LLMs can suggest scientific mechanisms, but this paper finds that letting the agent choose experiments and fit the mechanism is far less data-efficient.

So scientific agents may work better when the LLM proposes hypotheses but does not decide what the evidence means.

MDA turns the LLM into the hypothesis generator. Bayesian inference scores the candidate mechanisms, and value-of-information chooses the next experiment where those mechanisms disagree most.

That changes the experiment budget dramatically.

On FORCEBENCH, MDA reaches roughly the accuracy of an unthrottled Opus 4.7 agent using 8 experiments instead of about 41, while reaching a 93% numeric pass rate versus 31% for the budget-matched Opus 4.7 LLM agent.

The mechanism is easy to see in the examples: for Yukawa forces, it probes long range because the competing laws look identical nearby; for Coulomb, it changes source charge because moving the probe alone cannot separate the true law from a charge-blind fit.

When predictions still fail, MDA asks the LLM for new mechanisms and repeats the loop.

Let LLMs propose scientific ideas, but let explicit uncertainty and designed experiments decide what survives.