Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling

TL;DR

The paper addresses reward hacking in RLHF due to noisy human preferences and superficial biases. It proposes Bayesian Non-Negative Reward Model (BNRM), which integrates non-negative factor analysis with a Bayesian Bradley–Terry framework, yielding sparse, disentangled local factors and a global reward dictionary to capture uncertainty. BNMR uses amortized variational inference with Weibull posteriors to scale to large LLMs, and demonstrates improved robustness to distribution shifts, reduced reliance on spurious cues, and enhanced interpretability of reward decompositions compared with strong baselines. The work shows clear empirical gains in both reward modeling and RLHF tasks, suggesting principled sparsity and uncertainty modeling as effective inductive biases for reliable alignment in LLMs.

Abstract

Reward models learned from human preferences are central to aligning large language models (LLMs) via reinforcement learning from human feedback, yet they are often vulnerable to reward hacking due to noisy annotations and systematic biases such as response length or style. We propose Bayesian Non-Negative Reward Model (BNRM), a principled reward modeling framework that integrates non-negative factor analysis into Bradley-Terry (BT) preference model. BNRM represents rewards through a sparse, non-negative latent factor generative process that operates at two complementary levels: instance-specific latent variables induce disentangled reward representations, while sparsity over global latent factors acts as an implicit debiasing mechanism that suppresses spurious correlations. Together, this disentanglement-then-debiasing structure enables robust uncertainty-aware reward learning. To scale BNRM to modern LLMs, we develop an amortized variational inference network conditioned on deep model representations, allowing efficient end-to-end training. Extensive empirical results demonstrate that BNRM substantially mitigates reward over-optimization, improves robustness under distribution shifts, and yields more interpretable reward decompositions than strong baselines.

Figures (10)

• Figure 1: Motivation on disentanglement and debiasing for alleviating spurious correlations in reward modeling.
• Figure 2: Graphical model representations. (a) The standard BT model; (b) our proposed BNRM. Here, $\mathbf{x}$ denotes the prompt, $\mathbf{y}_1, \mathbf{y}_2$ are candidate responses, and nodes represent the predictive process for the preference $\mathbf{y}_1 \succ \mathbf{y}_2$.
• Figure 3: Variational Inferencer for BNRM.
• Figure 4: ID and OOD evaluation results for BT-BNRM. (a): performance when training on a varying number of samples. (b): performance under different label-noise ratios. Solid bars denote the BT baseline, hatched bars denote our BT-BNRM.
• Figure 5: Pearson correlation and mean reward score between response length and reward score on the RM-Bench Hard subset. The top plot shows the correlation between response length and reward score. The x-axis is log-scaled for better visual clarity. The bottom plot reports the average reward score within each length bucket, which visually highlights the non-negative property of our BT-BNRM.