Research map / Large language models / AI and society
Agentic AI Scientists Are Not Built For Autonomous Scientific Discovery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Position paper arguing agentic AI scientists are not built for autonomous discovery, identifying four challenges (McNamara fallacy in problem selection, tacit/failure knowledge gaps, diversity compression from preference optimisation, invalid benchmarks) and supporting diversity claims with a 12-model "hypothesis hivemind" embedding experiment.
Problem
Whether LLM-based agentic AI scientists can perform end-to-end autonomous scientific discovery in the natural sciences, given limits in problem selection, training data, output diversity, and evaluation.
Approach
Argues the obstacles are inherent to training and deployment design rather than scale or scaffolding, and proposes simulators as verifiers, a persistent mutable epistemic state, a centralized preregistration repository for AI-generated hypotheses, and demand-pull application.
Key results
- Cross-provider similarity of novel hypotheses on AI4Mat was 0.64 vs 0.72 intra-model across 12 models.
- Drawing from 10 models instead of one raised effective number of distinct hypotheses (Vendi) only from 2.9 to 3.6 on AI4Mat.
- On PLOS Biology, cross-provider similarity was 0.69 vs 0.76 intra-model for novel hypotheses.
Surprising findings
- Outputs from different providers are nearly as similar to each other as repeated samples from one model.
- Querying ten different models yields only marginally more effective diversity than one model's samples, far less than hypotheses about different papers.
Limitations
- Hivemind experiment uses a single embedding model and prompt setup, with embedding similarity as a proxy for semantic diversity.
- Argument is largely position-based, with the SSE running example being illustrative rather than empirically tested.
- Excluded 5 PLOS papers due to provider content filters, potentially biasing the sample.
Experimental setup
Datasets: 50 NeurIPS 2025 AI4Mat track papers 52 PLOS Biology papers (April-June 2026)
Benchmarks: hypothesis hivemind (AI4Mat and PLOS Biology paper sets)
Metrics: mean cosine similarity of embeddings Vendi score
Builds on: Hivemind hypothesis Vendi score dual-search framework
Closest papers by meaning
- AI-Researcher: Autonomous Scientific Innovation (NeurIPS 2025 · similarity 0.91)
- DiscoveryWorld: A Virtual Environment for Developing and Evaluating Automated Scientific Discovery Agents (NeurIPS 2024 · similarity 0.88)
- Agent-Native Research Artifacts (NeurIPS 2026 · similarity 0.88)
- AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research Suite (ICLR 2026 · similarity 0.88)
- InnovatorBench: Evaluating Agents’ Ability to Conduct Innovative AI Research (ICLR 2026 · similarity 0.88)
- From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking (ICLR 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.