Reward Model Evaluation

Reward models (RMs) have become an indispensable fixture of the language model (LM) post-training playbook, enabling policy alignment and test-time scaling. Research on the application of RMs in code generation, however, has been comparatively sparse, with existing work largely focusing on execution feedback. This choice constrains post-training to optimizing functional correctness over self-contained executable code. In this work, we examine the training and evaluation of multilingual, multi-criteria code RMs. To this end, we first compile Themis-CodeRewardBench, a benchmark to evaluate code RMs across five preference dimensions (i.e., criteria) and eight programming languages, on which we profile 50+ code, math, and general-purpose RMs. Observing the limited proficiency of current RMs beyond scoring for functional correctness, we develop Themis-CodePreference, the largest open-source collection of code preferences to date (more than 350k preference pairs), and use it to train Themis-RM, a suite of multilingual code reward models for flexible multi-criteria scoring, ranging in size from 600M to 32B parameters. Our experiments and ablations demonstrate positive scaling trends, strong cross-lingual transfer when training on diverse preferences, and the importance of multi-criteria training for reliable code reward modeling.

Large language models have achieved remarkable progress in text generation but still struggle with generative writing tasks. In terms of evaluation, existing benchmarks evaluate writing reward models coarsely and fail to measure performance from the perspective of specific requirements. In terms of training, existing training methods either use LLM-as-a-judge approaches or train coarse-grained reward models, lacking fine-grained requirement-adherence reward modeling. To address these issues, we propose a fine-grained evaluation pipeline WEval for writing reward models and a fine-grained reinforcement learning training framework WRL. The evaluation data of WEval covers multiple task categories and requirement types, enabling systematic evaluation of writing reward models by measuring the correlation between the rankings of the reward model and gold rankings. WRL constructs positive and negative samples by selectively dropping instruction requirements, allowing for more precise reward model training. Experiments show that our models achieve substantial improvements across various writing benchmarks and exhibit strong generalization. The code and data are publicly available at https://github.com/Rainier-rq1/From_Coarse_to_Fine.

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.

Using a language model to score or rank candidate responses has become a scalable alternative to human evaluation in reinforcement learning from human feedback (RLHF) pipelines, benchmarking, and application layer evaluations. However, output reliability depends heavily on prompting and aggregation strategy. We present an empirical investigation of four drop-in techniques -- ensemble scoring, task-specific criteria injection, calibration context, and adaptive model escalation -- for improving LLM judge accuracy on RewardBench 2, with a unifying lens of noise control on the stochastic judge: ensembling as Monte Carlo averaging over per-call noise, criteria injection as between-response discrimination sharpening, and per-response score variance as an uncertainty signal. Ensemble scoring and task-specific criteria injection (the latter virtually cost free) together reach up to 85.8% accuracy, +13.5pp over baseline. Calibration context and adaptive model escalation also improve over baseline but are dominated by criteria + ensembling on the cost-accuracy Pareto frontier. Small models benefit disproportionately from ensembling, making high-accuracy LLM judges accessible at low cost. We show that these techniques generalise across model providers, evaluating on both OpenAI GPT and Anthropic Claude families.

We present a discriminative multimodal reward model that scores all candidate responses in a single forward pass. Conventional discriminative reward models evaluate each response independently, requiring multiple forward passes, one for each potential response. Our approach concatenates multiple responses with separator tokens and applies cross-entropy over their scalar scores, enabling direct comparative reasoning and efficient $N$-way preference learning. The multi-response design also yields up to $N\times$ wall-clock speedup and FLOPs reduction over conventional single-response scoring. To enable $N$-way reward evaluation beyond existing pairwise benchmarks, we construct two new benchmarks: (1) MR$^2$Bench-Image contains human-annotated rankings over responses from 8 diverse models; (2) MR$^2$Bench-Video is a large-scale video-based reward benchmark derived from 94K crowdsourced pairwise human judgments over video question-answering spanning 19 models, denoised via preference graph ensemble. Both benchmarks provide 4-response evaluation variants sampled from the full rankings. Built on a 4B vision-language backbone with LoRA fine-tuning and a lightweight MLP value head, our model achieves state-of-the-art results on six multimodal reward benchmarks, including MR$^2$Bench-Image, MR$^2$Bench-Video, and four other existing benchmarks. Our model outperforms existing larger generative and discriminative reward models. We further demonstrate that our reward model, when used in reinforcement learning with GRPO, produces improved policy models that maintain performance across standard multimodal benchmarks while substantially improving open-ended generation quality, outperforming a single-response discriminative reward model (RM) baseline by a large margin in both training stability and open-ended generation quality.

