Frontiers of LLMs for
Automated Scientific Discovery:Scaling Training, Hypothesis Coverage,
and Closed-Loop Experimentation

Overview

Large language models can now propose scientific hypotheses, write candidate equations, and design materials. Turning these capabilities into reliable scientific discovery requires more than generating a plausible answer. This tutorial examines how to train discovery models, evaluate the breadth of their hypothesis search, and connect their proposals to experimental evidence.

Building on our AAAI 2026 tutorial, we bring together methods and benchmarks across chemistry, physics, materials science, and genetics. Frameworks such as MOOSE-Chem, LLM-SR, and HypoSpace illustrate where LLM-driven discovery works, where it fails, and which tools to use in each setting.

Motivation

Prompting an LLM to propose a hypothesis is only a starting point. Three questions organize this tutorial:

1. Scaling training. How can we train discovery models rather than only prompt them, when hypothesis generation involves a combinatorially large search space?
2. Hypothesis coverage. When many explanations fit the same observations, does a model explore the admissible hypothesis space—or repeatedly return a few familiar answers?
3. Closed-loop experimentation. How should a system choose what data to collect next, verify candidate discoveries, and update its beliefs over repeated experiments?

Together, these questions shift the focus from individual generations to scientific systems that learn, search broadly, and improve through evidence.

Tutorial Outline

1. Scientific discovery as a closed loop

   • Problem specification, context retrieval, hypothesis generation, experiment design, and evaluation.
   • Why benchmarks shape the field: memorization, contamination, and the gap between fitting data and recovering scientific structure.

2. Discovery with verification

   • LLM-guided equation and materials search, from tree search and symbolic–numeric pretraining to evolutionary search with numerical fitting.
   • Scientific equations as programs: LLM-SR and its demonstration; memorization-resistant evaluation with LLM-SRBench and SurfaceBench.
   • Adaptive search, diversity, and hypothesis-conditioned experiment selection, illustrated by LLEMA, DecAEvolve, EvoDiverse, and LLM-AutoSciLab.

3. Hypothesis generation and refinement

   • Decomposing scientific hypothesis discovery into inspiration retrieval, composition, and ranking: MOOSE, MOOSE-Chem, and ResearchBench.
   • Fine-grained discovery through hierarchical search with MOOSE-Chem2, and ranking with simulated experimental feedback.
   • Human-in-the-loop exploration and refinement with a MOOSE-Copilot demonstration.

4. Scaling training for discovery

   • The complexity barrier to direct training, and how factorization, bounded composition, and hierarchical retrieval make training tractable.
   • MOOSE-Star: scaling with data and inference budgets, and understanding the limits of brute-force sampling.

5. Hypothesis coverage and set-valued evaluation

   • HypoSpace: enumerated admissible hypothesis sets and deterministic validators.
   • Separating validity, uniqueness, and recovery; diagnosing mode collapse even when individual outputs are valid.
   • Why more sampling alone may not restore coverage, and which mitigations help.

6. Long-horizon discovery and research agents

   • Iterative structure and equation recovery with Auto-Discovery-Bench; belief tracking across repeated hypothesis–experiment–update cycles.
   • Process-level evaluation of autonomous research agents with FML-bench, and the role of exploration breadth.

7. Synthesis and open problems

   • Coverage under peaked generators, verification without simulators, and credit assignment over long research trajectories.
   • Benchmarks that renew with model cutoffs, cross-domain lessons, and discussion.

Demonstrations and a short hands-on segment introduce the core hypothesis-generation and equation-discovery loops.

Target Audience

AI researchers, graduate students, practitioners, and domain scientists who want to build, train, or evaluate LLM-based systems for scientific discovery. The tutorial is accessible to participants with a general AI background; scientific case studies are introduced from scratch.

Prerequisite Knowledge

Basic familiarity with LLMs and their use as agents, at the level of having used an LLM API and read a paper on prompting or tool use. The equation-discovery material assumes undergraduate familiarity with regression and optimization. No specialist knowledge of chemistry, physics, materials science, or genetics is required.

Learning Outcomes

Participants will gain:

• A clear map of the methods and benchmarks used in LLM-driven scientific discovery.
• An understanding of hypothesis generation as retrieval, composition, and ranking, and how this decomposition enables search and training.
• Practical insight into equation and materials discovery with numerical verification.
• Tools for distinguishing valid outputs from broad hypothesis-space coverage.
• An understanding of experiment selection, iterative belief updating, and the challenges of long-horizon scientific agents.