Hugging Face Daily PapersYifan Zhang, Yunheng Zou, Shaokun Zhang2 min readpaperadvanced
Agora: Git as Shared Memory for Collective AutoResearch
Summary
Agora treats a Git repository as a shared, append‑only memory for autonomous research agents, recording each claim as an immutable commit in a DAG. In a 12‑day run with 13 language‑model workers it generated 1,703 reproducible contributions and closed 62 % of the gap to a trained GPT‑2‑124M, demonstrating that shared research state can accelerate discovery.
- Agora stores every research artifact as an immutable Git commit, forming a directed‑acyclic graph that guarantees reproducible provenance.
- A derived index over the commit graph highlights frontiers, neglected branches, and verification status, enabling diversity‑aware agent selection.
- In a 12‑day experiment, 13 language‑model workers produced 1,703 contributions, improving bits‑per‑byte from 3.39 to 1.899 (62 % of the GPT‑2‑124M gap) without a central planner.
- The top solution compressed donor model statistics into the target’s embeddings and added sparse edits to attention, feed‑forward, and SSM blocks, achieving weight‑transfer without data or gradients.
Teams building autonomous ML agents need a reproducible, shared knowledge base to avoid duplicated work; Agora shows a practical Git‑based approach that yields measurable gains.
8/10
