Verifier Noise Is a Temperature: Testing Tilt Predictions in Finite-Budget LLM-Guided Search
Larger scoring minibatches improved search in six of seven settings, but the theory failed to predict how much.
After 150 steps on bin packing, search achieved 62–88% of the predicted gain when the LLM edited existing programs and 109–119% when it wrote new heuristics. Runs starting from good programs kept losing quality over 300 steps, and different starting points had not converged to a common level. That leaves open whether the proposed “temperature” model describes long-run LLM editing. A separate prompt-optimisation test also exceeded predictions, and the relationship found in bin packing did not transfer.
ArXiv · AI/CL/LG's note
After 150 steps on bin packing, search achieved 62–88% of the predicted gain when the LLM edited existing programs and 109–119% when it wrote new heuristics. Runs starting from good programs kept losing quality over 300 steps, and different starting points had not converged to a common level. That leaves open whether the proposed “temperature” model describes long-run LLM editing. A separate prompt-optimisation test also exceeded predictions, and the relationship found in bin packing did not transfer.
ArXiv · AI/CL/LG's note
score 4