Value-Conditioned Proposal Acquisition for Test-Time Reasoning
| Source: arXiv AI
Tags: ValueFormation, test-time compute, MATH-500, GSM8K, Qwen3-14B, Llama
ValueFormation decides at test time whether to stop, keep sampling, or call a reasoning specialist, beating eight baselines on MATH-500 by over 15 points for three 8B models.
Details
Sampling and selection can't recover a correct answer that never enters the candidate pool. ValueFormation is a test-time policy that, after each draw from the primary model, decides whether to stop, sample again, or acquire one proposal from a different reasoning specialist. The three choices are compared on one cost-regularized value scale learned from vote-state features. The calibration is frozen across tasks and primary models, and acquired proposals are accepted only through a frozen posterior comparison. On held-out MATH-500, one Llama-fitted calibration beats the strongest of eight baselines by more than 15 points for each of three 8B primaries and transfers unchanged to Qwen3-14B. Removing acquisition cuts Llama MATH accuracy by 20.43 points. On GSM8K, gains come mainly from cheaper allocation, and raising the token price reduces specialist calls without retraining.