SR-Agents
A benchmark and toolkit for skill-retrieval-augmented LLM agents that treats reusable agent skills as a large corpus to retrieve from, rather than listing every skill in the prompt.
SR-Agents studies Skill Retrieval Augmentation: letting an LLM agent retrieve reusable skills, such as natural-language instructions, tool-use procedures or executable code, from a large external corpus instead of listing every available skill in the prompt, which runs into context-window limits and selection failures as skill libraries grow.
The release includes SRA-Bench, a benchmark for the decomposed evaluation of skill retrieval, incorporation and application, with 5,400 capability-intensive test instances, 636 manually constructed gold skills mixed with over 25,000 web-collected distractor skills, and six task families covering theorem application, logical reasoning, tool use, medical calculation and code generation. The benchmark evaluates the full pipeline: retrieval, incorporating the retrieved skills into the agent's problem-solving state, and end-task execution.
It ships as a research toolkit with data, baseline agents and evaluation pipelines for studying retrieval-based skill augmentation.