DisCo research agent closes operational knowledge gap in large language models

The DisCo research agent closes a fundamental knowledge gap in large language models: the ability to turn theoretical knowledge into working code. Researchers identified a missing layer, "operational knowledge," that separates knowing a method from the ability to implement it in practice. That knowledge exists in GitHub repositories and papers, but in a format designed for human reading and too large to load at task time. The solution: distill the knowledge into compact, verified "skills" that can be retrieved at runtime.

How the mechanism works

DisCo operates through two complementary distillation pipelines. The first, task-independent, processes one thousand popular ML repositories and extracts a library of more than five thousand verified skills, organized into twenty domains and one hundred seventy-eight capability families — the AREX-Skill library. The second, task-dependent, generates the skills required for a specific task in real time. Both pipelines run on the same skeleton: a GPT-5.5 backbone, a fixed research harness, and an identical compute budget.

The numbers on benchmarks

In this configuration, the skill-equipped agent achieves significantly higher results than the exact same agent without them. On MLE-bench, a benchmark that tests end-to-end ML engineering ability, a 134.3% improvement was recorded. On PaperBench, which examines paper reproduction, the gap stands at 34.4%. FrontierCS, focused on advanced computer-science tasks, shows a 9.2% increase, and PassNet, which tests knowledge transfer between tasks, improves by 14.0%. All of these numbers are achieved while the model, the harness, and the budget remain constant; the only difference is the addition of distilled operational context.

What this means in practice

The implication is that the bottleneck for autonomous research agents is not necessarily the model's reasoning ability, but access to compressed practical knowledge. Instead of forcing the model to rediscover implementation tricks on every run — hyperparameter settings, training-stability tricks, efficient data structures — the system retrieves them as a ready-made reference library. The approach resembles the difference between a junior engineer who has read a textbook and a veteran who can recall the specific library's gotchas from memory.

Direction for future work

The work points toward a direction in which ML research automation relies not only on larger models, but on structured knowledge infrastructure built from the open ecosystem. The AREX-Skill library is available for use, and the dual-track methodology allows it to be extended to any new domain without retraining the model. The open question is how far the approach scales when tasks move beyond the boundaries of known open source, but as a starting point, the gap between "knows the method" and "can run it" has narrowed dramatically.