Language models (LMs) trained on vast quantities of text data can acquire sophisticated skills such as generating summaries, answering questions or generating code. However, they also manifest behaviors that violate human preferences, e.g., they can generate offensive content, falsehoods or perpetuate social biases. In this thesis, I explore several approaches to aligning LMs with human preferences. First, I argue that aligning LMs can be seen as Bayesian inference: conditioning a prior (base, pretrained LM) on evidence about human preferences (Chapter 2). Conditioning on human preferences can be implemented in numerous ways. In Chapter 3, I investigate the relation between two approaches to finetuning pretrained LMs using feedback given by a scoring function: reinforcement learning from human feedback (RLHF) and distribution matching. I show that RLHF can be seen as a special case of distribution matching but distributional matching is strictly more general. In chapter 4, I show how to extend the distribution matching to conditional language models. Finally, in chapter 5 I explore a different root: conditioning an LM on human preferences already during pretraining. I show that involving human feedback from the very start tends to be more effective than using it only during supervised finetuning. Overall, these results highlight the room for alignment techniques different from and complementary to RLHF.
This work identifies 18 foundational challenges in assuring the alignment and safety of large language models (LLMs). These challenges are organized into three different categories: scientific understanding of LLMs, development and deployment methods, and sociotechnical challenges. Based on the identified challenges, we pose $200+$ concrete research questions.
The proliferation of generative models, combined with pretraining on web-scale data, raises a timely question: what happens when these models are trained on their own generated outputs? Recent investigations into model-data feedback loops discovered that such loops can lead to model collapse, a phenomenon where performance progressively degrades with each model-fitting iteration until the latest model becomes useless. However, several recent papers studying model collapse assumed that new data replace old data over time rather than assuming data accumulate over time. In this paper, we compare these two settings and show that accumulating data prevents model collapse. We begin by studying an analytically tractable setup in which a sequence of linear models are fit to the previous models' predictions. Previous work showed if data are replaced, the test error increases linearly with the number of model-fitting iterations; we extend this result by proving that if data instead accumulate, the test error has a finite upper bound independent of the number of iterations. We next empirically test whether accumulating data similarly prevents model collapse by pretraining sequences of language models on text corpora. We confirm that replacing data does indeed cause model collapse, then demonstrate that accumulating data prevents model collapse; these results hold across a range of model sizes, architectures and hyperparameters. We further show that similar results hold for other deep generative models on real data: diffusion models for molecule generation and variational autoencoders for image generation. Our work provides consistent theoretical and empirical evidence that data accumulation mitigates model collapse.
Human feedback is commonly utilized to finetune AI assistants. But human feedback may also encourage model responses that match user beliefs over truthful ones, a behaviour known as sycophancy. We investigate the prevalence of sycophancy in models whose finetuning procedure made use of human feedback, and the potential role of human preference judgments in such behavior. We first demonstrate that five state-of-the-art AI assistants consistently exhibit sycophancy across four varied free-form text-generation tasks. To understand if human preferences drive this broadly observed behavior, 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 state-of-the-art AI assistants, likely driven in part by human preference judgments favoring sycophantic responses.
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. 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.
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 "Olaf Scholz was the ninth Chancellor of Germany", it will not automatically be able to answer the question, "Who was the ninth Chancellor of Germany?". Moreover, the likelihood of the correct answer ("Olaf Scholz") will not be higher than for a random name. Thus, models exhibit a basic failure of logical deduction and do not generalize a prevalent pattern in their training set (i.e. if "A is B'' occurs, "B is A" is more likely to occur). 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. This shows a failure of logical deduction that we hypothesize is caused by the Reversal Curse. Code is available at https://github.com/lukasberglund/reversal_curse.
We aim to better understand the emergence of `situational awareness' in large language models (LLMs). A model is situationally aware if it's aware that it's a model and can recognize whether it's currently in testing or deployment. Today's LLMs are tested for safety and alignment before they are deployed. An LLM could exploit situational awareness to achieve a high score on safety tests, while taking harmful actions after deployment. Situational awareness may emerge unexpectedly as a byproduct of model scaling. One way to better foresee this emergence is to run scaling experiments on abilities necessary for situational awareness. As such an ability, we propose `out-of-context reasoning' (in contrast to in-context learning). We study out-of-context reasoning experimentally. First, we finetune an LLM on a description of a test while providing no examples or demonstrations. At test time, we assess whether the model can pass the test. To our surprise, we find that LLMs succeed on this out-of-context reasoning task. Their success is sensitive to the training setup and only works when we apply data augmentation. For both GPT-3 and LLaMA-1, performance improves with model size. These findings offer a foundation for further empirical study, towards predicting and potentially controlling the emergence of situational awareness in LLMs. Code is available at: https://github.com/AsaCooperStickland/situational-awareness-evals.