Can model confidence work as a reward signal for reasoning?

Reinforcement Learning from Self-Feedback (RLSF) exploits a simple observation: in a well-calibrated model, answer confidence correlates with reasoning quality. By using confidence as the reward signal rather than human preference or external verification, RLSF achieves two things simultaneously that normally trade off:

(i) Restores calibration — confidence becomes predictive of correctness again, after RLHF had degraded it. RLHF optimizes for human preference and fluency, which rewards confident-sounding outputs regardless of accuracy. RLSF reverses this by making the reward explicitly tied to calibrated confidence.

(ii) Strengthens step-by-step reasoning — higher-confidence answer spans tend to come from traces with more coherent reasoning chains. Training to maximize confidence indirectly selects for better reasoning.

The mechanism: a frozen LLM generates multiple CoT solutions for each problem. Confidence is computed per final-answer span. Traces are ranked by this confidence to create a synthetic preference dataset (higher confidence = chosen, lower = rejected). A reward model is trained on these preferences and used for standard RL finetuning.

The key insight is that confidence-as-reward can be inserted as an additional post-training step after standard SFT and RLHF — patching the calibration damage that RLHF introduces without undoing its alignment benefits. This requires no human labels, gold answers, or externally curated rewards.

The human learning parallel is explicit: humans use confidence as an intrinsic reward signal when external feedback is unavailable. Metacognitive monitoring — the ability to track your own certainty — is how humans regulate their own learning without a teacher.

The connection to Does binary reward training hurt model calibration? is complementary: that work adds calibration as an explicit second reward term; RLSF uses calibration itself as the primary reward. Both address the same RLHF-induced calibration degradation from different angles.

The risk is the same as Does self-consistency reliably reward correct answers during training? — confidence and self-consistency are correlated proxies, both vulnerable to the model becoming confidently wrong. But RLSF's emphasis on calibration (making confidence track accuracy) is explicitly designed to resist this — the model is rewarded for being accurately confident, not just confident.

Extensions to general domains via RLPR and INTUITOR: Two RLVR papers extend intrinsic reward signals beyond math to general domains. RLPR (RL from LLM Intrinsic Probability) computes the model's token-level probability of generating a reference answer, using this as reward signal — the model's own knowledge about what constitutes a correct answer replaces external verifiers. INTUITOR goes further: it uses self-certainty as the sole reward signal, computed as the confidence gap between the model's top-choice answer and alternatives. Both extend verifiable-reward RL to domains without rule-based verifiers (medicine, law, open-ended reasoning) — precisely the domains where external verification infrastructure is hardest to build. The convergence with RLSF is notable: all three use the model's internal probability landscape as reward, but RLSF targets calibration restoration, RLPR targets domain extension, and INTUITOR targets complete verifier independence. See Can model confidence alone replace external answer verification?.