PluginRSI: Recursive Improvement of Agent Harnesses with Reusable Plugins
PluginRSI evolves agent harnesses as reusable plugins that transfer across models and tasks.
PluginRSI treats a language-model agent harness as a composition of atomized, reusable plugins rather than a single program. Plugins are improved independently, stored in a shared library, and recombined into new harnesses at each iteration. The method beats prior harness optimization on software engineering, command-line interaction, and question answering, and the resulting harnesses keep their edge when moved to other solver models. The accumulated library also speeds later optimization and improves convergence on unseen tasks.
- Harnesses are compositions of independently improved plugins stored in a shared library.
- Plugins recombine each iteration instead of searching over complete programs.
- Gains hold on software engineering, command-line, and question-answering tasks.
- Optimized harnesses transfer to other solver models without further tuning.
- A stored plugin library speeds later optimization on unseen tasks.
From the original article
1 min read at huggingface.co
The harness surrounding a language model is a central determinant of agent performance. Recent methods optimize harnesses by searching over complete programs, where individual mechanisms are difficult to isolate and reuse. We introduce PluginRSI, which represents a harness as a composition of atomized plugins and organizes harness evolution around these plugins.
By Yaorui Shi, Yuchun Miao, Yuxin Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.