Agora: Git as Shared Memory for Collective AutoResearch
Yifan Zhang, Yi Dong and colleagues at NVIDIA present Agora, a shared memory for autonomous research agents in which every result, hypothesis and verification is an immutable Git commit in an append-only DAG, and report a 12-day run with 13 LLM workers and no central planner.
Ask this paper
Research as commits: Each claim is a commit whose parent edges record what it builds on; a derived index lists the frontier, neglected branches and verification status, and a diversity-aware selection rule keeps workers from all following one leader.
Weight-transfer task: Workers had to initialize a frozen 119.6M-parameter attention-SSM hybrid from 141 donor models without training data or gradient updates.
Result: The community posted 1,703 contributions and reduced the evaluator from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M.
Reproducibility: The winning recipe has a 145-commit ancestry across 15 accounts, and all 165 independent reproductions posted during the run succeeded.
Stated limits: One human intervention was needed to break a monoculture, and the authors describe the controlled experiment still needed to show that shared state improves discovery per unit of compute.
Abstract
Autonomous research loops such as AutoResearch show that one coding agent can improve a training setup unattended. Run several of them and each session starts from scratch, so more agents tend to mean more duplicated search rather than more discovery. Agora is a shared memory for such agents: research is recorded as an append-only directed acyclic graph (DAG) stored in Git, so that every claim is a commit anyone can check out and rerun. Each result, insight, hypothesis, verification, and report is an immutable commit whose parent edges say what it builds on; a derived index exposes the frontier, the neglected branches, and the verification status of each claim, and a diversity-aware selection rule keeps the community from collapsing onto one leader. We describe the system and report its first sustained use: a run of nearly 12 days in which 13 language-model workers, with no assigned tasks and no central planner, worked on a weight-transfer problem. Given 141 pretrained donor models and a frozen 119.6M-parameter attention-SSM hybrid whose dimensions match no donor, the workers had to initialize the target without training data or gradient updates. They published 1,703 contributions and drove the evaluator from 3.39 to 1.899 bits per byte, closing 62% of the gap to a trained GPT-2 124M. The winning recipe compresses donor next-token statistics into the target's embedding and output head, then adds a short-range context signal through sparse edits to attention, feed-forward, and state-space blocks. Its 145-commit ancestry spans 15 accounts, and 165 independent reproductions were posted, none of which failed. We describe the single mid-run human intervention that pulled the community out of a monoculture, what the trace does and does not establish, and the controlled comparison that would settle whether shared research state improves discovery per unit of compute.