Sparse autoencoders provide a promising unsupervised approach for extracting interpretable features from a language model by reconstructing activations from a sparse bottleneck layer. Since language models learn many concepts, autoencoders need to be very large to recover all relevant features. However, studying the properties of autoencoder scaling is difficult due to the need to balance reconstruction and sparsity objectives and the presence of dead latents. We propose using k-sparse autoencoders [Makhzani and Frey, 2013] to directly control sparsity, simplifying tuning and improving the reconstruction-sparsity frontier. Additionally, we find modifications that result in few dead latents, even at the largest scales we tried. Using these techniques, we find clean scaling laws with respect to autoencoder size and sparsity. We also introduce several new metrics for evaluating feature quality based on the recovery of hypothesized features, the explainability of activation patterns, and the sparsity of downstream effects. These metrics all generally improve with autoencoder size. To demonstrate the scalability of our approach, we train a 16 million latent autoencoder on GPT-4 activations for 40 billion tokens. We release training code and autoencoders for open-source models, as well as a visualizer.

Large language models often possess latent capabilities that lie dormant unless explicitly elicited, or surfaced, through fine-tuning or prompt engineering. Predicting, assessing, and understanding these latent capabilities pose significant challenges in the development of effective, safe AI systems. In this work, we recast elicitation as an information-constrained fine-tuning problem and empirically characterize upper bounds on the minimal number of parameters needed to achieve specific task performances. We find that training as few as 10–100 randomly chosen parameters—several orders of magnitude fewer than state-of-the-art parameter-efficient methods—can recover up to 50\% of the performance gap between pretrained-only and full fine-tuned models, and 1,000s to 10,000s of parameters can recover 95\% of this performance gap. We show that a logistic curve fits the relationship between the number of trained parameters and model performance gap recovery. This scaling generalizes across task formats and domains, as well as model sizes and families, extending to reasoning models and remaining robust to increases in inference compute. To help explain this behavior, we consider a simplified picture of elicitation via fine-tuning where each trainable parameter serves as an encoding mechanism for accessing task-specific knowledge. We observe a relationship between the number of trained parameters and how efficiently relevant model capabilities can be accessed and elicited, offering a potential route to distinguish elicitation from teaching.

In recent years, large language models have greatly improved in their ability to perform complex multi-step reasoning. However, even state-of-the-art models still regularly produce logical mistakes. To train more reliable models, we can turn either to outcome supervision, which provides feedback for a final result, or process supervision, which provides feedback for each intermediate reasoning step. Given the importance of training reliable models, and given the high cost of human feedback, it is important to carefully compare the both methods. Recent work has already begun this comparison, but many questions still remain. We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision. To support related research, we also release PRM800K, the complete dataset of 800,000 step-level human feedback labels used to train our best reward model.

Widely used alignment techniques, such as reinforcement learning from human feedback (RLHF), rely on the ability of humans to supervise model behavior—for example, to evaluate whether a model faithfully followed instructions or generated safe outputs. However, future superhuman models will behave in complex ways too difficult for humans to reliably evaluate; humans will only be able to weakly supervise superhuman models. We study an analogy to this problem: can weak model supervision elicit the full capabilities of a much stronger model? We test this using a range of pretrained language models in the GPT-4 family on natural language processing (NLP), chess, and reward modeling tasks. We find that when we naively finetune strong pretrained models on labels generated by a weak model, they consistently perform better than their weak supervisors, a phenomenon we call weak-to-strong generalization. However, we are still far from recovering the full capabilities of strong models with naive finetuning alone, suggesting that techniques like RLHF may scale poorly to superhuman models without further work. We find that simple methods can often significantly improve weak-to-strong generalization: for example, when finetuning GPT-4 with a GPT-2-level supervisor and an auxiliary confidence loss, we can recover close to GPT-3.5-level performance on NLP tasks. Our results suggest that it is feasible to make empirical progress today on a fundamental challenge of aligning superhuman models.

Making language models bigger does not inherently make them better at following a user's intent. For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are not aligned with their users. In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by fine-tuning with human feedback. Starting with a set of labeler-written prompts and prompts submitted through a language model API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback. We call the resulting models InstructGPT. In human evaluations on our prompt distribution, outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters. Moreover, InstructGPT models show improvements in truthfulness and reductions in toxic output generation while having minimal performance regressions on public NLP datasets. Even though InstructGPT still makes simple mistakes, our results show that fine-tuning with human feedback is a promising direction for aligning language models with human intent.

We seek to align agent behavior with a user’s objectives in a reinforcement learning setting with unknown dynamics, an unknown reward function, and unknown unsafe states. The user knows the rewards and unsafe states, but querying the user is expensive. We propose an algorithm that safely and efficiently learns a model of the user’s reward function by posing ’what if?’ questions about hypothetical agent behavior. We start with a generative model of initial states and a forward dynamics model trained on off-policy data. Our method uses these models to synthesize hypothetical behaviors, asks the user to label the behaviors with rewards, and trains a neural network to predict the rewards. The key idea is to actively synthesize the hypothetical behaviors from scratch by maximizing tractable proxies for the value of information, without interacting with the environment. We call this method reward query synthesis via trajectory optimization (ReQueST). We evaluate ReQueST with simulated users on a state-based 2D navigation task and the image-based Car Racing video game. The results show that ReQueST significantly outperforms prior methods in learning reward models that transfer to new environments with different initial state distributions. Moreover, ReQueST safely trains the reward model to detect unsafe states, and corrects reward hacking before deploying the agent.

