Paper proposes a confidence-calibrated three-action controller for evaluator-guided test-time reasoning
The paper (arXiv:2610.20836v1) formulates evaluator-guided reasoning as a compute-aware three-action decision problem (trust, fallback, or expand) and trains a controller that scores terminal actions by predicted success and expansion by a rollout-based continuation-success surrogate while explicitly pricing compute. The method learns action-specific calibrated confidences and local uncertainty penalties; on held-out width expansion and revision tasks it reduces normalized online compute by about 30% and 16% respectively versus calibration-selected fixed policies, with mixed but small utility differences across four Qwen3-8B seeds.