Transforming and Combining Rewards for Aligning Large Language Models

• URL: http://arxiv.org/abs/2402.00742v2
• Date: Fri, 19 Jul 2024 05:12:15 GMT
• Title: Transforming and Combining Rewards for Aligning Large Language Models
• Authors: Zihao Wang, Chirag Nagpal, Jonathan Berant, Jacob Eisenstein, Alex D'Amour, Sanmi Koyejo, Victor Veitch,
• Abstract summary: A common approach for aligning language models to human preferences is to first learn a reward model from preference data, and then use this reward model to update the language model. We use a log-sigmoid function to transform rewards learned from Bradley-Terry preference models. Experiments aligning language models to be both helpful and harmless using RLHF show substantial improvements over the baseline (non-transformed) approach.
• Score: 69.44634017612798
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: A common approach for aligning language models to human preferences is to first learn a reward model from preference data, and then use this reward model to update the language model. We study two closely related problems that arise in this approach. First, any monotone transformation of the reward model preserves preference ranking; is there a choice that is ``better'' than others? Second, we often wish to align language models to multiple properties: how should we combine multiple reward models? Using a probabilistic interpretation of the alignment procedure, we identify a natural choice for transformation for (the common case of) rewards learned from Bradley-Terry preference models. The derived transformation is straightforward: we apply a log-sigmoid function to the centered rewards, a method we term ``LSC-transformation'' (log-sigmoid-centered transformation). This transformation has two important properties. First, it emphasizes improving poorly-performing outputs, rather than outputs that already score well. This mitigates both underfitting (where some prompts are not improved) and reward hacking (where the model learns to exploit misspecification of the reward model). Second, it enables principled aggregation of rewards by linking summation to logical conjunction: the sum of transformed rewards corresponds to the probability that the output is ``good'' in all measured properties, in a sense we make precise. Experiments aligning language models to be both helpful and harmless using RLHF show substantial improvements over the baseline (non-transformed) approach.

Related papers

• Rethinking Bradley-Terry Models in Preference-Based Reward Modeling: Foundations, Theory, and Alternatives [14.401557416713315]
  We revisit the foundations of using Bradley-Terry (BT) models in reward modeling. We argue that the BT model is not a necessary choice from the perspective of downstream optimization. We propose a simple and straightforward upper-bound algorithm, compatible with off-the-shelf binary classifiers.
  arXiv  Detail & Related papers  (2024-11-07T18:57:03Z)
• A Probability--Quality Trade-off in Aligned Language Models and its Relation to Sampling Adaptors [50.046717886067555]
  We show that when sampling corpora from an aligned language model, there exists a trade-off between the strings' average reward and average log-likelihood. We provide a formal treatment of this phenomenon and demonstrate how a choice of sampling adaptor allows for a selection of how much likelihood we exchange for the reward.
  arXiv  Detail & Related papers  (2024-06-14T17:38:21Z)
• RewardBench: Evaluating Reward Models for Language Modeling [100.28366840977966]
  We present RewardBench, a benchmark dataset and code-base for evaluation of reward models. The dataset is a collection of prompt-chosen-rejected trios spanning chat, reasoning, and safety. On the RewardBench leaderboard, we evaluate reward models trained with a variety of methods.
  arXiv  Detail & Related papers  (2024-03-20T17:49:54Z)
• Secrets of RLHF in Large Language Models Part II: Reward Modeling [134.97964938009588]
  We introduce a series of novel methods to mitigate the influence of incorrect and ambiguous preferences in the dataset. We also introduce contrastive learning to enhance the ability of reward models to distinguish between chosen and rejected responses.
  arXiv  Detail & Related papers  (2024-01-11T17:56:59Z)
• Quark: Controllable Text Generation with Reinforced Unlearning [68.07749519374089]
  Large-scale language models often learn behaviors that are misaligned with user expectations. We introduce Quantized Reward Konditioning (Quark), an algorithm for optimizing a reward function that quantifies an (un)wanted property. For unlearning toxicity, negative sentiment, and repetition, our experiments show that Quark outperforms both strong baselines and state-of-the-art reinforcement learning methods.
  arXiv  Detail & Related papers  (2022-05-26T21:11:51Z)
• A Sparsity-promoting Dictionary Model for Variational Autoencoders [16.61511959679188]
  Structuring the latent space in deep generative models is important to yield more expressive models and interpretable representations. We propose a simple yet effective methodology to structure the latent space via a sparsity-promoting dictionary model.
  arXiv  Detail & Related papers  (2022-03-29T17:13:11Z)
• AvgOut: A Simple Output-Probability Measure to Eliminate Dull Responses [97.50616524350123]
  We build dialogue models that are dynamically aware of what utterances or tokens are dull without any feature-engineering. The first model, MinAvgOut, directly maximizes the diversity score through the output distributions of each batch. The second model, Label Fine-Tuning (LFT), prepends to the source sequence a label continuously scaled by the diversity score to control the diversity level. The third model, RL, adopts Reinforcement Learning and treats the diversity score as a reward signal.
  arXiv  Detail & Related papers  (2020-01-15T18:32:06Z)

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.