurations, into a reusable skill an agent can load.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source Applied to about 1,000 widely used machine learning repositories, it produced the AREX-Skill Library: more than 5,000 verified skills across 20 research areas and 178 package families.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source 3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source With the GPT-5.5 backbone, the research harness and the execution budget all held fixed, the skill equipped agent scored 134.3 percent higher on MLE-bench, moving from 31.11 to 72.89 percent, and gained 34.4 percent on PaperBench, 9.2 on FrontierCS and 14.0 on PassNet.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source 3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source Our assessment, with high confidence, is that this is the clearest demonstration yet that procedural knowledge, not model weights, is the binding constraint on machine learning agents; with moderate confidence, that the same method transfers to any domain with a large body of working code and documentation; and with moderate confidence that the finding is a problem for labs whose pricing assumes capability lives in the model.
What a skill is
The library's unit is a folder: a SKILL.md file describing when and how to do something, plus references and scripts.3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source DisCo produces them two ways. Task agnostic distillation reads a repository and extracts patterns that apply broadly, such as how a training framework expects data to be laid out. Task oriented distillation starts from a downstream task and extracts the procedure that solves it.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source 2 AI Weekly 2026-09-04 Two skill modes; gains with weights unchanged; framing that frontier models have not captured procedural knowledge. Open source The result is closer to a senior engineer's notes than to a manual: what works, what breaks, what the README does not say.
The library was built in public. It launched on 3 August 2026 and reached 1,000 repositories on 27 August; the repository is Apache 2.0 at the top level, with each skill carrying the license of the code it was distilled from.3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source The paper itself is under a more restrictive CC BY-NC-SA 4.0.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source Coverage spans ML engineering, large language models, computer vision, data science, scientific computing, deployment, training infrastructure, robotics, generative media and biomedical AI.3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source
Why the fixed backbone matters
The experimental design is the finding. Everything about the model was held constant, and the only variable was whether the agent could load skills.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source On MLE-bench, a benchmark of Kaggle style machine learning competitions, the score more than doubled.3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source On PaperBench, reproducing research papers, it rose by a third. On FrontierCS and PassNet the gains were smaller but present.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source
AI Weekly's reading is that frontier models are far from capturing what domain specific procedural knowledge can add.2 AI Weekly 2026-09-04 Two skill modes; gains with weights unchanged; framing that frontier models have not captured procedural knowledge. Open source That is the polite version. The blunt version is that a model trained on all of GitHub still does not know how to use GitHub's contents, and that a few thousand structured notes close much of the gap. The knowledge was always there; it was not in a form the agent could act on.
Who gains and who loses
Anyone running coding agents on machine learning tasks gains an Apache 2.0 library that, on the authors' numbers, more than doubles success on a standard benchmark for the cost of a download.3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source Agent framework vendors gain a template for skills as a product category. The eleven authors, whose affiliations the arXiv page does not list, gain a citation magnet; the VectorSpaceLab organization hosting the library gains a widely used artifact.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source 3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source
The frontier labs lose a little of the argument that capability is what you pay them for. If a fixed GPT-5.5 can be made twice as good at a task class by an external library, the value of the next model increment on that task class is reduced, and the value of whoever curates the skills is increased.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source Benchmark designers lose signal: MLE-bench with skills and without are now different tests, and leaderboards will have to say which. Repository maintainers lose control of how their operational knowledge is repackaged, though the license passthrough is designed to respect that.3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source
The counter case
The gains are the authors' own evaluation, and the largest one is on a benchmark whose tasks are close in kind to the repositories the skills were distilled from; machine learning competitions are exactly what a machine learning skill library should help with.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source The smaller gains on FrontierCS and PassNet may be the honest measure of transfer beyond the home domain.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source There is also a contamination question the paper will need to answer: if any of the 1,000 repositories contain solutions to MLE-bench tasks or their close relatives, a skill is a shortcut rather than knowledge. And the fixed backbone design proves the library helps GPT-5.5; it does not prove that a model trained with these skills in its data would not have learned them, in which case the next model absorbs the gain and the library's advantage is temporary.2 AI Weekly 2026-09-04 Two skill modes; gains with weights unchanged; framing that frontier models have not captured procedural knowledge. Open source
What to watch
- Independent replication on MLE-bench. A third party reproducing a score near 73 percent with the public library and a GPT-5.5 agent within two months would confirm the headline; a much smaller gain would suggest harness effects.3 GitHub, VectorSpaceLab 2026-08-27 Skill folder structure SKILL.md plus references and scripts; raw scores 31.11 to 72.89 MLE-bench, 29.45 to 39.59 PaperBench, 70.63 to 77.14 FrontierCS, 1.343 to 1.531 PassNet; Apache 2.0 at repo level; launched 3 August, 1,000 repos on 27 August; VectorSpaceLab. Open source
- A contamination audit. Publication of an overlap analysis between the 1,000 source repositories and benchmark tasks would settle the shortcut question.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source
- Skills in a lab product. If OpenAI, Anthropic or Google ship a first party skill library or absorb AREX style skills into an agent product within six months, the finding has been accepted as real.2 AI Weekly 2026-09-04 Two skill modes; gains with weights unchanged; framing that frontier models have not captured procedural knowledge. Open source
- The next model closes the gap. If a successor to GPT-5.5 scores near 73 percent on MLE-bench without skills, the knowledge has moved into the weights and the library's edge was a snapshot.1 arXiv 2026-09-02 Eleven authors; submitted 2 September 2026; DisCo with task agnostic and task oriented distillation; 5,000 plus skills from about 1,000 repos in 20 areas and 178 families; GPT-5.5 fixed; 134.3 percent MLE-bench, 34.4 PaperBench, 9.2 FrontierCS, 14.0 PassNet; CC BY-NC-SA 4.0 paper. Open source
The model did not get smarter. It got the notes. For anyone selling intelligence by the token, that is the uncomfortable part.