Rational Metareasoning for Large Language Models

Being prompted to engage in reasoning has emerged as a core technique for using large language models (LLMs), deploying additional inference-time compute to improve task performance. However, as LLMs increase in both size and adoption, inference costs are correspondingly becoming increasingly burdensome. How, then, might we optimize reasoning's cost-performance tradeoff? This work introduces a novel approach based on computational models of metareasoning used in cognitive science, training LLMs to selectively use intermediate reasoning steps only when necessary. We first develop a reward function that incorporates the Value of Computation by penalizing unnecessary reasoning, then use this reward function with Expert Iteration to train the LLM. Compared to few-shot chain-of-thought prompting and STaR, our method significantly reduces inference costs (20-37\% fewer tokens generated across three models) while maintaining task performance across diverse datasets.

When a language model is optimized for reasoning, does it still show embers of autoregression? An analysis of OpenAI o1

In "Embers of Autoregression" (McCoy et al., 2023), we showed that several large language models (LLMs) have some important limitations that are attributable to their origins in next-word prediction. Here we investigate whether these issues persist with o1, a new system from OpenAI that differs from previous LLMs in that it is optimized for reasoning. We find that o1 substantially outperforms previous LLMs in many cases, with particularly large improvements on rare variants of common tasks (e.g., forming acronyms from the second letter of each word in a list, rather than the first letter). Despite these quantitative improvements, however, o1 still displays the same qualitative trends that we observed in previous systems. Specifically, o1 - like previous LLMs - is sensitive to the probability of examples and tasks, performing better and requiring fewer "thinking tokens" in high-probability settings than in low-probability ones. These results show that optimizing a language model for reasoning can mitigate but might not fully overcome the language model's probability sensitivity.

Capturing the Complexity of Human Strategic Decision-Making with Machine Learning

Understanding how people behave in strategic settings--where they make decisions based on their expectations about the behavior of others--is a long-standing problem in the behavioral sciences. We conduct the largest study to date of strategic decision-making in the context of initial play in two-player matrix games, analyzing over 90,000 human decisions across more than 2,400 procedurally generated games that span a much wider space than previous datasets. We show that a deep neural network trained on these data predicts people's choices better than leading theories of strategic behavior, indicating that there is systematic variation that is not explained by those theories. We then modify the network to produce a new, interpretable behavioral model, revealing what the original network learned about people: their ability to optimally respond and their capacity to reason about others are dependent on the complexity of individual games. This context-dependence is critical in explaining deviations from the rational Nash equilibrium, response times, and uncertainty in strategic decisions. More broadly, our results demonstrate how machine learning can be applied beyond prediction to further help generate novel explanations of complex human behavior.

Building Machines that Learn and Think with People

What do we want from machine intelligence? We envision machines that are not just tools for thought, but partners in thought: reasonable, insightful, knowledgeable, reliable, and trustworthy systems that think with us. Current artificial intelligence (AI) systems satisfy some of these criteria, some of the time. In this Perspective, we show how the science of collaborative cognition can be put to work to engineer systems that really can be called ``thought partners,'' systems built to meet our expectations and complement our limitations. We lay out several modes of collaborative thought in which humans and AI thought partners can engage and propose desiderata for human-compatible thought partnerships. Drawing on motifs from computational cognitive science, we motivate an alternative scaling path for the design of thought partners and ecosystems around their use through a Bayesian lens, whereby the partners we construct actively build and reason over models of the human and world.

Deciphering the Factors Influencing the Efficacy of Chain-of-Thought: Probability, Memorization, and Noisy Reasoning

Chain-of-Thought (CoT) prompting has been shown to enhance the multi-step reasoning capabilities of Large Language Models (LLMs). However, debates persist about whether LLMs exhibit abstract generalization or rely on shallow heuristics when given CoT prompts. To understand the factors influencing CoT reasoning we provide a detailed case study of the symbolic reasoning task of decoding shift ciphers, where letters are shifted forward some number of steps in the alphabet. GPT-4 achieves zero accuracy on most shift ciphers with standard prompting, but with CoT its accuracy improves to an average of 32%. By focusing on a single relatively simple task, we are able to identify three factors that systematically affect CoT performance: the probability of the task's expected output (probability), what the model has implicitly learned during pre-training (memorization), and the number of intermediate operations involved in reasoning (noisy reasoning). We show that these factors can drastically influence the task accuracy; e.g., varying the output's probability of occurrence can shift accuracy from 26% to 70%. We also demonstrate that it is essential for the model to explicitly produce intermediate steps as output that can be conditioned on to increase the probability of the correct answer. Our experiments indicate that as long as the model does so, the validity of the demonstrations in the prompt does not matter. Overall, we conclude that CoT prompting performance reflects both memorization and a probabilistic version of genuine reasoning.

