GOLLuM Fixes LLM Overconfidence in Science via Uncertainty Training
bravo_abad · x · 2026-08-29
Addressing LLM overconfidence in science, GOLLuM trains a language model jointly with a Gaussian process to learn from experimental outcomes under uncertainty.
Key Results:
- Ranked first on average across 23 tasks in synthesis, materials, and molecular design.
- Achieves results comparable to traditional Bayesian optimization with 40% fewer experiments.
- Nearly doubles the discovery of high-yield reactions vs expert descriptors and frontier LLMs (43% vs 24–25%) starting from just ten failed experiments.
In contrast, prompting GPT-5, Gemini, and Claude to optimize directly showed failure rates up to 80% due to hallucinated molecules, duplicates, and invalid suggestions.
More from Research
- Class Weights 101: The One-Line Fix for Imbalanced Data That Beats SMOTE — dima806_dima · 2026-10-15
- One Diagram Maps the Gauge Freedom of Legendre Transforms from Convex Analysis to Information Geometry — FrnkNlsn · 2026-10-15
- Japan's Humanitext wins $816,667 Schmidt Sciences HAVI grant to build verified classical-text knowledge graph — SebastianNehrd2 · 2026-10-15
- StereoQueerEval joins SemEval 2027: spotting LGBTQIA+ stereotypes and hate speech in four languages — debora_nozza · 2026-10-15
- RLVR fine-tuning collapses solution space, replay buffer fix boosts diversity and pass@8 — manoelribeiro · 2026-10-15
- ModeBench shows RLVR collapses LLMs into single-solution mode; Re:Max training preserves diversity — manoelribeiro · 2026-10-15