GitSwarm paper: multi-agent swarm compounds inference in a shared Git repo, solves all 30 IMOProofBench problems
NandoDF · x · 2026-10-11
A 14-author paper from Ruslan Salakhutdinov, Aaron Courville, Jason Weston and others introduces GitSwarm, a new "compounding inference" paradigm: intermediate work from inference-time compute persists and can be inspected, extended, combined, or challenged by later computation.
Key points:
- An asynchronous multi-agent system where agents independently explore, experiment, verify, refine, and synthesize, collaborating via structured persistent memory
- A shared, branchable Git repository serves as memory: atomic commits preserve intermediate artifacts, while explicit semantic dependencies track how contributions build on each other
- Results: solves all 30 IMOProofBench-Advanced problems in one run with GPT-5.5; scores 79.4% mean on ProgramBench vs. 65.1% for the strongest baseline under the same budget; also validated on three GPU-backed neural architecture research tasks
The core motivation: parallel attempts repeat work and long reasoning chains lose detail through compaction — GitSwarm makes long-horizon problem solving compound like interest.
More from coding & agent
- Codex agent battles Verizon's broken billing system in surreal hour-long browser session — rickasaurus · 2026-10-11
- Anatomy: one-line prompt turns Claude into an interactive 3D machine wireframe generator, fully open source — _AustinCalvert_ · 2026-10-11
- Claude Code fast mode reportedly blocks usage despite 97% of weekly credits remaining — RileyRalmuto · 2026-10-11
- REA: an agentic reverse-engineering toolkit for dissecting arbitrary hardware — jazir55 · 2026-10-11
- Cowcraft MCP and WoWBench go live, testing LLM agents inside World of Warcraft — djcows · 2026-10-11
- exe team's open-source web-based coding agent is standalone and useful — davidcrawshaw · 2026-10-11