RSIHub: A Composable Framework for Agent-Evolution Methods

One-line takeaway

RSIHub enables composable, comparable agent evolution methods with up to 13.5pp improvement, though gains vary across tasks.

RSIHub框架通过可组合的组件设计，支持多种智能体进化方法在共享基础设施上高效对比与迭代，最大提升达13.5个百分点，但性能迁移存在差异。

Key points

RSIHub is a composable, open-source framework that enables researchers to build, compare, and extend agent evolution methods using reusable recipes and shared infrastructure. It supports multiple methods like AHE and Hyperagents without core modifications, offering standardized evaluation, lineage tracking, and component reusability. Experiments on MiniSWE and Codex show up to 13.5 percentage point gains in pass@1, though performance transfer across tasks is inconsistent.

• RSIHub enables composable agent evolution methods with shared infrastructure and reusable recipes.
• Multiple evolution strategies (AHE, Hyperagents, etc.) are implemented without modifying core framework code.
• Up to 13.5pp improvement in pass@1 on Terminal Bench 2 and τ³ Banking is observed under fixed model and evaluator settings.
• Performance gains do not consistently transfer across held-out tasks, indicating setting-specific behavior.

Written only from the official arXiv abstract by an open-source model, then checked automatically (every number and name must appear in the abstract). The paper is authoritative.