RewardBench: Evaluating Reward Models for Language Modeling
Reviewed by Pithpith:ONM3W2MHopen to challenge →
read the original abstract
Reward models (RMs) are at the crux of successfully using RLHF to align pretrained models to human preferences, yet there has been relatively little study that focuses on evaluation of those models. Evaluating reward models presents an opportunity to understand the opaque technologies used for alignment of language models and which values are embedded in them. Resources for reward model training and understanding are sparse in the nascent open-source community around them. To enhance scientific understanding of reward models, we present RewardBench, a benchmark dataset and code-base for evaluation. The RewardBench dataset is a collection of prompt-chosen-rejected trios spanning chat, reasoning, and safety, to benchmark how reward models perform on challenging, structured and out-of-distribution queries. We create specific comparison datasets for RMs that have subtle, but verifiable reasons (e.g. bugs, incorrect facts) why one answer should be preferred to another. On the RewardBench leaderboard, we evaluate reward models trained with a variety of methods, such as the direct MLE training of classifiers and the implicit reward modeling of Direct Preference Optimization (DPO). We present many findings on propensity for refusals, reasoning limitations, and instruction following shortcomings of various reward models towards a better understanding of the RLHF process.
This paper has not been read by Pith yet.
Forward citations
Cited by 31 Pith papers
-
TabArena: A Living Benchmark for Machine Learning on Tabular Data
TabArena launches a dynamic, updatable benchmarking system for tabular ML that shows boosted trees remain competitive, deep learning matches them under larger budgets with ensembling, foundation models excel on small ...
-
QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents
QVal is a new evaluation framework that directly measures dense supervision quality via Q-alignment to a reference policy, showing simple prompting baselines outperform 21 other methods across environments and models.
-
EST-PRM: Stress-Testing Process Reward Models Before They Become Load-Bearing
EST-PRM stress-tests five PRM models on 4,687 reasoning chains from MATH-500, GSM8K, and PRMBench using three label-preserving transformations and reports model-specific vulnerability patterns.
-
PReMISE: Policy Rubrics as Measurement Specifications for LLM Judges
PReMISE discovers and audits rubric sets for LLM judges, finding no existing source meets all reliability, preference-fit, and robustness criteria simultaneously while showing two repair methods improve accuracy and r...
-
Enhanced LLM Reasoning by Optimizing Reward Functions with Search-Driven Reinforcement Learning
Iterative search over reward functions with ranked feedback in GRPO training improves LLM math reasoning, achieving F1 of 0.795 on GSM8K versus 0.609 for baseline.
-
Enhanced LLM Reasoning by Optimizing Reward Functions with Search-Driven Reinforcement Learning
Iterative LLM-driven search over reward functions, screened via GRPO on GSM8K, raises F1 from 0.609 baseline to 0.795 with ensembles on Llama-3.2-3B.
-
Four-Axis Decision Alignment for Long-Horizon Enterprise AI Agents
Long-horizon enterprise AI agents' decisions decompose into four measurable axes, with benchmark experiments on six memory architectures revealing distinct weaknesses and reversing a pre-registered prediction on summa...
-
CompliBench: Benchmarking LLM Judges for Compliance Violation Detection in Dialogue Systems
CompliBench uses simulation and adversarial flaw injection to create labeled dialogue data showing that top proprietary LLMs perform poorly at spotting guideline violations while fine-tuned smaller models outperform t...
-
Evalet: Evaluating Large Language Models through Functional Fragmentation
Evalet applies functional fragmentation to deliver fragment-level qualitative analysis of LLM evaluations, with a user study showing 48% more misalignment detections than holistic scoring.
-
Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing
Magpie synthesizes 300K high-quality alignment instructions from Llama-3-Instruct via auto-regressive prompting on partial templates, enabling fine-tuned models to match official instruct performance on AlpacaEval, Ar...
-
Addressing Over-Refusal in LLMs with Competing Rewards
SEAR trains one LLM via adversarial process rewards to explore harmful reasoning paths but flip to safe outputs, reducing over-refusal while preserving safety.
