UCL's Large Discovery Models hit SOTA, 2.4x LLM reflection on experiment search
jiqizhixin · x · 2026-09-08
Jun Wang's group at UCL presents Large Discovery Models (LDM), turning model-based open-ended search into an empirical loop: given experimental data and a foundation-model context, it builds a Bayesian reward signal over the search space, scoring each candidate by performance or uncertainty reduction. A fast loop iterates in real time from experimental data; a slow loop distills the reward signal into the foundation model via post-training. On H100, LDM achieves 2.4x the BPB reduction of pure LLM reflection under identical conditions; on B200 it actively widens the search space, driving BPB down to 0.902291 — world #1 on the Auto-R leaderboard.
More from Research
- Bio ML critique: many published models score within ~10% of a naive linear baseline — iskander · 2026-09-08
- Pangram's trick: repeating text twice so LLMs can attend backward — diffusion tech creeping in — Birchlabs · 2026-09-08
- Sander Dieleman on Diffusion Models, Typicality, and Why ML Breakthroughs Are Closer Than They Look — sedielem · 2026-09-08
- Critical review of agentic AI offers framework for how much authority to delegate — dair_ai · 2026-09-08
- AI evals need higher effective rank: researchers push 'continuous benchmarks' for robustness — ajratner · 2026-09-08
- LlamaIndex CEO: The Fix for Benchmaxxing Is More Benchmarks, Not Fewer — ajratner · 2026-09-08