Skip to main content
Back

Meta's GitSwarm Solves All 30 Olympiad Proofs by Giving AI Agents Shared Memory

A new multi-agent framework uses a shared Git repository as working memory, letting parallel LLM workers inspect, extend, and build on each other's attempts.

  • GitSwarm uses a shared Git repo as persistent memory for parallel LLM agents working long-horizon tasks.
  • Scores 79.4% on ProgramBench vs 65.1% for the strongest baseline under matched budget.
  • Solves all 30 IMOProofBench-Advanced olympiad problems in a single run with GPT-5.5.
  • 94.7% of worker contributions get built upon, and selected solutions inherit from 82-93% of the graph.
  • Homogeneous workers emerge into roles like construction, repair, verification without any explicit assignment.
  • Authored by researchers at Meta FAIR, Mila, CMU, Princeton, and Oxford including Weston, Salakhutdinov, and Arora.

Inference-time scaling, the compute spent after a model has been trained, commonly uses independent best-of-N samples or one extended chain-of-thought trajectory. Independent samples repeat dead ends because they cannot inspect their peers’ work, while long trajectories can lose details when context is compacted. GitSwarm stores intermediate work in a shared Git repository so later model workers can inspect, extend, test, or reject it.

Researchers from Meta FAIR, Mila, Carnegie Mellon, Princeton, and Oxford describe the system in the GitSwarm paper. The team includes Jason Weston, Ruslan Salakhutdinov, Sanjeev Arora, Aaron Courville, and Anirudh Goyal. Using GPT-5.5, the authors report solving all 30 problems in IMOProofBench-Advanced in one run. GitSwarm also records a 79.4% mean score on ProgramBench, compared with 65.1% for the strongest reported baseline under the stated compute budget.

How cumulative inference works

The authors call the approach compounding inference: intermediate results persist so subsequent workers can reuse evidence, challenge assumptions, combine branches, and repair earlier attempts. A failed experiment remains discoverable hours later, complete with its files, history, and relationship to other work.

A Git repository provides that institutional memory through familiar mechanisms. Commits preserve bounded changes, branches isolate competing approaches, and history records how the project developed. GitSwarm adds semantic dependency metadata so workers can credit useful work from branches outside a commit’s direct ancestry.

Git supplies the coordination layer

GitSwarm separates task selection from capacity management through three design choices:

  • Decentralized task selection: Workers inspect the repository and choose their own contribution. No central planner assigns subtasks.
  • Homogeneous workers: Every worker receives the same prompt and tools, without fixed roles such as planner, implementer, or critic.
  • Asynchronous execution: A scheduler maintains C concurrent workers and starts a replacement whenever one exits.

A worker turns the current repository state into one bounded contribution through a cycle that keeps provenance attached to every change: