Robert Lange

Sakana AI Verified account @SakanaAILabs · Oct 15

The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery

Blog: pub.sakana.ai/agentic-scient…
GitHub: github.com/SakanaAI/AI-Sc…
arXiv: arxiv.org/abs/2610.18933

Since 2024, we’ve used The AI Scientist as a way to test how far AI alone can take scientific research, from v1 to v2 to the paper published in Nature earlier this year. The Agentic Scientist is the next step in that line of work, built for the era of coding agents.

Coding agents such as Codex and Claude Code now carry out long software tasks on their own. Handing them a research problem, however, tends to produce a strong hyperparameter optimizer rather than a scientist: scores go up, but ideas stay narrow, and the reasons for the gain are rarely examined.

The Agentic Scientist makes the research loop explicit. It wraps off-the-shelf coding agents in a scaffold that controls what they see, what they write, how they are evaluated, and how results are handed to the next stage.

Parallel workers test hypotheses against rubrics defined before any experiment runs, share successes and failures through a common memory, and then ablate the winning program to generate evidence.

We used an early version of the system to study Error Diffusion, a biologically motivated alternative to backpropagation. Our researchers set the question and the rubric, chose directions at each stage, and verified the results, references, and claims before submission. The resulting paper was accepted as an oral presentation at ALIFE 2026. In autonomous runs, the system also raised Doc-to-LoRA retrieval accuracy from 52% to 98% on multi-needle tasks and designed a learned evolution strategy that improves over its CMA-ES reference on unseen tasks.

Alongside it we release AutoIllustrate, a pipeline for scientific figures that keeps an editable draw.io file as the source of truth, so every diagram can be corrected by hand.

Humans define the question and the standards of evidence; agents expand the experiments researchers can run. We call this Co-Discovery, and we are releasing both systems as open source for the community to build on.

View quoted post on X

Robert Lange Verified account @RobertTLange · 9h

Fastino's GLiDE at a glance 👇 a recurrent depth decision model:

– Qwen3.8-27B's 3:1 Gated DeltaNet/attention hybrid, attention made bidirectional
– LM head + MTP dropped (−1.7B) → 255×5120 readout over option codes, 1 pass per question
– "Thinking" = recurrent depth: layers 24–47 run twice (½x + ½G(x)), 64 → 88 layer passes, 0 new params, 0 output tokens
– 40K tokens/question, ≤255 options, text/JSON/images
– Decision Index 0.3.1: 72.49 (27B) … 51.10 (0.4B encoder)
– 112 ms median on 1× RTX PRO 6000

Blog: fastino.ai/blog/introduci…
Weights: huggingface.co/collections/fa…

x.com/george_onx/sta…

View quoted post on X

George Maloney Verified account @george_onx · 4d

Today we’re releasing the GLiDE family: five open-weight decision models, from 0.4B to 27B.

Together, they set a new Pareto frontier for decision models.

Most decisions inside an app don’t need a model that thinks out loud. They need a fast, clear, and calibrated answer. That's why we built GLiDE.

The part I’m most excited about is the recurrent depth in 27B and 2B models. Selected middle layers run twice, giving the model more computation before it answers without adding parameters.

On the Hugging Face Decision Index, GLiDE 27B is expected to score 73.3 vs. KnowLine 31B's 71.86, a 2% higher score with approximately ~11% fewer served params.

All five models use the System One schema. Test the same tasks across sizes, evaluate them on your own workload, and run the smallest model that meets your needs.

View quoted post on X