ModularRSI
A modular, contrastive framework for recursively evolving an AI agent's harness (its scaffolding of tools and control logic) rather than its model weights.
ModularRSI is a method for recursive self-improvement (RSI) applied to an AI agent's harness: the scaffolding of prompts, tools, and control logic around a model, distinct from the model's weights. It was introduced in a paper posted to arXiv on 14 September 2026 (arXiv:2609.14857) by Siwei Wu and 13 coauthors. It differs from the general RSI entry in that it is a specific, benchmark-disjoint, contrastive, and modular technique aimed at the harness rather than the weights or, as in Dream-RSI, the exploration policy.
The method contrasts successful and failed trajectories from the same task and aggregates the differences across many tasks to isolate recurring behavioral deficiencies. It splits the harness into five functional modules: agent loop, tool use, observation management, context management, and task completion detection. Each module evolves independently within a restricted scope. An integration stage then merges the modules and resolves conflicts. The authors curated 2,000 executable evolution tasks from external sources, disjoint from the evaluation benchmarks, to measure generalization rather than memorization.
On TerminalBench 2.0 the evolved harness raised accuracy from 47.57% to 52.43%. On SWE-Bench Verified it went from 73.40% to 76.45%, with gains holding on unseen in-domain and cross-domain tasks. It also transferred to foundation models beyond the one used during evolution. The authors released the code and a 2,000-task evolution dataset on GitHub.
Sources
Research this in Signals
Signals turns a topic into a sourced research record you can inspect and rerun. Your first scan is free, and this one starts with ModularRSI already loaded, so edit it or scan as is.