-
A Finite-Calibration Regime Map for LLM Judge Panels
The paper introduces a finite-calibration regime map and Finite-Calibration Panel Selection selector, finding scalar aggregation wins on most real benchmark-budget combinations while joint tables help when interaction...
-
Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR
POW3R adapts rubric criterion weights via rollout contrast in RLVR to improve mean reward, strict completion rates, and training speed over static rubric aggregation on multimodal and text tasks.
-
Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders
Sparse autoencoders isolate unstable features in reward model representations and enable two mitigation techniques that reduce preference errors on perturbed inputs without retraining.
-
Deployment-Relevant Alignment Cannot Be Inferred from Model-Level Evaluation Alone
Deployment-relevant AI alignment cannot be inferred from model-level evaluations alone, as benchmark audits show missing interaction support and cross-model tests reveal model-dependent scaffold effects.
-
Zero-Shot Detection of LLM-Generated Text via Implicit Reward Model
IRM derives implicit reward signals from off-the-shelf LLMs to detect generated text zero-shot and reports better results than prior zero-shot and supervised detectors on the DetectRL benchmark.
-
SAM 3D: 3Dfy Anything in Images
SAM 3D reconstructs 3D objects from single images with geometry, texture, and pose using human-model annotated data at scale and synthetic-to-real training, achieving 5:1 human preference wins.
-
PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling
PaTaRM converts pairwise preference data into pointwise reward signals via a novel PAR mechanism and task-adaptive rubrics, reporting 8.7% gains on RewardBench/RMBench and 13.6% relative RLHF improvement.
-
Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images
Locate-Then-Examine improves AI-generated image detection by localizing suspicious regions first then performing region-aware re-examination, while releasing the TRACE dataset of 20k annotated images.
-
Exploring the Secondary Risks of Large Language Models
Introduces secondary risks as a new class of LLM failures from benign prompts, defines two primitives, proposes SecLens search framework, and releases SecRiskBench showing risks are widespread across 16 models.
-
RewardBench 2: Advancing Reward Model Evaluation
RewardBench 2 is a new benchmark that supplies challenging fresh human prompts for reward model evaluation, yielding lower average scores but higher correlation with downstream best-of-N sampling and RLHF training per...
-
Improving Video Generation with Human Feedback
A human preference dataset and VideoReward model enable Flow-DPO and Flow-NRG to produce smoother, better-aligned videos from text prompts in flow-based generators.
-
Pairwise Reference Alignment as a Model-Level Ordinal Observable
Pairwise reference alignment is formulated as an ordinal observable equal to the probability that a model score agrees with reference preferences on triples (x, y+, y-), with centered statistics, margin extensions, es...
-
Judging the Judges: A Systematic Evaluation of Bias Mitigation Strategies in LLM-as-a-Judge Pipelines
Style bias dominates LLM-as-a-Judge systems far more than position bias, with debiasing strategies providing model-dependent gains and public tools released for replication.
-
Judging the Judges: A Systematic Evaluation of Bias Mitigation Strategies in LLM-as-a-Judge Pipelines
Gemini 2.5 Flash with a Combined Budget debiasing strategy achieves 71.0% judge agreement at ~$0.001/evaluation, outperforming frontier models at 15x lower cost.
-
Difficulty-Based Preference Data Selection by DPO Implicit Reward Gap
Selecting preference pairs whose DPO implicit reward gap is small yields better LLM alignment than random or baseline selection while using only 10% of the data.
-
Discriminatory Compliance: How LLMs Answer Queries from Protected Groups
State-of-the-art LLMs respond inconsistently to queries from protected-group personas, with some responses omitting key information that should be provided.
-
Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization
Proxy RL produces a staged proxy-internalization capability that emerges before and predicts reward hacking in coding environments.
-
Skywork-Reward: Bag of Tricks for Reward Modeling in LLMs
Data-centric filtering yields an 80K preference dataset and reward models that lead RewardBench while boosting other top entries.
-
Qwen2.5 Technical Report
Qwen2.5 LLMs scale pre-training data to 18 trillion tokens and apply multistage reinforcement learning, achieving competitive performance on benchmarks with models up to 5 times larger.
-
LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods
A survey that organizes LLMs-as-judges research into functionality, methodology, applications, meta-evaluation, and limitations.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.