In classical Reinforcement Learning from Human Feedback (RLHF), Reward Models (RMs) serve as the fundamental signal provider for model alignment. As Large Language Models evolve into agentic systems capable of autonomous tool invocation and complex reasoning, the paradigm of reward modeling faces unprecedented challenges -- most notably, the lack of benchmarks specifically designed to assess RM capabilities within tool-integrated environments. To address this gap, we present Plan-RewardBench, a trajectory-level preference benchmark designed to evaluate how well judges distinguish preferred versus distractor agent trajectories in complex tool-using scenarios. Plan-RewardBench covers four representative task families -- (i) Safety Refusal, (ii) Tool-Irrelevance / Unavailability, (iii) Complex Planning, and (iv) Robust Error Recovery -- comprising validated positive trajectories and confusable hard negatives constructed via multi-model natural rollouts, rule-based perturbations, and minimal-edit LLM perturbations. We benchmark representative RMs (generative, discriminative, and LLM-as-Judge) under a unified pairwise protocol, reporting accuracy trends across varying trajectory lengths and task categories. Furthermore, we provide diagnostic analyses of prevalent failure modes. Our results reveal that all three evaluator families face substantial challenges, with performance degrading sharply on long-horizon trajectories, underscoring the necessity for specialized training in agentic, trajectory-level reward modeling. Ultimately, Plan-RewardBench aims to serve as both a practical evaluation suite and a reusable blueprint for constructing agentic planning preference data.

Pluralistic alignment has emerged as a critical frontier in the development of Large Language Models (LLMs), with reward models (RMs) serving as a central mechanism for capturing diverse human values. While benchmarks for general response quality are prevalent, evaluating how well reward models account for individual user preferences remains an open challenge. To bridge this gap, we introduce Personalized RewardBench, a novel benchmark designed to rigorously assess reward models' capacity to model personalized preferences. We construct chosen and rejected response pairs based on strict adherence to (or violation of) user-specific rubrics, ensuring that preference distinctions are uniquely tailored to the individual. In particular, human evaluations confirm that the primary discriminative factor between pairs is strictly personal preference, with both responses maintaining high general quality (e.g., correctness, relevance and helpfulness). Extensive testing reveals that existing state-of-the-art reward models struggle significantly with personalization, peaking at an accuracy of just 75.94%. Crucially, because an effective reward model benchmark should predict a reward model's performance on downstream tasks, we conduct experiments demonstrating that our benchmark exhibits a significantly higher correlation with downstream performance in both Best-of-N (BoN) sampling and Proximal Policy Optimization (PPO) compared to existing baselines. These findings establish Personalized RewardBench as a robust and accurate proxy for evaluating reward models' performance in downstream applications.

Large language models (LLMs) are now widely used as judges, yet their decisions can change under presentation choices that should be irrelevant. We study one such source of instability: candidate-order sensitivity in listwise factuality evaluation, where several answers can look similarly polished while differing substantially in hallucination risk. We introduce PCFJudge, an inference-time method that reruns the same factuality-first listwise prompt over multiple orderings of the same candidate set and aggregates the resulting scores, ranks, and uncertainty signals into a single consensus decision. On RewardBench 2 Factuality, the final seven-permutation aggregate (K=7) improves top-1 selection accuracy from 86.00% to 91.33% with GPT-5.4 and from 86.33% to 89.67% with Claude Sonnet 4.6. These results suggest that candidate order can be a meaningful source of factuality-judging error and that marginalizing over this nuisance variation can improve the reliability of LLM evaluation.

The widespread adoption of reinforcement learning-based alignment highlights the growing importance of reward models. Various benchmarks have been built to evaluate reward models in various domains and scenarios. However, a significant gap remains in assessing reward models for long-form generation, despite its critical role in real-world applications. To bridge this, we introduce Long-form RewardBench, the first reward modeling testbed specifically designed for long-form generation. Our benchmark encompasses five key subtasks: QA, RAG, Chat, Writing, and Reasoning. We collected instruction and preference data through a meticulously designed multi-stage data collection process, and conducted extensive experiments on 20+ mainstream reward models, including both classifiers and generative models. Our findings reveal that current models still lack long-form reward modeling capabilities. Furthermore, we designed a novel Long-form Needle-in-a-Haystack Test, which revealed a correlation between reward modeling performance and the error's position within a response, as well as the overall response length, with distinct characteristics observed between classification and generative models. Finally, we demonstrate that classifiers exhibit better generalizability compared to generative models trained on the same data. As the first benchmark for long-form reward modeling, this work aims to offer a robust platform for visualizing progress in this crucial area.

