Hypothesis-driven multi-objective optimization of the Belousov–Zhabotinsky reaction

Self-driving laboratories can optimize experiments, but whether they can reason scientifically about what to test remains unclear. Bayesian optimization (BO) provides uncertainty-aware search without explicit scientific interpretation, whereas language-model agents can formulate hypotheses but lack calibrated uncertainty. Here we introduce maLLaMar , which combines these capabilities in a physical self-driving laboratory. Using the Belousov–Zhabotinsky reaction, we compare BO-only, agent-led and cooperative experiment selection under matched experimental budgets. Agent-led selection raised the fraction of hour-long oscillations from 24% to 64%, while cooperation yielded the largest number of near-optimal conditions, eleven compared with two for either strategy alone. The best conditions sustained oscillations for the full hour with mean periods up to 4.7 times shorter than the highest-scoring literature formulation re-measured on our platform. Hypotheses recorded prospectively during experimentation were evaluated against an exhaustive 294-condition reference map that was never available to the agent. The map supported or refined several predictions and falsified generalizations drawn from sparsely sampled regions. Statistical optimization and explicit scientific reasoning thus contribute complementary capabilities, allowing an autonomous laboratory both to find improved conditions and to generate claims that independent experiments can test.