Reward Model Overoptimisation in Iterated RLHF

• URL: http://arxiv.org/abs/2505.18126v1
• Date: Fri, 23 May 2025 17:36:13 GMT
• Title: Reward Model Overoptimisation in Iterated RLHF
• Authors: Lorenz Wolf, Robert Kirk, Mirco Musolesi,
• Abstract summary: Reinforcement learning from human feedback (RLHF) is a widely used method for aligning large language models with human preferences.<n> RLHF often suffers from reward model overoptimisation, in which models overfit to the reward function.<n>We present the first comprehensive study of overoptimisation in iterated RLHF.
• Score: 3.6701456157280052
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Reinforcement learning from human feedback (RLHF) is a widely used method for aligning large language models with human preferences. However, RLHF often suffers from reward model overoptimisation, in which models overfit to the reward function, resulting in non-generalisable policies that exploit the idiosyncrasies and peculiarities of the reward function. A common mitigation is iterated RLHF, in which reward models are repeatedly retrained with updated human feedback and policies are re-optimised. Despite its increasing adoption, the dynamics of overoptimisation in this setting remain poorly understood. In this work, we present the first comprehensive study of overoptimisation in iterated RLHF. We systematically analyse key design choices - how reward model training data is transferred across iterations, which reward function is used for optimisation, and how policies are initialised. Using the controlled AlpacaFarm benchmark, we observe that overoptimisation tends to decrease over successive iterations, as reward models increasingly approximate ground-truth preferences. However, performance gains diminish over time, and while reinitialising from the base policy is robust, it limits optimisation flexibility. Other initialisation strategies often fail to recover from early overoptimisation. These findings offer actionable insights for building more stable and generalisable RLHF pipelines.

Related papers

• Mitigating Preference Hacking in Policy Optimization with Pessimism [32.58012040199723]
  This work tackles the problem of overoptimization in reinforcement learning from human feedback (RLHF)<n>RLHF relies on reward or preference models trained on emphfixed preference datasets, and these models are unreliable when evaluated outside the support of this preference data.<n>We propose novel, pessimistic objectives for RLHF which are provably robust to overoptimization through the use of pessimism in the face of uncertainty.
  arXiv  Detail & Related papers  (2025-03-10T00:13:19Z)
• PILAF: Optimal Human Preference Sampling for Reward Modeling [14.336058926701432]
  We propose Policy-Interpolated Learning for Aligned Feedback (PILAF), a novel response sampling strategy for preference labeling.<n>PILAF explicitly aligns preference learning with maximizing the underlying oracle reward.
  arXiv  Detail & Related papers  (2025-02-06T18:09:00Z)
• REINFORCE++: An Efficient RLHF Algorithm with Robustness to Both Prompt and Reward Models [9.950083479263293]
  REINFORCE++ is a novel approach that removes the critic model while using the global advantage normalization.<n>It exhibits robust performance across various reward models without requiring prompt set truncation.<n>It achieves superior generalization in both RLHF and long chain-of-thought settings.
  arXiv  Detail & Related papers  (2025-01-04T02:08:06Z)
• Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms [50.808123629394245]
  Direct Alignment Algorithms (DDAs) like Direct Preference Optimization have emerged as alternatives to the classical RLHF pipeline. This work formulates and formalizes the reward over-optimization or hacking problem for DAAs and explores its consequences across objectives, training regimes, and model scales.
  arXiv  Detail & Related papers  (2024-06-05T03:41:37Z)
• Fine-Tuning Language Models with Reward Learning on Policy [68.70065254564642]
  Reinforcement learning from human feedback (RLHF) has emerged as an effective approach to aligning large language models (LLMs) to human preferences. Despite its popularity, (fixed) reward models may suffer from inaccurate off-distribution. We propose reward learning on policy (RLP), an unsupervised framework that refines a reward model using policy samples to keep it on-distribution.
  arXiv  Detail & Related papers  (2024-03-28T10:02:10Z)
• Improving Reinforcement Learning from Human Feedback with Efficient Reward Model Ensemble [67.4269821365504]
  Reinforcement Learning from Human Feedback (RLHF) is a widely adopted approach for aligning large language models with human values. However, RLHF relies on a reward model that is trained with a limited amount of human preference data. We contribute a reward ensemble method that allows the reward model to make more accurate predictions.
  arXiv  Detail & Related papers  (2024-01-30T00:17:37Z)
• Iterative Data Smoothing: Mitigating Reward Overfitting and Overoptimization in RLHF [79.98542868281471]
  Reinforcement Learning from Human Feedback (RLHF) is a technique that aligns language models closely with human-centric values. It is observed that the performance of the reward model degrades after one epoch of training, and optimizing too much against the learned reward model eventually hinders the true objective. This paper delves into these issues, leveraging the theoretical insights to design improved reward learning algorithm termed 'Iterative Data Smoothing' (IDS)
  arXiv  Detail & Related papers  (2024-01-29T17:43:42Z)
• Contrastive Preference Learning: Learning from Human Feedback without RL [71.77024922527642]
  We introduce Contrastive Preference Learning (CPL), an algorithm for learning optimal policies from preferences without learning reward functions. CPL is fully off-policy, uses only a simple contrastive objective, and can be applied to arbitrary MDPs.
  arXiv  Detail & Related papers  (2023-10-20T16:37:56Z)
• Direct Preference Optimization: Your Language Model is Secretly a Reward Model [119.65409513119963]
  We introduce a new parameterization of the reward model in RLHF that enables extraction of the corresponding optimal policy in closed form. The resulting algorithm, which we call Direct Preference Optimization (DPO), is stable, performant, and computationally lightweight. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods.
  arXiv  Detail & Related papers  (2023-05-29T17:57:46Z)

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.