We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is trained on a sentence of the form ''_A_ is _B_'', it will not automatically generalize to the reverse direction ''_B_ is _A_''. This is the **Reversal Curse**. For instance, if a model is trained on ''Valentina Tereshkova was the first woman to travel to space'', it will not automatically be able to answer the question, ''Who was the first woman to travel to space?''. Moreover, the likelihood of the correct answer (''Valentina Tershkova'') will not be higher than for a random name. Thus, models do not generalize a prevalent pattern in their training set: if ''_A_ is _B_'' occurs, ''_B_ is _A_'' is more likely to occur. It is worth noting, however, that if ''_A_ is _B_'' appears _in-context_, models can deduce the reverse relationship. We provide evidence for the Reversal Curse by finetuning GPT-3 and Llama-1 on fictitious statements such as ''Uriah Hawthorne is the composer of _Abyssal Melodies_'' and showing that they fail to correctly answer ''Who composed _Abyssal Melodies?_''. The Reversal Curse is robust across model sizes and model families and is not alleviated by data augmentation. We also evaluate ChatGPT (GPT-3.5 and GPT-4) on questions about real-world celebrities, such as ''Who is Tom Cruise's mother? [A: Mary Lee Pfeiffer]'' and the reverse ''Who is Mary Lee Pfeiffer's son?''. GPT-4 correctly answers questions like the former 79\% of the time, compared to 33\% for the latter. Code available at: https://github.com/lukasberglund/reversal_curse.

As language models (LMs) become more capable, it is increasingly important to align them with human preferences. However, the dominant paradigm for training Preference Models (PMs) for that purpose suffers from fundamental limitations, such as lack of transparency and scalability, along with susceptibility to overfitting the preference dataset. We propose Compositional Preference Models (CPMs), a novel PM framework that decomposes one global preference assessment into several interpretable features, obtains scalar scores for these features from a prompted LM, and aggregates these scores using a logistic regression classifier. Through these simple steps, CPMs allow to control which properties of the preference data are used to train the preference model and to build it based on features that are believed to underlie the human preference judgment. Our experiments show that CPMs not only improve generalization and are more robust to overoptimization than standard PMs, but also that best-of-n samples obtained using CPMs tend to be preferred over samples obtained using conventional PMs. Overall, our approach demonstrates the benefits of endowing PMs with priors about which features determine human preferences while relying on LM capabilities to extract those features in a scalable and robust way.

Reinforcement learning from human feedback (RLHF) is a popular technique for training high-quality AI assistants. However, RLHF may also encourage model responses that match user beliefs over truthful responses, a behavior known as sycophancy. We investigate the prevalence of sycophancy in RLHF-trained models and whether human preference judgments are responsible. We first demonstrate that five state-of-the-art AI assistants consistently exhibit sycophancy behavior across four varied free-form text-generation tasks. To understand if human preferences drive this broadly observed behavior of RLHF models, we analyze existing human preference data. We find that when a response matches a user's views, it is more likely to be preferred. Moreover, both humans and preference models (PMs) prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time. Optimizing model outputs against PMs also sometimes sacrifices truthfulness in favor of sycophancy. Overall, our results indicate that sycophancy is a general behavior of RLHF models, likely driven in part by human preference judgments favoring sycophantic responses.

Aligning language models with preferences can be posed as approximating a target distribution representing some desired behavior. Existing approaches differ both in the functional form of the target distribution and the algorithm used to approximate it. For instance, Reinforcement Learning from Human Feedback (RLHF) corresponds to minimizing a reverse KL from an implicit target distribution arising from a KL penalty in the objective. On the other hand, Generative Distributional Control (GDC) has an explicit target distribution and minimizes a forward KL from it using the Distributional Policy Gradient (DPG) algorithm. In this paper, we propose a new approach, $f$-DPG, which allows the use of any $f$-divergence to approximate any target distribution that can be evaluated. $f$-DPG unifies both frameworks (RLHF, GDC) and the approximation methods (DPG, RL with KL penalties). We show the practical benefits of various choices of divergence objectives and demonstrate that there is no universally optimal objective but that different divergences present different alignment and diversity trade-offs. We show that Jensen-Shannon divergence strikes a good balance between these objectives, and frequently outperforms forward KL divergence by a wide margin, leading to significant improvements over prior work. These distinguishing characteristics between divergences persist as the model size increases, highlighting the importance of selecting appropriate divergence objectives.

