Large Discovery Models: Empirically-grounded Model-Based Open-Ended Search
The paper proposes pairing a generative model with a Bayesian reward surrogate so discovery search is guided by measured performance and uncertainty.
Large Discovery Models keep a memory of experimental results and update the surrogate as new observations arrive. The setup is tested on neural-network training, antibody design, and molecular optimization. The authors report gains over LLM-only reflection and traditional statistical search, including better validation BPB reduction, lower binding energy, and stronger molecular multi-objective performance. HF Daily Papers' note
Large Discovery Models keep a memory of experimental results and update the surrogate as new observations arrive. The setup is tested on neural-network training, antibody design, and molecular optimization. The authors report gains over LLM-only reflection and traditional statistical search, including better validation BPB reduction, lower binding energy, and stronger molecular multi-objective performance. HF Daily Papers' note
score 5