Agentic AI Scientists Are Not Built For Autonomous Scientific Discovery | arXiv cs.AI

Agentic AI Scientists Are Not Built For Autonomous Scientific Discovery | arXiv cs.AI

The authors argue that agentic AI scientists can already function as co-scientists, but are not built for autonomous scientific discovery. They identify four challenges. Problem selection is influenced by the McNamara fallacy; large language models (LLMs) omit tacit procedural and failure knowledge from laboratory practice in their training corpora. Preference optimisation during post-training compresses output diversity toward consensus. Most scientific benchmarks measure single-turn prediction accuracy without feedback from physical experiments to computational models. In the authors’ view, these are not merely problems of scale or scaffolding, but call for revisiting fundamental design choices. They recommend scientific simulations as verifiers for training and a persistent, mutable epistemic state that carries beliefs and shifting objectives across an investigation. They also recommend a centralized preregistration repository for AI-generated hypotheses, and applying AI scientists in response to scientific need rather than tool affordance. The abstract presents these as design recommendations; it does not spell out how to implement or evaluate them.

Publication: b19f4942-288b-4958-8923-7f313c97b0a4

attribution — pass: The reading consistently presents the position as the authors' argument ("The authors argue", "In the authors' view"). It does not present those claims as established fact. The source is correctly identified as an arXiv cs.AI position paper by Bisht, Kumar, Jablonka, Mausam and Krishnan.

This position paper argues that although they already function as co-scientists, agentic AI scientists are not built for autonomous scientific discovery.

summary — uncertain: The reading tracks the abstract faithfully, but it leaves central jargon unexplained for a non-specialist: McNamara fallacy, agentic, preference optimisation/post-training, scaffolding, epistemic state, preregistration and tool affordance. A reader cannot grasp the first challenge without knowing that the McNamara fallacy means favoring what is easily measured. Smallest repair: add tooltips or one-clause glosses for these terms, plus one sentence on why autonomy versus co-scientist matters. Restoring the abstract's condition 'To build truly autonomous AI scientists' would also help. The NeurIPS 2026 position-track framing is omitted, though that is minor.

Problem selection is influenced by the McNamara fallacy

date — pass: originalPublished is 2026-05-09, which matches the v1 submission date. It does not match the later v3 revision (6 Nov 2026) or the feed date (9 Nov 2026).

[Submitted on 9 May 2026 (v1), last revised 6 Nov 2026 (this version, v3)]

support — pass: Each claim is source-stated and backed by its excerpt. The note that the abstract does not explain implementation or evaluation is an accurate limitation, not an invented claim. The question raises a genuine unresolved issue that follows from the recommendations. No unsupported benefits or costs are implied.

To build truly autonomous AI scientists, we recommend scientific simulations as verifiers for training

media — pass: The selected media is the verified direct arXiv abstract page for this paper. That is the most direct way for readers to reach the source.