REVIEW 12 cited by
Helping or Herding? Reward Model Ensembles Mitigate but do not Eliminate Reward Hacking
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Reward models play a key role in aligning language model applications towards human preferences. However, this setup creates an incentive for the language model to exploit errors in the reward model to achieve high estimated reward, a phenomenon often termed \emph{reward hacking}. A natural mitigation is to train an ensemble of reward models, aggregating over model outputs to obtain a more robust reward estimate. We explore the application of reward ensembles to alignment at both training time (through reinforcement learning) and inference time (through reranking). First, we show that reward models are \emph{underspecified}: reward models that perform similarly in-distribution can yield very different rewards when used in alignment, due to distribution shift. Second, underspecification results in overoptimization, where alignment to one reward model does not improve reward as measured by another reward model trained on the same data. Third, overoptimization is mitigated by the use of reward ensembles, and ensembles that vary by their \emph{pretraining} seeds lead to better generalization than ensembles that differ only by their \emph{fine-tuning} seeds, with both outperforming individual reward models. However, even pretrain reward ensembles do not eliminate reward hacking: we show several qualitative reward hacking phenomena that are not mitigated by ensembling because all reward models in the ensemble exhibit similar error patterns.
Forward citations
Cited by 12 Pith papers
-
What do Reward Models Memorize?
Counterfactual memorization maps show RMs misallocate capacity to easy pairs, memorize dataset artifacts, and overgeneralize length/compliance on unseen pairs.
-
Bridging the Gap in Vision Language Models in Identifying Unsafe Concepts Across Modalities
Vision-language models consistently recognize unsafe content better from text than from images, and a simplified reinforcement learning fine-tune narrows that gap.
-
Bradley-Terry and Multi-Objective Reward Modeling Are Complementary
Jointly training a Bradley-Terry preference head and a multi-attribute regression head on a shared embedding improves reward-model robustness to reward hacking and boosts multi-objective scoring performance.
-
Best-of-N through the Smoothing Lens: KL Divergence and Regret Analysis
Smoothed Best-of-N has finite-sample KL and regret bounds under imperfect reward models, and tuning its temperature can make its regret bound beat hard Best-of-N in the overoptimization regime.
-
RewardAnything: Generalizable Principle-Following Reward Models
RewardAnything follows natural-language reward principles at inference time and, with the new RABench benchmark, demonstrates that principle-conditioned listwise training beats fixed-preference reward models on held-o...
-
Learning a Pessimistic Reward Model in RLHF
Pessimistic fine-tuning of reward models against rejection-sampling policies lets RLHF agents optimize greedily without KL regularization and still avoid reward hacking.
-
T2I-Eval-R1: Reinforcement Learning-Driven Reasoning for Interpretable Text-to-Image Evaluation
Training an open-source 7B multimodal LLM with GRPO and a continuous reward, using only coarse quality scores, produces human-aligned interpretable text-to-image evaluation scores and rationales.
-
Think-RM: Enabling Long-Horizon Reasoning in Generative Reward Models
A generative reward model trained with long chain-of-thought and rule-based RL outperforms standard and vertically scaled reward baselines on RM-Bench and RewardBench.
-
Activation Reward Models for Few-Shot Model Alignment
Mean attention-head activations from a few labeled examples, injected into selected heads, turn a frozen vision-language model into a few-shot reward model that beats prompting and scoring baselines and a new reward-h...
-
Token-level Accept or Reject: A Micro Alignment Approach for Large Language Models
MARA aligns LLMs with human preferences by training a 4M-parameter MLP to accept or reject candidate tokens, avoiding full-model fine-tuning, with measured gains based on the same reward models used in training.
-
Reward Modeling for Reinforcement Learning-Based LLM Reasoning: Design, Challenges, and Evaluation
A taxonomy-driven survey arguing that reward design is the central mechanism shaping reliable LLM reasoning, with maps of reward paradigms, reward-hacking failure modes, and benchmark pitfalls.
-
Towards Reliable, Uncertainty-Aware Alignment
Variance-aware RLHF adds a variance-weighted KL penalty to PPO and reduces reward variance and the risk of underperforming the reference policy in the paper's experiments.
Discussion (0). Continue with ORCID to comment.