Papers with RewardBench
RewardBench: Evaluating Reward Models for Language Modeling (2025.findings-naacl)
Copied to clipboard
Nathan Lambert, Valentina Pyatkin, Jacob Morrison, Lester James Validad Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, Noah A. Smith, Hannaneh Hajishirzi
| Challenge: | Evaluating reward models presents an opportunity to understand the opaque technologies used for alignment of language models. |
| Approach: | They present a benchmark dataset and code-base for evaluation of reward models . they use prompt-chosen-rejected trios to benchmark how they perform on queries . |
| Outcome: | The proposed dataset compares RMs with other models on a set of questions. |
reWordBench: Benchmarking and Improving the Robustness of Reward Models with Transformed Inputs (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing reward models have a high performance on benchmarks, but performance degradation is often due to overfitting. |
| Approach: | They propose to explicitly train reward models to assign similar scores to paraphrases to improve their robustness. |
| Outcome: | The proposed model reduces degradation by half for the Chat Hard subset in RewardBench. |
LMUNIT: Fine-grained Evaluation with Natural Language Unit Tests (2025.findings-emnlp)
Copied to clipboard
Jon Saad-Falcon, Rajan Pathe Vivek, William Berrios, Nandita Shankar Naik, Matija Franklin, Bertie Vidgen, Amanpreet Singh, Douwe Kiela, Shikib Mehri
| Challenge: | Using natural language unit tests, language models are costly and noisy, and automated metrics provide only coarse, difficult-to-interpret signals. |
| Approach: | They propose a paradigm that decomposes response quality into explicit, testable criteria and a unified scoring model, LMUnit, which combines multi-objective training across preferences, direct ratings, and natural language rationales. |
| Outcome: | The proposed paradigm significantly improves inter-annotator agreement and enables more effective LLM development workflows. |
Hybrid Preferences: Learning to Route Instances for Human vs. AI Feedback (2025.acl-long)
Copied to clipboard
Lester James Validad Miranda, Yizhong Wang, Yanai Elazar, Sachin Kumar, Valentina Pyatkin, Faeze Brahman, Noah A. Smith, Hannaneh Hajishirzi, Pradeep Dasigi
| Challenge: | Learning from human feedback has enabled the alignment of language models (LMs) with human preferences. |
| Approach: | They propose a Hybrid Preference routER that defers an annotation to either humans or LMs, achieving better annotation quality while reducing the cost of human-only annotation. |
| Outcome: | The proposed model achieves better annotation quality while reducing the cost of human-only annotation. |
Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Reinforcement learning from human feedback (RLHF) is the primary method for aligning large language models with human preferences. |
| Approach: | They propose to train an Absolute-Rating Multi-Objective Reward Model with multi-dimensional absolute-rating data. |
| Outcome: | The proposed model outperforms the LLM-as-a-judge method on RewardBench . it achieves state-of-the-art performance on the benchmark . |
Act-Adaptive Margin: Dynamically Calibrating Reward Models for Subjective Ambiguity (2026.acl-long)
Copied to clipboard
Feiteng Fang, Dingwei Chen, Xiang Huang, Ting-En Lin, Yuchuan Wu, Xiong Liu, Jing Ye, Ziqiang Liu, Haonan Zhang, Liang Zhu, Hamid Alinejad-Rokny, Min Yang, Yongbin Li
| Challenge: | Existing approaches to reward modeling in reinforcement learning tasks are limited when dealing with ambiguous preferences. |
| Approach: | They propose to use AAM to dynamically calibrate preference margins using the Bradley-Terry model's internal parameter knowledge to improve reward modeling in subjective tasks. |
| Outcome: | The proposed approach improves reward modeling by dynamically calibrating preference margins using the model’s internal parameter knowledge. |
Improve LLM-as-a-Judge Ability as a General Ability (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies focus on generative judges, but only on their judge ability. |
| Approach: | They propose a method that leverages the generative and reasoning capabilities of large language models to evaluate LLM responses across diverse scenarios, providing accurate preference signals. |
| Outcome: | The proposed model performs on RewardBench with only 2% to 40% of the data required by other training frameworks. |
Can External Validation Tools Improve Annotation Quality for LLM-as-a-Judge? (2025.acl-long)
Copied to clipboard
| Challenge: | Pairwise feedback is widely used to evaluate and provide feedback to large language models (LLMs). |
| Approach: | They propose a tool-using agentic system to provide higher quality feedback on three challenging response domains: long-form factual, math and code tasks. |
| Outcome: | The proposed system can provide higher quality pairwise comparisons on three domains, independent of the LLM’s internal knowledge and biases. |
Margin Matching Preference Optimization: Enhanced Model Alignment with Granular Feedback (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for large language models rely on binary labels that fail to capture the subtle differences in relative quality between pairs. |
| Approach: | They propose a method that incorporates relative quality margins into optimization to improve LLM policies and reward models. |
| Outcome: | The proposed approach outperforms baseline methods on popular benchmarks including MT-bench and RewardBench. |
Mutual-Taught for Co-adapting Policy and Reward Models (2025.acl-long)
Copied to clipboard
Tianyuan Shi, Canbin Huang, Fanqi Wan, Longguang Zhong, Ziyi Yang, Weizhou Shen, Xiaojun Quan, Ming Yan
| Challenge: | Experimental results show that this iterative approach leads to consistent improvements in both the policy model and reward model. |
| Approach: | They propose a method that iteratively improves both the policy model and reward model without requiring additional human annotation. |
| Outcome: | The proposed method improves both the policy model and reward model without human annotation. |
PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling (2026.acl-long)
Copied to clipboard
| Challenge: | Existing reward models lack generative and reasoning capabilities, resulting in poor performance. |
| Approach: | They propose a reward-aware task-adaptive reward model that enables pointwise training using readily available pairwise data via a novel Preference-Aware Reward mechanism. |
| Outcome: | The proposed reward model achieves an average relative improvement of 8.7% over the base models on RewardBench and RMBench. |
Foundational Autoraters: Taming Large Language Models for Better Automatic Evaluation (2024.emnlp-main)
Copied to clipboard
| Challenge: | evaluating large language models' output is difficult due to the high cost of human evaluation. |
| Approach: | They propose a family of foundational large autorater models that train on over 100 quality assessment tasks. |
| Outcome: | The proposed model outperforms models on 8 of 12 autorater benchmarks on 53 quality assessment tasks. |
IPO: Your Language Model is Secretly a Preference Classifier (2025.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning from human feedback (RLHF) is the primary method for aligning large language models with human preferences, but it often incurs significant computational and financial costs due to its reliance on training external reward models or human-labeled preferences. |
| Approach: | They propose an alternative approach that leverages generative LLMs as preference classifiers to reduce the dependence on external reward models or human-labeled preferences. |
| Outcome: | The proposed approach reduces the dependence on external reward models or human-labeled preferences by using generative LLMs as preference classifiers. |
DORM: Preference Data Weights Optimization for Reward Modeling in LLM Alignment (2025.findings-emnlp)
Copied to clipboard
Rongzhi Zhang, Chenwei Zhang, Xinyang Zhang, Liang Qiu, Haoming Jiang, Yuchen Zhuang, Qingru Zhang, Hyokun Yun, Xian Li, Bing Yin, Tuo Zhao, Chao Zhang
| Challenge: | Existing approaches to align large language models with human preferences are noisy and varying in importance of preference samples. |
| Approach: | a new method enhances reward modeling by learning to dynamically weigh preference data. |
| Outcome: | a new method improves the performance of large language models with human preferences . it initializes data importance and iteratively refines them to maximize validation performance. |