Computational theories of attention in decision making

Constraining a reinforcement learning and sequential sampling model with visual fixation data can enhance its predictive ability. The key is to allow learned option values and relative gaze to jointly influence the accumulation of evidence prior to choice. Using data from two eye-tracking experiments, we tested several model variants that differ in how they integrate values and gaze in the formation of drift rates (e.g., linear versus nonlinear integration, additive versus multiplicative gaze). The models capture a wide range of choice-RT effects, learning patterns, and valuation biases. We find consistent support for a nonlinear relationship between learned values and mean drift rates, one that induces competition between the accumulators. The evidence for additive versus multiplicative gaze was mixed; however, unlike other models, our best multiplicative model does not predict stronger gaze effects for high value options. Advantages and limitations of our modeling approach will be discussed, as well as directions for future research.

This is an in-person presentation on July 20, 2026 (10:40 ~ 11:00 EDT).

Consider a choice of whether to accept or reject a gamble with three equiprobable states/outcomes, e.g., based on the winner of a rock-paper-scissors game with -$100 if rock, $10 if paper, $150 if scissors. Standard decision theory says that this decision should depend only on the curvature of the utility function for money. What this decision should not depend on is the outcomes assigned to the states (rock, paper, scissors) in other choices. However, some recent work has argued that people’s decisions are based on noisy representations of their options, constrained by limited attention. If people face such constraints, they need to choose how to allocate their attention across states. Consider the extreme case where they can only attend to one outcome – which state should they consider? We argue that people should allocate more attention to the states with more variance in their outcomes across decisions, and thus might put more weight on those high-variance states in their choices (Roe et al. 2001; Yang & Krajbich 2023). We tested this idea with two eye-tracking experiments, one in person and one online, where participants chose whether to accept or reject three-outcome gambles. We presented these outcomes at three distinct locations on the screen and manipulated the across-trial variance of outcomes by location. Surprisingly, people appear to allocate more gaze and more decision weight to the low-variance state. This puzzling result forces us to rethink the factors that drive attention in risky choice.

This is an in-person presentation on July 20, 2026 (11:00 ~ 11:20 EDT).

Whether it be buying a phone or choosing a route to work, many everyday decisions require sampling and integrating information from multiple attributes to make a choice. Existing computational models of this process are limited, because they either treat attention as a passive input to predict behaviour, or they assume information search to follow overly complex rationality principles. We introduce the Multi-Attribute Search and Choice (MASC) model, a Bayesian cognitive model which addresses these limitations by proposing that people search in an efficient, though not necessarily optimal, manner. Formally, MASC represents subjective values of attributes and options as belief distributions that are iteratively updated as attention selectively samples information. Once there is sufficient evidence in favour of one option, a choice is made. Central to the model is its myopic search rule, which assumes observers plan only one step ahead, sampling the attribute that maximizes the probability of choosing the associated option at the next step. Despite its simplicity, this search rule gives rise to attention being guided by value, uncertainty and attribute importance, enabling the model to account for variety of empirical evidence and search dynamics while outperforming alternative search rules. Using a preregistered study combining EEG and eye-tracking, we further show that MASC-derived estimates of belief updating, at both the level of attributes and of options, differentially predict fixation-locked neural activity. Together, these findings provide converging behavioral and neural evidence that multi-attribute decision making involves dynamic, fixation-driven updating consistent with Bayesian inference, as formalized by MASC.

This is an in-person presentation on July 20, 2026 (11:20 ~ 11:40 EDT).