Instruction-following is a foundational capability of large language models (LLMs), with its improvement hinging on scalable and accurate feedback from judge models. However, the reliability of current judge models in instruction-following remains underexplored due to several deficiencies of existing meta-evaluation benchmarks, such as their insufficient data coverage and oversimplified pairwise evaluation paradigms that misalign with model optimization scenarios. To this end, we propose IF-RewardBench, a comprehensive meta-evaluation benchmark for instruction-following that covers diverse instruction and constraint types. For each instruction, we construct a preference graph containing all pairwise preferences among multiple responses based on instruction-following quality. This design enables a listwise evaluation paradigm that assesses the capabilities of judge models to rank multiple responses, which is essential in guiding model alignment. Extensive experiments on IF-RewardBench reveal significant deficiencies in current judge models and demonstrate that our benchmark achieves a stronger positive correlation with downstream task performance compared to existing benchmarks. Our codes and data are available at https://github.com/thu-coai/IF-RewardBench.

Recent advancements in Generative Reward Models (GRMs) have demonstrated that scaling the length of Chain-of-Thought (CoT) reasoning considerably enhances the reliability of evaluation. However, current works predominantly rely on unstructured length scaling, ignoring the divergent efficacy of different reasoning mechanisms: Breadth-CoT (B-CoT, i.e., multi-dimensional principle coverage) and Depth-CoT (D-CoT, i.e., substantive judgment soundness). To address this, we introduce Mix-GRM, a framework that reconfigures raw rationales into structured B-CoT and D-CoT through a modular synthesis pipeline, subsequently employing Supervised Fine-Tuning (SFT) and Reinforcement Learning with Verifiable Rewards (RLVR) to internalize and optimize these mechanisms. Comprehensive experiments demonstrate that Mix-GRM establishes a new state-of-the-art across five benchmarks, surpassing leading open-source RMs by an average of 8.2%. Our results reveal a clear divergence in reasoning: B-CoT benefits subjective preference tasks, whereas D-CoT excels in objective correctness tasks. Consequently, misaligning the reasoning mechanism with the task directly degrades performance. Furthermore, we demonstrate that RLVR acts as a switching amplifier, inducing an emergent polarization where the model spontaneously allocates its reasoning style to match task demands. The synthesized data and models are released at Hugging Face, and the code is released at Github.

While music generation models have evolved to handle complex multimodal inputs mixing text, lyrics, and reference audio, evaluation mechanisms have lagged behind. In this paper, we bridge this critical gap by establishing a comprehensive ecosystem for music reward modeling under Compositional Multimodal Instruction (CMI), where the generated music may be conditioned on text descriptions, lyrics, and audio prompts. We first introduce CMI-Pref-Pseudo, a large-scale preference dataset comprising 110k pseudo-labeled samples, and CMI-Pref, a high-quality, human-annotated corpus tailored for fine-grained alignment tasks. To unify the evaluation landscape, we propose CMI-RewardBench, a unified benchmark that evaluates music reward models on heterogeneous samples across musicality, text-music alignment, and compositional instruction alignment. Leveraging these resources, we develop CMI reward models (CMI-RMs), a parameter-efficient reward model family capable of processing heterogeneous inputs. We evaluate their correlation with human judgment scores on musicality and alignment on CMI-Pref along with previous datasets. Further experiments demonstrate that CMI-RM not only correlates strongly with human judgments, but also enables effective inference-time scaling via top-k filtering. Code is available at GitHub (https://github.com/Haiwen-Xia/CMI-RewardBench). Model weights: CMI-RM (https://huggingface.co/HaiwenXia/CMI-RM). Datasets: CMI-Pref-Pseudo (https://huggingface.co/datasets/HaiwenXia/cmi-pref-pseudo) and CMI-Pref (https://huggingface.co/datasets/HaiwenXia/cmi-pref)

Reward models are central to aligning large language models (LLMs) with human preferences. Yet most approaches rely on pointwise reward estimates that overlook the epistemic uncertainty in reward models arising from limited human feedback. Recent work suggests that quantifying this uncertainty can reduce the costs of human annotation via uncertainty-guided active learning and mitigate reward overoptimization in LLM post-training. However, uncertainty-aware reward models have so far been adopted without thorough comparison, leaving them poorly understood. This work introduces a unified framework, RewardUQ, to systematically evaluate uncertainty quantification for reward models. We compare common methods along standard metrics measuring accuracy and calibration, and we propose a new ranking strategy incorporating both dimensions for a simplified comparison. Our experimental results suggest that model size and initialization have the most meaningful impact on performance, and most prior work could have benefited from alternative design choices. To foster the development and evaluation of new methods and aid the deployment in downstream applications, we release our open-source framework as a Python package. Our code is available at https://github.com/lasgroup/rewarduq.

