GitSwarm paper: agents collaborate via shared Git repo, GPT-5.5 solves all IMOProofBench
anirudhg9119 · x · 2026-10-10
A new arXiv paper from a heavyweight team (Vedant Shah, Anirudh Goyal, Jason Weston, Aaron Courville, Rob Fergus, Ruslan Salakhutdinov, etc.) proposes "compounding inference": organizing inference-time compute so intermediate work persists and can be extended by later computation. GitSwarm instantiates this with asynchronous homogeneous agents collaborating through a branchable shared Git repository, with atomic commits preserving artifacts and explicit semantic dependencies tracking how work builds across branches. Results: on IMOProofBench-Advanced it solved all 30 problems in one run with GPT-5.5; on ProgramBench it scored 79.4% mean vs 65.1% for the strongest baseline under the same budget, plus sustained GPU-backed neural architecture research. The authors frame "swarm scaling" as the next inference scaling paradigm.
Related event: GitSwarm: Decentralized Compounding Inference via Shared Git Repositories(11 posts)→
More from coding & agent
- Creator finds hand-tweaking generative models faster than prompts, sees room beyond text UIs — keenanisalive · 2026-10-10
- Telling an LLM to "believe in yourself" helps it write 3D SDF models, but not enough — keenanisalive · 2026-10-10
- "Do better!" prompting stalls fast; even top VLMs understand images unevenly — keenanisalive · 2026-10-10
- Full prompt revealed: making an LLM build procedural SDF 3D models from one image — keenanisalive · 2026-10-10
- LLM took 45 minutes to model a dragon; a diffusion model did far better in 3 — keenanisalive · 2026-10-10
- Reconstructing 3D from 2D is ill-posed; Nano Banana made the reference views — keenanisalive · 2026-10-10