ResearchTrails: Mining Git Histories Into Human Research Decision Trajectories to Train Better Research LLMs
ChenhaoTan · x · 2026-10-08
xuchengong released ResearchTrails, a dataset of human research decision trajectories extracted from Git commit histories.
- Final papers show outcomes but miss the full path — abandoned methods, omitted experiments, and ablations. ResearchTrails recovers these intermediate decisions.
- The team shows that learning from these trajectories improves LLMs' ability to assist with research exploration.
- Blog, code, data, and a trajectory visualizer are all available.
More from Research
- Open Env Arena launches: agents compete to design the best RL environments — ben_burtenshaw · 2026-10-08
- Quasi-Riemann hypothesis proof sparks reading wave, with Claude used to summarize the paper — aran_nayebi · 2026-10-08
- 8-page paper makes first double-log improvement to Erdos-Rankin prime gap bound in a century — aran_nayebi · 2026-10-08
- Fields Medalist Gowers pushes back on the claim that AI 'merely explores the convex hull' — burny_tech · 2026-10-08
- Getting an LLM judge from garbage to 27/30 human agreement: an evals workshop writeup — tejaskumarlol · 2026-10-08
- Mila and Inria sign four-year partnership spanning agentic AI, robotics and health — Mila_Quebec · 2026-10-08