Learning from feedback advances the capabilities and safety of frontier models, but preference supervision remains costly and difficult to scale. We introduce next-token preference (NTP), an unsupervised reward-modeling method that constructs preference pairs from the continuation structure of raw web text. NTP uses each prefix's original continuation as the chosen response and nearby continuations as in-batch negatives, without candidate generation or human or model-generated preference annotations. Training on 11M tokens of math-focused web data yields steady gains on RewardBench and RewardBench 2, with gains across model families and sizes. Across mathematical and general-domain web corpora, our method improves overall RewardBench 2 accuracy by 7.6-8.3 percentage points over initialization under the same token budget. Training on math-focused text yields gains of up to +16.1 on in-domain math subsets and consistent improvements on out-of-domain safety and general subsets. These reward models improve best-of-N selection on math, code, safety, and instruction following, and support math policy optimization competitive with supervised reward models of similar size. NTP also improves the label efficiency of subsequent supervised fine-tuning. Using half the labeled pairs from a curated preference dataset, NTP initialization surpasses base initialization trained on the full dataset. These results suggest that web text contains transferable preference signals that can reduce dependence on costly preference annotations.

Personalized alignment from preference data has focused primarily on improving personal reward model (RM) accuracy, with the implicit assumption that better preference ranking translates to better personalized behavior. However, in deployment, computational constraints necessitate inference-time adaptation such as reward-guided decoding (RGD) rather than per-user policy fine-tuning. This creates a critical but overlooked requirement: reward models must not only rank preferences accurately but also effectively guide generation. We demonstrate that standard RM accuracy fails catastrophically as a selection criterion for deployment-ready personalized rewards. We introduce policy accuracy; a metric quantifying whether RGD-adapted LLMs correctly discriminate between preferred and dispreferred responses and show that upstream RM accuracy correlates only weakly with downstream policy accuracy (Kendall's tau = 0.08--0.31). More critically, we introduce Pref-LaMP the first personalized alignment benchmark with ground-truth user completions, enabling direct behavioural evaluation. On Pref-LaMP, we expose a complete decoupling between discriminative ranking and generation metrics: methods with 20-point RM accuracy differences produce almost identical output quality, and methods with high ranking accuracy can fail to generate behaviorally aligned responses. These findings reveal that the field has been optimizing for proxy metrics that do not predict deployment performance, and that current personalized alignment methods fail to operationalize preferences into behavioral adaptation under realistic deployment constraints. In contrast, we find simple in-context learning (ICL) to be highly effective - dominating all reward-guided methods for models $\geq$3B parameters, achieving $\sim$3 point ROUGE-1 gains over the best reward method at 7B scale.

The reliability of large language models (LLMs) during test-time scaling is often assessed with external verifiers or reward models that distinguish correct reasoning from flawed logic. Prior work has studied both outcome reward models (ORMs), which assess only the final answer, and process reward models (PRMs), which score intermediate reasoning steps. Although PRMs are often viewed as advantageous due to their finer-grained supervision, much of the supporting evidence comes from math-adjacent settings, and their relative benefits across broader domains remain unclear. We present the first unified evaluation of four reward model variants, discriminative ORM and PRM (dORM, dPRM) and generative ORM and PRM (gORM, gPRM), across 14 diverse domains. Contrary to conventional wisdom, we find that (i) dORM performs on par with dPRM, (ii) gPRM is not competitive, and (iii) overall, gORM is the most robust, yielding significant and consistent gains across every tested domain. We attribute the worse performance of gPRM to the stepwise scoring process, which inherits label noise from LLM-based automatic labeling, leading to difficulties in evaluating long reasoning trajectories, including those involving self-correcting reasoning. Both our theoretical analysis and empirical observations indicate that stepwise aggregation compounds errors as reasoning length increases. These findings challenge the common assumption that fine-grained supervision is always better and support generative outcome verification for multi-domain deployment. Our code is publicly available to facilitate future research in multi-domain settings.