UCL Introduces Large Discovery Models (LDM): Automating Experimental Design with AI
机器之心 · wechat · 2026-08-30
A team led by Wang Jun at UCL has released Large Discovery Models (LDM), designed to solve the problem of "how to design the next experiment" in scientific discovery. LDM integrates generative foundation models with Gaussian Process (GP) reward models, showing strong results in Auto-Research, antibody design, and molecular optimization.
Core Architecture: Fast and Slow Loops
- Fast Loop: The LLM generates open-ended candidates (code, sequences), while the GP evaluates value based on experimental feedback (exploration vs. exploitation), selecting the best experiment via acquisition functions.
- Slow Loop: High-value search strategies are distilled into model parameters, enabling the model to learn "how to propose new designs based on experience," a meta-skill transferable across tasks.
Key Concept: Discovery
- Traditional optimization focuses on "exploitation" (finding best in known space) and "exploration" (gathering info in known space).
- LDM emphasizes "discovery": when the current search space is saturated, the model alters its representation to include novel mechanisms or structures.
Experimental Results
- Auto-Research: Introduced new mechanisms like Muon optimizer; BPB reduction is 2.4x that of pure LLM reflection, reaching 0.902291 on B200 hardware.
- Antibody Design: Achieved an average 18.2% reduction in binding energy in a 20^11 sequence space.
- Molecular Design: Pareto hypervolume improved by 60% over pure LLM and Bayesian optimization.
Cross-Task Generalization
- A model fine-tuned only on code and molecule data matched or outperformed supervised models on 4 out of 5 unseen antibody tasks, proving the acquisition of general research decision strategies.
More from Research
- A beginner-friendly introduction to causal inference kicks off, Part 1 — Nasereliver · 2026-10-16
- FBI method trains for Pass@K with one sample per update at 12.3% of the cost — EmtiyazKhan · 2026-10-16
- Berkeley researchers sample humanoid robot policies directly from weight space with diffusion — berkeley_ai · 2026-10-16
- OneWM: unified causal training shows world and action modeling improve each other — IAAI233 · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- 5.8M-param Speech Event Detector adds laughs, breaths and背景 sounds to Whisper transcripts — Ayush Kumar Bar · 2026-10-16