In some agent designs like inverse reinforcement learning an agent needs to learn its own reward function. Learning the reward function and optimising for it are typically two different processes, usually performed at different stages. We consider a continual (``one life'') learning approach where the agent both learns the reward function and optimises for it at the same time. We show that this comes with a number of pitfalls, such as deliberately manipulating the learning process in one direction, refusing to learn, ``learning'' facts already known to the agent, and making decisions that are strictly dominated (for all relevant reward functions). We formally introduce two desirable properties: the first is `unriggability', which prevents the agent from steering the learning process in the direction of a reward function that is easier to optimise. The second is `uninfluenceability', whereby the reward-function learning process operates by learning facts about the environment. We show that an uninfluenceable process is automatically unriggable, and if the set of possible environments is sufficiently large, the converse is true too.

To solve complex real-world problems with reinforcement learning, we cannot rely on manually specified reward functions. Instead, we need humans to communicate an objective to the agent directly. In this work, we combine two approaches to this problem: learning from expert demonstrations and learning from trajectory preferences. We use both to train a deep neural network to model the reward function and use its predicted reward to train an DQN-based deep reinforcement learning agent on 9 Atari games. Our approach beats the imitation learning baseline in 7 games and achieves strictly superhuman performance on 2 games. Additionally, we investigate the fit of the reward model, present some reward hacking problems, and study the effects of noise in the human labels.

Many state-of-the-art reinforcement learning (RL) algorithms typically assume that the environment is an ergodic Markov Decision Process (MDP). In contrast, the field of universal reinforcement learning (URL) is concerned with algorithms that make as few assumptions as possible about the environment. The universal Bayesian agent AIXI and a family of related URL algorithms have been developed in this setting. While numerous theoretical optimality results have been proven for these agents, there has been no empirical investigation of their behavior to date. We present a short and accessible survey of these URL algorithms under a unified notation and framework, along with results of some experiments that qualitatively illustrate some properties of the resulting policies, and their relative performance on partially-observable gridworld environments. We also present an open- source reference implementation of the algorithms which we hope will facilitate further understanding of, and experimentation with, these ideas.

We discuss some recent results on Thompson sampling for nonparametric reinforcement learning in countable classes of general stochastic environments. These environments can be non-Markovian, non-ergodic, and partially observable. We show that Thompson sampling learns the environment class in the sense that (1) asymptotically its value converges in mean to the optimal value and (2) given a recoverability assumption regret is sublinear. We conclude with a discussion about optimality in reinforcement learning.

For sophisticated reinforcement learning (RL) systems to interact usefully with real-world environments, we need to communicate complex goals to these systems. In this work, we explore goals defined in terms of (non-expert) human preferences between pairs of trajectory segments. Our approach separates learning the goal from learning the behavior to achieve it. We show that this approach can effectively solve complex RL tasks without access to the reward function, including Atari games and simulated robot locomotion, while providing feedback on about 0.1% of our agent's interactions with the environment. This reduces the cost of human oversight far enough that it can be practically applied to state-of-the-art RL systems. To demonstrate the flexibility of our approach, we show that we can successfully train complex novel behaviors with about an hour of human time. These behaviors and environments are considerably more complex than any which have been previously learned from human feedback.

This paper establishes for the first time the predictive performance of speed priors and their computational complexity. A speed prior is essentially a probability distribution that puts low probability on strings that are not efficiently computable. We propose a variant to the original speed prior (Schmidhuber, 2002), and show that our prior can predict sequences drawn from probability measures that are estimable in polynomial time. Our speed prior is computable in doubly-exponential time, but not in polynomial time. On a polynomial time computable sequence our speed prior is computable in exponential time. We show that Schmidhuber’s speed prior has better complexity under the same conditions; however, the question of its predictive properties remains open.

A big open question of algorithmic information theory is the choice of the universal Turing machine (UTM). For Kolmogorov complexity and Solomonoff induction we have invariance theorems: the choice of the UTM changes bounds only by a constant. For the universally intelligent agent AIXI (Hutter, 2005) no invariance theorem is known. Our results are entirely negative: we discuss cases in which unlucky or adversarial choices of the UTM cause AIXI to misbehave drastically. We show that Legg-Hutter intelligence and thus balanced Pareto optimality is entirely subjective, and that every policy is Pareto optimal in the class of all computable environments. This undermines all existing optimality properties for AIXI. While it may still serve as a gold standard for AI, our results imply that AIXI is a \emphrelative theory, dependent on the choice of the UTM.

Ultimate Automizer is a software verification tool that is able to analyze reachability of an error label, memory safety, and termination of C programs. For all three tasks, our tool follows an automatabased approach where interpolation is used to compute proofs for traces. The interpolants are generated via a new scheme that requires only the post operator, unsatisfiable cores and live variable analysis. This new scheme enables our tool to use the SMT theory of arrays in combination with interpolation.