Large Language Models Assume People are More Rational than We Really are

In order for AI systems to communicate effectively with people, they must understand how we make decisions. However, people's decisions are not always rational, so the implicit internal models of human decision-making in Large Language Models (LLMs) must account for this. Previous empirical evidence seems to suggest that these implicit models are accurate -- LLMs offer believable proxies of human behavior, acting how we expect humans would in everyday interactions. However, by comparing LLM behavior and predictions to a large dataset of human decisions, we find that this is actually not the case: when both simulating and predicting people's choices, a suite of cutting-edge LLMs (GPT-4o & 4-Turbo, Llama-3-8B & 70B, Claude 3 Opus) assume that people are more rational than we really are. Specifically, these models deviate from human behavior and align more closely with a classic model of rational choice -- expected value theory. Interestingly, people also tend to assume that other people are rational when interpreting their behavior. As a consequence, when we compare the inferences that LLMs and people draw from the decisions of others using another psychological dataset, we find that these inferences are highly correlated. Thus, the implicit decision-making models of LLMs appear to be aligned with the human expectation that other people will act rationally, rather than with how people actually act.

Representational Alignment Supports Effective Machine Teaching

A good teacher should not only be knowledgeable; but should be able to communicate in a way that the student understands -- to share the student's representation of the world. In this work, we integrate insights from machine teaching and pragmatic communication with the burgeoning literature on representational alignment to characterize a utility curve defining a relationship between representational alignment and teacher capability for promoting student learning. To explore the characteristics of this utility curve, we design a supervised learning environment that disentangles representational alignment from teacher accuracy. We conduct extensive computational experiments with machines teaching machines, complemented by a series of experiments in which machines teach humans. Drawing on our findings that improved representational alignment with a student improves student learning outcomes (i.e., task accuracy), we design a classroom matching procedure that assigns students to teachers based on the utility curve. If we are to design effective machine teachers, it is not enough to build teachers that are accurate -- we want teachers that can align, representationally, to their students too.

What Should Embeddings Embed? Autoregressive Models Represent Latent Generating Distributions

Autoregressive language models have demonstrated a remarkable ability to extract latent structure from text. The embeddings from large language models have been shown to capture aspects of the syntax and semantics of language. But what {\em should} embeddings represent? We connect the autoregressive prediction objective to the idea of constructing predictive sufficient statistics to summarize the information contained in a sequence of observations, and use this connection to identify three settings where the optimal content of embeddings can be identified: independent identically distributed data, where the embedding should capture the sufficient statistics of the data; latent state models, where the embedding should encode the posterior distribution over states given the data; and discrete hypothesis spaces, where the embedding should reflect the posterior distribution over hypotheses given the data. We then conduct empirical probing studies to show that transformers encode these three kinds of latent generating distributions, and that they perform well in out-of-distribution cases and without token memorization in these settings.

Analyzing the Benefits of Prototypes for Semi-Supervised Category Learning

Categories can be represented at different levels of abstraction, from prototypes focused on the most typical members to remembering all observed exemplars of the category. These representations have been explored in the context of supervised learning, where stimuli are presented with known category labels. We examine the benefits of prototype-based representations in a less-studied domain: semi-supervised learning, where agents must form unsupervised representations of stimuli before receiving category labels. We study this problem in a Bayesian unsupervised learning model called a variational auto-encoder, and we draw on recent advances in machine learning to implement a prior that encourages the model to use abstract prototypes to represent data. We apply this approach to image datasets and show that forming prototypes can improve semi-supervised category learning. Additionally, we study the latent embeddings of the models and show that these prototypes allow the models to form clustered representations without supervision, contributing to their success in downstream categorization performance.

Eliciting the Priors of Large Language Models using Iterated In-Context Learning

As Large Language Models (LLMs) are increasingly deployed in real-world settings, understanding the knowledge they implicitly use when making decisions is critical. One way to capture this knowledge is in the form of Bayesian prior distributions. We develop a prompt-based workflow for eliciting prior distributions from LLMs. Our approach is based on iterated learning, a Markov chain Monte Carlo method in which successive inferences are chained in a way that supports sampling from the prior distribution. We validated our method in settings where iterated learning has previously been used to estimate the priors of human participants -- causal learning, proportion estimation, and predicting everyday quantities. We found that priors elicited from GPT-4 qualitatively align with human priors in these settings. We then used the same method to elicit priors from GPT-4 for a variety of speculative events, such as the timing of the development of superhuman AI.