RLHS: Mitigating Misalignment in RLHF with Hindsight Simulation

• URL: http://arxiv.org/abs/2501.08617v2
• Date: Mon, 10 Feb 2025 21:17:01 GMT
• Title: RLHS: Mitigating Misalignment in RLHF with Hindsight Simulation
• Authors: Kaiqu Liang, Haimin Hu, Ryan Liu, Thomas L. Griffiths, Jaime Fernández Fisac,
• Abstract summary: We show that Reinforcement Learning from Human Feedback can cause severe, systematic misalignment. We introduce Reinforcement Learning from Hindsight Simulation (RLHS), which presents plausible simulated outcomes to evaluators before eliciting feedback. We evaluate post-hoc on the TruthfulQA benchmark and find that, even after single-task fine-tuning, both RLHF misalignment and RLHS alignment carry over to substantially different settings.
• Score: 3.998312409829935
• License:
• Abstract: While Reinforcement Learning from Human Feedback (RLHF) has shown promise in aligning generative AI, we present empirical evidence that it can also cause severe, systematic misalignment. We hypothesize that this stems from evaluator feedback depending on downstream outcome predictions (foresight) that can be influenced by the AI's output, inducing Goodhart's law dynamics. Conversely, our theoretical analysis shows that conditioning evaluator feedback on downstream observations (hindsight) inhibits this effect by decoupling the alignment signal from potentially compromised predictions-crucially, the result holds even if the observed outcomes are sampled from the AI's own world model. Building on this insight, we introduce Reinforcement Learning from Hindsight Simulation (RLHS), which presents plausible simulated outcomes to evaluators before eliciting feedback. We demonstrate RLHS on online (PPO) and offline (DPO) large language model fine-tuning, obtaining superior alignment over RLHF in controlled consultancy-type experiments and user studies. We evaluate post-hoc on the TruthfulQA benchmark and find that, even after single-task fine-tuning, both RLHF misalignment and RLHS alignment carry over to substantially different settings.

Related papers

• Solving the Inverse Alignment Problem for Efficient RLHF [0.0]
  We define the 'inverse alignment problem' in language model training. We investigate whether repeatedly fine-tuning a reward model on subsets of the offline preference dataset aligned with a periodically frozen policy improves upon vanilla RLHF.
  arXiv  Detail & Related papers  (2024-12-13T19:47:38Z)
• UDQL: Bridging The Gap between MSE Loss and The Optimal Value Function in Offline Reinforcement Learning [10.593924216046977]
  We first theoretically analyze overestimation phenomenon led by MSE and provide the theoretical upper bound of the overestimated error. At last, we propose the offline RL algorithm based on underestimated operator and diffusion policy model.
  arXiv  Detail & Related papers  (2024-06-05T14:37:42Z)
• Assessing the Impact of Distribution Shift on Reinforcement Learning Performance [0.0]
  Reinforcement learning (RL) faces its own set of unique challenges. Comparison of point estimates, and plots that show successful convergence to the optimal policy during training, may obfuscate overfitting or dependence on the experimental setup. We propose a set of evaluation methods that measure the robustness of RL algorithms under distribution shifts.
  arXiv  Detail & Related papers  (2024-02-05T23:50:55Z)
• Beyond Imitation: Leveraging Fine-grained Quality Signals for Alignment [105.34140537748546]
  We propose an improved alignment approach named FIGA. Different from prior methods, we incorporate fine-grained quality signals that are derived by contrasting good and bad responses. Our approach has made two major contributions. Firstly, we curate a refined alignment dataset that pairs initial responses and the corresponding revised ones. Secondly, we devise a new loss function can leverage fine-grained quality signals to instruct the learning of LLMs for alignment.
  arXiv  Detail & Related papers  (2023-11-07T15:36:40Z)
• Understanding, Predicting and Better Resolving Q-Value Divergence in Offline-RL [86.0987896274354]
  We first identify a fundamental pattern, self-excitation, as the primary cause of Q-value estimation divergence in offline RL. We then propose a novel Self-Excite Eigenvalue Measure (SEEM) metric to measure the evolving property of Q-network at training. For the first time, our theory can reliably decide whether the training will diverge at an early stage.
  arXiv  Detail & Related papers  (2023-10-06T17:57:44Z)
• Mitigating the Alignment Tax of RLHF [76.4300447532456]
  aligning LLMs under Reinforcement Learning with Human Feedback can lead to forgetting pretrained abilities, also known as the alignment tax. We propose model averaging to maximize alignment performance while incurring minimal alignment tax. We validate HMA's performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B.
  arXiv  Detail & Related papers  (2023-09-12T14:16:54Z)
• 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)
• Supervised Advantage Actor-Critic for Recommender Systems [76.7066594130961]
  We propose negative sampling strategy for training the RL component and combine it with supervised sequential learning. Based on sampled (negative) actions (items), we can calculate the "advantage" of a positive action over the average case. We instantiate SNQN and SA2C with four state-of-the-art sequential recommendation models and conduct experiments on two real-world datasets.
  arXiv  Detail & Related papers  (2021-11-05T12:51:15Z)
• Causal Inference Q-Network: Toward Resilient Reinforcement Learning [57.96312207429202]
  We consider a resilient DRL framework with observational interferences. Under this framework, we propose a causal inference based DRL algorithm called causal inference Q-network (CIQ) Our experimental results show that the proposed CIQ method could achieve higher performance and more resilience against observational interferences.
  arXiv  Detail & Related papers  (2021-02-18T23:50:20Z)

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.