Does the Prior Pay Off? Scaling a Protein Language Model Lets Reinforcement Learning Beat Directed Evolution on GB1
A deliberately fair, matched-query-budget comparison of GRPO reinforcement-learning fine-tuning of a protein language model (ESM-2) against classical directed evolution on the GB1 four-site fitness landscape, using an exact-lookup oracle (nothing to game), a strong simulated-annealing baseline, a novelty floor, and five seeds. The protein-LM prior is a poor fitness predictor (masked-marginal Spearman 0.04 at 35M, 0.15 at 150M; its top-ranked variants have near-zero fitness), so a no-RL masked-marginal proposer is weak. Simulated annealing nearly solves the landscape, and tuned GRPO over a 35M prior only matches it, missing a pre-registered 10% bar. Scaling the prior to 150M changes the outcome: GRPO then beats annealing by about 10% in the sample-limited regime where queries are scarce (significant at a 1k-query budget; the gap narrows to a tie as both methods saturate). The gain is decoupled from the prior's fitness-ranking quality, which stays poor, indicating the larger model helps as an initialization and inductive bias rather than as a fitness oracle. GRPO also collapses without discovery-shaped reward, an entropy bonus, and a low learning rate.