Language models (LMs) are pretrained to imitate text from large and diverse datasets that contain content that would violate human preferences if generated by an LM: falsehoods, offensive comments, personally identifiable information, low-quality or buggy code, among others. Here, we explore alternative objectives for pretraining LMs in a way that also guides them to generate text aligned with human preferences. We benchmark five objectives for pretraining with human feedback across three tasks and study how they affect the alignment and capabilities of pretrained LMs. We find a Pareto-optimal and simple approach among those we explored: conditional training, or learning distribution over tokens conditional on their human preference scores. Conditional training reduces the rate of undesirable content by up to an order of magnitude, both when generating without a prompt and with an adversarially-chosen prompt. Moreover, conditional training maintains the downstream task performance of standard LM pretraining, both before and after task-specific finetuning. Pretraining with human feedback results in much better preference satisfaction than standard LM pretraining followed by finetuning with feedback, i.e., learning and then unlearning undesirable behavior. Our results suggest that we should move beyond imitation learning when pretraining LMs and incorporate human preferences from the start of training.

Machine learning is shifting towards general-purpose pretrained generative models, trained in a self-supervised manner on large amounts of data, which can then be applied to solve a large number of tasks. However, due to their generic training methodology, these models often fail to meet some of the downstream requirements (e.g., hallucinations in abstractive summarization or style violations in code generation). This raises the important question of how to adapt pre-trained generative models to meet all requirements without destroying their general capabilities ("catastrophic forgetting"). Recent work has proposed to solve this problem by representing task-specific requirements through energy-based models (EBMs) and approximating these EBMs using distributional policy gradients (DPG). Despite its effectiveness, this approach is however limited to unconditional distributions. In this paper, we extend DPG to conditional tasks by proposing Conditional DPG (CDPG). We evaluate CDPG on four different control objectives across three tasks (translation, summarization and code generation) and two pretrained models (T5 and GPT-Neo). Our results show that fine-tuning using CDPG robustly moves these pretrained models closer towards meeting control objectives and — in contrast with baseline approaches — does not result in catastrophic forgetting.

The availability of large pre-trained models is changing the landscape of Machine Learning research and practice, moving from a "training from scratch" to a "fine-tuning'' paradigm. While in some applications the goal is to "nudge'' the pre-trained distribution towards preferred outputs, in others it is to steer it towards a different distribution over the sample space. Two main paradigms have emerged to tackle this challenge: Reward Maximization (RM) and, more recently, Distribution Matching (DM). RM applies standard Reinforcement Learning (RL) techniques, such as Policy Gradients, to gradually increase the reward signal. DM prescribes to first make explicit the target distribution that the model is fine-tuned to approximate. Here we explore the theoretical connections between the two paradigms and show that methods such as KL-control developed in the RM paradigm can also be construed as belonging to DM. We further observe that while DM differs from RM, it can suffer from similar training difficulties, such as high gradient variance. We leverage connections between the two paradigms to import the concept of baseline into DM methods. We empirically validate the benefits of adding a baseline on an array of controllable language generation tasks such as constraining topic, sentiment, and gender distributions in texts sampled from a language model. We observe superior performance in terms of constraint satisfaction, stability, and sample efficiency.

Communication is compositional if complex signals can be represented as a combination of simpler subparts. In this paper, we theoretically show that inductive biases on both the training framework and the data are needed to develop a compositional communication. Moreover, we prove that compositionality spontaneously arises in the signaling games, where agents communicate over a noisy channel. We experimentally confirm that a range of noise levels, which depends on the model and the data, indeed promotes compositionality. Finally, we provide a comprehensive study of this dependence and report results in terms of recently studied compositionality metrics: topographical similarity, conflict count, and context independence.