Reinforcement learning from human feedback (RLHF) has become a core post-training step for aligning large language models, yet the reward signal used in RLHF is only a learned proxy for true human utility. From an operations research perspective, this creates a decision problem under objective misspecification: the policy is optimized against an estimated reward, while deployment performance is determined by an unobserved objective. The resulting gap leads to reward over-optimization, or Goodharting, where proxy reward continues to improve even after true quality deteriorates. Existing mitigations address this problem through uncertainty penalties, pessimistic rewards, or conservative constraints, but they can be computationally burdensome and overly pessimistic. We propose Wasserstein distributionally robust regret optimization (DRRO) for RLHF. Instead of pessimizing worst-case value as in standard DRO, DRRO pessimizes worst-case regret relative to the best policy under the same plausible reward perturbation. We study the promptwise problem through a simplex allocation model and show that, under an $\ell_1$-ground-cost Wasserstein ambiguity set, the inner worst-case regret admits an exact solution and the optimal policy has a water-filling structure. These results lead to a practical policy-gradient algorithm with a simple sampled-bonus interpretation and only minor changes to GRPO-style RLHF training. The framework also clarifies theoretically why DRRO is less pessimistic than DRO, and our experiments show that DRRO mitigates over-optimization more effectively than existing baselines while standard DRO is systematically over-pessimistic.

Standard Bradley--Terry (BT) reward models are limited when human preferences are pluralistic. Although soft preference labels preserve disagreement information, BT can only express it by shrinking reward margins. Gaussian reward models provide an alternative by jointly predicting a reward mean and a reward variance, but suffer from a fundamental non-identifiability from pairwise preferences alone. We propose Anchor-guided Variance-aware Reward Modeling, a framework that resolves this non-identifiability by augmenting preference data with two coarse response-level anchor labels. Building on this, we prove that two anchors are sufficient for identification, develop a joint training objective and establish a non-asymptotic convergence rate for both the estimated reward mean and variance functions. Across simulation studies and four real-world diverging-preference datasets, our method consistently improves reward modeling performance and downstream RLHF, including PPO training and best-of-$N$ selection.

Large language models are increasingly deployed with test-time strategies: sample $N$ responses, score them with a reward model or verifier, and return the best. This deployment rule exposes a mismatch in post-training: standard objectives optimize the mean reward of a single response, whereas best-of-$N$ performance is governed by the upper tail of the reward distribution. Recent test-time-aware objectives partly address this mismatch, but typically assume that training can use the same per-prompt rollout budget as deployment, which is impractical when post-training must cover many prompts while deployment can allocate much larger per-prompt test-time compute. We study this budget-mismatch regime, where only $m\ll N$ per-prompt rollouts are available during training but the target objective is best-of-$N$ deployment. Under structural assumptions on the reward tails, we show that the policy gradient of the best-of-$N$ objective can be approximated from a much smaller rollout group by extrapolating upper-tail statistics. This yields a family of Tail-Extrapolated estimators for best-of-$N$-oriented post-training: a simple direct estimator, Tail-Extrapolated Advantage (TEA), and a fixed-order debiased Prefix-TEA estimator based on moment cancellation. Experiments on instruction-following tasks show that TEA and Prefix-TEA improve best-of-$N$ performance across different language models, reward models and datasets under various training and test-time budget settings.

Offline evaluation of language models from usage logs is biased when model choice is confounded: the same user-side factors that influence which model is used can also influence how its output is judged, so raw comparisons of logged scores mix self-selected populations rather than estimating a common quantity of interest. A small randomized experiment can break this bias by overriding model choice, but in practice such experiments are scarce and costly. We study a three-source design that combines a large confounded observational log (OBS) for scale, a small randomized experiment (EXP) for unconfounded scoring, and an offline simulator (SIM) that replays candidate models on cached contexts. Our main result is an identification theorem showing that the randomized experiment and the simulator are together enough to recover causal model values; the observational log enters only afterward, to reduce estimation error rather than to make the causal comparison valid. Six estimator families are evaluated in a controlled semi-synthetic validation and in two real-task cached benchmarks for summarization and coding. No family dominates every regime; relative performance depends on the amount of unbiased EXP supervision and on how closely the target reward aligns with OBS-derived structure.

Having defined and validated the pairwise feedback simulator and evaluations in AlpacaFarm, we569 now turn our attention to studying methods that learn from pairwise feedback on AlpacaFarm.570 Unfortunately, the lack of existing benchmarks for learning from pairwise feedback for instruction571 following means that there has not been any open study of these methods in the instruction-following572 setting. In the remainder of this section, we will introduce our reference methods, which fall into two575 categories based on whether they fit a surrogate reward model as part of the learning process.576 FeedME is a method proposed by OpenAI [45] that incorporates human feedback578 with supervised fine-tuning on model generations that are rated 7/7 by human labelers. We adapt579 this approach to the pairwise feedback setting and call this baseline binary FeedME. This approach580 fine-tunes the SFT model on the chosen response in each preference pair with supervised learning.581 Motivated by controllable generation through conditioning [27, 34,582 29, 21], we propose binary reward conditioning, a baseline method that fine-tunes the SFT model583 with the feedback data Dpairwise by conditioning instances with either a positive or negative control584 token. Specifically, for each instance (x,y0,y1,z) 2D pairwise, the string concatenation of instruction585 x and response yz denoted as [x,yz] is prepended with the positive token and used in supervised586 fine-tuning (similarly [x,y1 z]is prepended with the negative token). This process creates a modified587 demonstration dataset that is double the size of Dpairwise. At test time, we draw samples from the588 fine-tuned model conditioned on the positive token.589 A.2 Methods that optimize a surrogate reward function590 We now describe methods that incorporate feedback by first building a surrogate reward model with591 pairwise feedback data. To start, we describe the step of training the surrogate reward model.592 While this can be a powerful approach,596 we will see that it can also lead to over-optimization [19] where models learn to exploit the reward597 model rather than achieve high true reward. We now describe 4 methods that leverage the surrogate598 reward model.599

Training language models via reinforcement learning often relies on imperfect proxy rewards, since ground truth rewards that precisely define the intended behavior are rarely available. Standard metrics for assessing the quality of proxy rewards, such as ranking accuracy, treat incorrect rewards as strictly harmful. In this work, however, we highlight that not all deviations from the ground truth are equal. By theoretically analyzing which outputs attract probability during policy gradient optimization, we categorize reward errors according to their effect on the increase in ground truth reward. The analysis establishes that reward errors, though conventionally viewed as harmful, can also be benign or even beneficial by preventing the policy from stalling around outputs with mediocre ground truth reward. We then present two practical implications of our theory. First, for reinforcement learning from human feedback (RLHF), we develop reward model evaluation metrics that account for the harmfulness of reward errors. Compared to standard ranking accuracy, these metrics typically correlate better with the performance of a language model after RLHF, yet gaps remain in robustly evaluating reward models. Second, we provide insights for reward design in settings with verifiable rewards. A key theme underlying our results is that the effectiveness of a proxy reward function depends heavily on its interaction with the initial policy and learning algorithm.