Variance-aware Reward Modeling with Anchor Guidance

Pairwise human comparisons are a primary interface through which modern AI systems learn human preferences. RLHF and related alignment pipelines typically model such comparisons with Bradley--Terry log-odds, where choice probabilities are governed by latent reward differences. This paper examines what this assumption misses through a reduced-form model motivated by rational inattention, in which each label is generated by a low-capacity evaluation channel. The model separates two forms of ambiguity that standard reward modeling tends to conflate: a comparison may be difficult because the two candidates are genuinely close in value, or because the relevant distinction is hard to detect under limited attention. We show that limited attention can fundamentally distort what pairwise comparisons reveal. In particular, passive comparison data cannot generally distinguish reward, attention, and default tendencies, and heterogeneous attention can make standard Bradley--Terry reward modeling recover misleading rankings. Our analysis shows that learning is governed not by the raw number of labels, but by the amount of attended information each label carries. A case study on human votes over language-model pairs from Chatbot Arena exhibits the predicted signature, a cyclic component of the comparison data that exceeds sampling noise and that no scalar reward can represent; a second case study on perceptual comparisons shows that response times and gaze carry gap information that the labels do not. This perspective suggests that human feedback should be treated not as direct revealed preference, but as an attention-limited measurement process: a weak preference signal may reflect hidden evaluation difficulty rather than genuine indifference.

Reward models in RLHF are trained to score only the final token of a response - a choice that discards rich signal from every intermediate position and produces models whose token-level outputs are noise. We argue this is a missed opportunity: a well-trained reward model's output at any token should represent the conditional expectation of the final reward given the response so far. We introduce Temporally Coherent Reward Modeling (TCRM), which induces this property via two regularization terms on top of the standard Bradley-Terry loss, with minimizers provably equal to conditional expectations. The regularizers correspond to Monte Carlo and TD value-learning objectives, establishing a direct connection to RL value functions. TCRM requires zero changes to architecture, data, or inference, yet unlocks three capabilities from one principle: interpretable token-level reward trajectories (middle-token pairwise accuracy improved from 50% to 88.9%, final-token accuracy preserved); state-of-the-art PRM performance on ProcessBench (44.9% average F1) among models trained only on outcome data; and unified reward/value modeling in PPO, reducing peak GPU memory by 27% and step time by 19% with matching LLM quality.

Current critic-less RLHF methods aggregate multi-objective rewards via an arithmetic mean, leaving them vulnerable to constraint neglect: high-magnitude success in one objective can numerically offset critical failures in others (e.g., safety or formatting), masking low-performing "bottleneck" rewards vital for reliable multi-objective alignment. We propose Reward-Variance Policy Optimization (RVPO), a risk-sensitive framework that penalizes inter-reward variance during advantage aggregation, shifting the objective from "maximize sum" to "maximize consistency." We show via Taylor expansion that a LogSumExp (SoftMin) operator effectively acts as a smooth variance penalty. We evaluate RVPO on rubric-based medical and scientific reasoning with up to 17 concurrent LLM-judged reward signals (Qwen2.5-3B/7B/14B) and on tool-calling with rule-based constraints (Qwen2.5-1.5B/3B). By preventing the model from neglecting difficult constraints to exploit easier objectives, RVPO improves overall scores on HealthBench (0.261 vs. 0.215 for GDPO at 14B, $p < 0.001$) and maintains competitive accuracy on GPQA-Diamond without the late-stage degradation observed in other multi-reward methods, demonstrating that variance regularization mitigates constraint neglect across model scales without sacrificing general capabilities.