Why it mattersResearch groups that let one model both propose and review its own experiments now have an open pipeline that separates the two roles, persists every run to disk, and preserves negative results instead of forcing every experiment into a success.
EvoMap published AutoResearch on GitHub, an open-source agent workflow that runs the full loop from an ML research idea to a package of experiment code, logs, metrics, critic reports and blind reviews. The repository was created on 15 August 2026, was last pushed on 1 September, and has reached 1,362 stars and 91 forks. It is Apache-2.0 licensed. A named team of six researchers is on the byline, including project leaders Haoyang Zhang and Junjie Wang from the Infinite Evolution Lab at EvoMap, and the README links to an arXiv preprint at 2608.17906.
What it runs
The pipeline takes one of three entry points. Give it a research idea and it plans the experiment, writes code, executes runs, analyses results, and produces an independent review. Ask it to discover directions and it collects recent papers, developer discussions and open-source trends, filters them, and generates candidate ideas with experiment plans. Ask for the full workflow and it does both. Every stage writes to disk: research plans, code, run logs, metrics, failure causes, critic reports and blind reviews all land in named directories under data/, so a run can be inspected, taken over by a human, or resumed after an interruption.
Three models, not one
The core design rule is that idea generation and critic stages require at least three distinct model identities, counted by underlying model rather than by endpoint. The README states it as a hard independence requirement: the preflight script fails if the configured providers do not offer enough distinct models for the multi-model stages. This is the mechanism against one model both proposing an idea and reviewing it favourably. The Idea Forge stage is where the three-way check runs, and the model roster is declared in config/providers.local.json rather than hard-coded to Gemini, GPT or Claude.
Where negative results live
The capability table calls out one design choice worth naming. AutoResearch preserves evidence and stops when a hypothesis fails, rather than continuing to search for a framing that turns the run into a success. Failed runs stay in data/experiments/ with their logs and metrics, and the blind-review outputs treat a negative result as a finished record rather than a task to fix. Pilot runs test feasibility at lower cost before a full experiment starts, so a plan can be dropped before it burns a large model budget.
For a team running ML experiments with a coding agent, the workflow makes two things concrete that are usually informal. First, the separation of proposer and reviewer becomes a preflight check that has to pass. Second, the record on disk is the primary artefact, so a later reader can walk back from a claim to the run that produced it. The trade-off is a Linux machine with Python 3.10 or newer, multiple provider credentials, and a real API budget for the multi-model stages. The value is that idea generation and experiment execution become inspectable rather than embedded in a single conversation.
Source
EvoMap/AutoResearch on GitHub and the arXiv preprint at 2608.17906
Source: EvoMap on GitHub
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually shipping with them. Short, and only when there is something worth reading.
