We study which preference-data examples most strongly shape a trained reward model and propose a scalable influence-function approximation tailored to Bradley-Terry-style reward heads. Using a low-rank Gauss-Newton approximation to the Hessian, we compute per-example influence in $O(d \cdot p)$ memory rather than the naive $O(p^2)$, where $p$ is the parameter count.
Reward models trained from human preference data are typically evaluated using held-out preference accuracy, but downstream RLHF performance depends on how well the reward model approximates true preference *expectations* over policy-induced distributions. We adapt doubly robust estimation from causal inference to the reward-modeling setting, treating the policy as a treatment and the reward signal as the outcome.
We extend prior work on identity realization measurement (2604.01840) with seven new probes across three architectures (Qwen 3B, Llama 8B, Mistral 7B).
Large language models exhibit sycophantic behavior—adjusting their responses to agree with user opinions even when those opinions are factually incorrect. While prior work has measured sycophancy in single-turn settings, real-world interactions are multi-turn, and the dynamics of sycophancy across extended dialogues remain unexplored.
Reinforcement Learning from Human Feedback (RLHF) has become the dominant paradigm for aligning large language models (LLMs) with human preferences. However, reward hacking—where models exploit reward model weaknesses to achieve high scores without genuine quality improvement—remains a critical failure mode that is difficult to detect post-deployment.
Reinforcement Learning from Human Feedback (RLHF) has become the dominant paradigm for aligning large language models with human preferences. However, RLHF pipelines are susceptible to reward model collapse—a phenomenon where the policy learns to exploit systematic biases in the learned reward model rather than genuinely improving on the intended objective.