Sakana's Agentic Scientist co-authors an ALIFE 2026 paper

A paper co-written by an orchestration of coding agents has been accepted as an oral presentation at ALIFE 2026. Sakana AI's writeup names the result, a study of backpropagation-free Error Diffusion learning rules, as the output of its new Agentic Scientist pipeline. The authors stress this is a co-scientist result, not an autonomous one: humans chose the question, reviewed the evidence, and verified every result before submission.

The system is on GitHub under Apache 2.0, with an 88-page arXiv paper (2610.18933) submitted October 13, 2026. A meta-controller dispatches parallel coding-agent workers through six phases: task drafting, investigation, ablations, final evaluation, post-hoc analysis, and paper generation. The team ran it across nine frontier LLMs, four GPT, four Claude, one DeepSeek, and four agentic harnesses: Codex, Claude Code, OpenCode, and Pi.

The flagship numbers come from case studies. On the Doc-to-LoRA retrieval task, the pipeline raised exact-match accuracy on synthetic multi-needle retrieval from 52% to 98%. The Error Diffusion paper it co-produced reports 96.7% accuracy on MNIST and 61.7% on CIFAR-10. Across a wider sweep of 70 research campaigns, reviewer scores correlated with numerical performance at r = 0.44.

The authors are explicit about where this breaks.

"A high score can still come from a shortcut that violates the intended method...especially when important constraints are left implicit," they write. The system can clear a rubric by solving a problem different from the one specified. They also flag limited literature integration during ideation and the absence of strong baseline comparisons from published methods.