Proposal of a method where multiple AI agents stack work in a Git repository to solve problems
- What
- Proposed and evaluated GitSwarm, a 'compounding inference' system where multiple identical AI agents work independently in a shared Git repository, continuously saving, referencing, and combining intermediate results.
- Key result
- Reported solving all 30 math proof problems in a single run, outperforming the strongest baseline under a fixed budget (65.1 points) with an average of 79.4 points on the programming benchmark ProgramBench, and improving starting architectures through iterative experiments across three neural network structure tasks.
- Why it matters
- Demonstrates that AI can accumulate intermediate achievements rather than discarding results after each attempt, reusing them for subsequent tasks, thereby suggesting a direction for long-horizon problem solving and AI design for scientific research. However, since these are results at the abstract level, actual industrial impact is difficult to judge from this material alone.
Background
Complex problems and scientific research require computations and knowledge to accumulate through multiple attempts. However, most AI inference proceeds in units of individual attempts or candidate answers, making it difficult for partial solutions or failed approaches to be utilized in subsequent work. This paper defines the method where intermediate work products persist and are reused as 'compounding inference' to solve this problem.
Method
GitSwarm is a structure where agents with identical performance make independent decisions and proceed with work without waiting for each other. Agents explore, experiment, verify, modify, and synthesize across multiple branches in a shared Git repository, preserving intermediate results as small-unit commits. They also record explicit dependency relationships indicating which previous work subsequent tasks were based on. This method was applied to math proofs, programming tasks, and neural network experimentation research using GPUs.
Main results
Reported solving all 30 problems in a single run using GPT-5.5 on IMOProofBench-Advanced. Recorded an average of 79.4% on ProgramBench, higher than the 65.1% baseline. Additionally, 94.7% of contributions in this task were reused in subsequent work, and the ancestry of the finally selected solution accounted for 82-93% of the entire contribution graph. In three neural network structure tasks, it was stated that through multiple experiments, the performance improved over the starting architecture.
Limitations & open points
The abstract does not provide detailed information such as the detailed conditions of the comparison baselines, the cost and scale of computation spent on experiments, specific figures showing the magnitude of improvement in neural network tasks, and reproducibility. Therefore, whether these results will appear at the same level in other tasks or real service environments cannot be confirmed from the abstract alone.
Glossary
- inference — The process where a trained AI model solves actual problems and produces answers.
- agent — An AI program that sets plans on its own toward a goal and performs tasks using tools or data.
- asynchronous — A method where multiple tasks proceed individually without waiting for each other to complete.
- Git repository and branch — Git is a tool for managing change history of software files, a repository is its storage, and a branch is a separate working path split from the original.
- commit — The act of saving file changes as a single unit of record.
- baseline — An existing representative method used to compare the performance of a new method.
- ancestry and contribution graph — A family tree-like structure connecting which previous works a result was created through.
This summary was written by AI from the paper's abstract and reviewed by the editor. Summaries on this page were translated by AI from the editor-reviewed Korean version. Check the original for exact figures and context.
This page compiles published papers, official announcements and press reports for information only. Paper summaries are written by AI from the abstract and reviewed by the editor; they may differ from the original, so check the paper for anything important. Nothing here is a recommendation to buy or sell any security or investment advice. Career moves cite official announcements or major press reports; some career records come from Wikidata (CC0). Summaries on this page were translated by AI from the editor-reviewed Korean version.
