RRM: Robust Reward Model Training Mitigates Reward Hacking

• URL: http://arxiv.org/abs/2409.13156v1
• Date: Fri, 20 Sep 2024 01:46:07 GMT
• Title: RRM: Robust Reward Model Training Mitigates Reward Hacking
• Authors: Tianqi Liu, Wei Xiong, Jie Ren, Lichang Chen, Junru Wu, Rishabh Joshi, Yang Gao, Jiaming Shen, Zhen Qin, Tianhe Yu, Daniel Sohn, Anastasiia Makarova, Jeremiah Liu, Yuan Liu, Bilal Piot, Abe Ittycheriah, Aviral Kumar, Mohammad Saleh,
• Abstract summary: Reward models (RMs) play a pivotal role in aligning large language models with human preferences. We introduce a causal framework that learns preferences independent of these artifacts. Experiments show that our approach successfully filters out undesirable artifacts, yielding a more robust reward model.
• Score: 51.12341734942797
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. However, traditional RM training, which relies on response pairs tied to specific prompts, struggles to disentangle prompt-driven preferences from prompt-independent artifacts, such as response length and format. In this work, we expose a fundamental limitation of current RM training methods, where RMs fail to effectively distinguish between contextual signals and irrelevant artifacts when determining preferences. To address this, we introduce a causal framework that learns preferences independent of these artifacts and propose a novel data augmentation technique designed to eliminate them. Extensive experiments show that our approach successfully filters out undesirable artifacts, yielding a more robust reward model (RRM). Our RRM improves the performance of a pairwise reward model trained on Gemma-2-9b-it, on RewardBench, increasing accuracy from 80.61% to 84.15%. Additionally, we train two DPO policies using both the RM and RRM, demonstrating that the RRM significantly enhances DPO-aligned policies, improving MT-Bench scores from 7.27 to 8.31 and length-controlled win-rates in AlpacaEval-2 from 33.46% to 52.49%.

Related papers

• Energy-Based Reward Models for Robust Language Model Alignment [9.843359827321194]
  We introduce Energy-Based Reward Model (EBRM), a lightweight post-hoc refinement framework for Reward Models (RMs) EBRM models the reward distribution explicitly, capturing uncertainty in human preferences and mitigating the impact of noisy or misaligned annotations. Empirical evaluations demonstrate significant improvements in robustness and generalization, achieving up to a 5.97% improvement in safety-critical alignment tasks.
  arXiv  Detail & Related papers  (2025-04-17T17:47:15Z)
• Free Process Rewards without Process Labels [55.14044050782222]
  We show that an textitimplicit PRM can be obtained at no additional cost, by simply training an ORM on the cheaper response-level labels. We show that our implicit PRM, when instantiated with the cross-entropy (CE) loss, is more data-efficient and can keep improving generation models even when trained with only one response per instruction.
  arXiv  Detail & Related papers  (2024-12-02T21:20:02Z)
• LASeR: Learning to Adaptively Select Reward Models with Multi-Armed Bandits [56.93583799109029]
  We introduce LASeR (Learning to Adaptively Select Rewards), which iteratively trains LLMs using multiple Reward Models (RMs) Our results demonstrate that LASeR can boost iterative LLM optimization by optimizing for multiple RMs. We also verify the presence of conflicting preferences from multiple RMs that can be mitigated using LASeR.
  arXiv  Detail & Related papers  (2024-10-02T16:46:38Z)
• Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts [23.27203570485055]
  Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models with human preferences. We propose a two-stage approach to train a reward model (RM) with multi-dimensional absolute-rating data. We efficiently trained an ArmoRM with Llama-3 8B and a gating network consisting of a shallow on top of the ArmoRM.
  arXiv  Detail & Related papers  (2024-06-18T17:58:28Z)
• MetaRM: Shifted Distributions Alignment via Meta-Learning [52.94381279744458]
  Reinforcement Learning from Human Feedback (RLHF) in language model alignment is critically dependent on the capability of the reward model (RM) We introduce MetaRM, a method leveraging meta-learning to align the RM with the shifted environment distribution. Extensive experiments demonstrate that MetaRM significantly improves the RM's distinguishing ability in iterative RLHF optimization.
  arXiv  Detail & Related papers  (2024-05-01T10:43:55Z)
• WARM: On the Benefits of Weight Averaged Reward Models [63.08179139233774]
  We propose Weight Averaged Reward Models (WARM) to mitigate reward hacking. Experiments on summarization tasks, using best-of-N and RL methods, shows that WARM improves the overall quality and alignment of LLM predictions.
  arXiv  Detail & Related papers  (2024-01-22T18:27:08Z)
• Confronting Reward Model Overoptimization with Constrained RLHF [114.71591361764547]
  We show that correlation between component RMs has a significant effect on the locations of these points. Our method addresses the problem of weighting component RMs by learning dynamic weights, naturally expressed by Lagrange multipliers.
  arXiv  Detail & Related papers  (2023-10-06T16:59:17Z)
• The Trickle-down Impact of Reward (In-)consistency on RLHF [71.37987812944971]
  We show that reward inconsistency exhibits a trickle-down effect on the downstream Reinforcement Learning from Human Feedback process. We propose Contrast Instructions -- a benchmarking strategy for the consistency of RM. We show that RLHF models trained with a more consistent RM yield more useful responses.
  arXiv  Detail & Related papers  (2023-09-28T04:05:13Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.