Ben Prystawski

Humans have a powerful and mysterious capacity to reason. Working through a set of mental steps enables us to make inferences we would not be capable of making directly even though we get no additional data from the world. Similarly, when large language models generate intermediate steps (a chain of thought) before answering a question, they often produce better answers than they would directly. We investigate why and how chain-of-thought reasoning is useful in language models, testing the hypothesis that reasoning is effective when training data consists of overlapping local clusters of variables that influence each other strongly. These training conditions enable the chaining of accurate local inferences to estimate relationships between variables that were not seen together in training. We prove that there will exist a "reasoning gap", where reasoning through intermediate variables reduces bias, for the simple case of an autoregressive density estimator trained on local samples from a chain-structured probabilistic model. We then test our hypothesis experimentally in more complex models, training an autoregressive language model on samples from Bayes nets but only including a subset of variables in each sample. We test language models’ ability to match conditional probabilities with and without intermediate reasoning steps, finding that intermediate steps are only helpful when the training data is locally structured with respect to dependencies between variables. The combination of locally structured observations and reasoning is much more data-efficient than training on all variables. Our results illustrate how the effectiveness of reasoning step by step is rooted in the local statistical structure of the training data.

Humans’ capacity for cumulative culture is remarkable: we can build up vast bodies of knowledge over generations. Communication, particularly via language, is a key component of this process. Previous work has described language as enabling posterior passing, where one Bayesian agent transmits a posterior distribution to the next. In practice, we cannot exactly copy our beliefs into the minds of others–we must communicate over the limited channel language provides. In this paper, we analyze cumulative culture as Bayesian reinforcement learning with communication over a rate-limited channel. We implement an agent that solves a crafting task and communicates to the next agent by approximating the optimal rate-distortion trade-off. Our model produces documented effects, such as the benefits of abstraction and selective social learning. It also suggests a new hypothesis: selective social learning can be harmful in tasks where initial exploration is required.

Abstract Goal-directed behavior is a deeply important part of human psychology. People constantly set goals for themselves and pursue them in many domains of life. In this paper, we develop computational models that characterize how humans pursue goals in a complex dynamic environment and test how well they describe human behavior in an experiment. Our models are motivated by the principle of resource rationality and draw upon psychological insights about people’s limited attention and planning capacities. We find that human goal pursuit is qualitatively different and substantially less efficient than optimal goal pursuit in our simulated environment. Models of goal pursuit based on the principle of resource rationality capture human behavior better than both a model of optimal goal pursuit and heuristics that are not resource-rational. We conclude that the way humans pursue goals is shaped by the need to achieve goals effectively as well as cognitive costs and constraints on planning and attention. Our findings are an important step toward understanding humans’ goal pursuit as cognitive limitations play a crucial role in shaping people’s goal-directed behavior.