When people choose between options they can see, two things are known to shape the outcome. The first is where they look: evidence for the item under the eyes counts more than evidence for items off to the side. The second is what they expected before looking, such as a brand's reputation or the average quality of the menu. What is not known is how the second fades as the first proceeds. Does an expectation about an item lose its grip only as that item is inspected, does it fade on a clock regardless of where the eyes go, or does it persist throughout the decision? This note proposes to answer that question with eye tracking, forced-fixation manipulations, and misleading prior cues. The competing models are gaze-dependent evidence accumulators (models in which a noisy running tally of evidence for each option grows faster while that option is fixated, until one tally hits a threshold) with different rules for how prior expectations decay.
The core idea borrows from machine learning. In Bayesian optimization with expert priors, the prior's influence is deliberately faded according to a schedule so that a wrong prior cannot trap the search, and some methods also reject priors that the data contradict. Those schedules are indexed by a global counter, the iteration number. The human analogue has never been tested: whether prior influence is indexed per item by how long that item has been looked at, globally by elapsed time, or by how surprising the fixated evidence was. Each rule makes a different and testable prediction about choices on trials with the same response time but a different split of gaze.
Core Claim
We claim that the influence of a prior expectation about an option decays as a function of the gaze dwell on that specific option, not as a function of elapsed decision time. We further claim that this decay speeds up when fixated evidence contradicts the prior, which is a human counterpart to the data-driven safeguard in Dynamic Priors in Bayesian Optimization for Hyperparameter Optimization. If true, this explains two puzzles within one mechanism. The first is the sign inversion reported by Considering What We Know and What We Don't Know: longer exposure helps better-than-expected items and hurts worse-than-expected ones. The second is the underpredicted first-fixation durations and overpredicted single-fixation trials in Fixation patterns in simple choice reflect optimal information sampling. Either claim can fail cleanly, and the experiments below are built so that each alternative rule wins if it is the right one.
The Model Family
All models share a gaze-weighted accumulator in the style of the attentional drift diffusion model. That model discounts unattended evidence by a factor $\theta$, estimated at 0.36 in perceptual choice and about 0.30 in value choice. On top of this, the momentary value estimate that drives accumulation for item $i$ at time $t$ is a mixture of a prior mean $m_i$ (set by a cue or by the choice context) and the running average $\bar{x}_i(t)$ of noisy samples taken so far:
$$\hat{v}_i(t) = w_i(t)\, m_i + \big(1 - w_i(t)\big)\, \bar{x}_i(t).$$
This is the same mixing form DynMeanBO applies to a Gaussian process prior mean (Incorporating Expert Priors into Bayesian Optimization via Dynamic Mean Decay). The models differ only in the prior weight $w_i(t)$.
The static model holds $w_i$ constant. It corresponds to a persistent drift bias, which Choice History Biases Subsequent Evidence Accumulation found to be the dominant form of across-trial history bias.
The clock model uses $w(t) = \tau_0/(\tau_0 + \lambda t)$ for all items. Here $\tau_0$ is the prior's precision (how confident the prior is), $\lambda$ is the precision gained per unit time, and $t$ is elapsed time in the trial. This is the direct analogue of πBO's global iteration-indexed schedule (πBO).
The dwell model replaces $t$ with the effective dwell on item $i$, $d_i(t) = D_i(t) + \theta\,(t - D_i(t))$, where $D_i(t)$ is the foveal dwell on item $i$ so far. This is what an ideal Bayesian observer with gaze-scaled sample precision implies, and it is the implicit mechanism of the Frömer et al. model once free gaze is added.
The surprise model uses the dwell form but makes $\tau_0$ itself uncertain. Prior precision is scaled by a reliability term $\rho_i(t)$, which is updated from the likelihood of the fixated samples under the prior:
$$w_i(t) = \frac{\rho_i(t)\,\tau_0}{\rho_i(t)\,\tau_0 + \lambda\, d_i(t)}.$$
A contradicting first fixation therefore collapses the prior faster than fixed-precision Bayes would.
All four models are fit with the same boundary, the same noise, the same non-decision time, and an optional urgency term. Differences in fit then cannot come from generic time-varying speed-accuracy tradeoffs. Gaze is treated as an observed input, following GLAM and aDDM practice (GLAMbox). For the fixation-duration analysis, we also embed each prior rule inside the Callaway et al. sampling policy, so that gaze is generated by the model rather than supplied to it.
Research Program
The work has three parts.
First, we reanalyze existing free-gaze data. These are the binary and trinary food-choice datasets used by Callaway et al., the 9 to 36 item snack data of Uncovering the computational mechanisms underlying many-alternative choice, and the three-lottery data of Gaze-dependent evidence accumulation predicts multi-alternative risky choice behaviour. In these data the prior can only be defined from context: the session's mean rating, or the display's mean rating. This part is cheap. Its limitation is that dwell and value are confounded, because people look longer at better items.
Second, we run a new experiment that breaks that confound. About 50 participants make binary and trinary snack choices under free gaze with 60 Hz or faster eye tracking. Before each item appears, a cue states an expected rating (for example, "shoppers rated this brand 8/10"). The cue is valid on 70% of trials and misleading by a controlled amount on 30%, mirroring the good and adversarial prior conditions of the optimization literature.
Third, in a forced-exposure block modeled on Experiment 2 of Tavares et al., one item is shown alone for a fixed dwell of 300, 700, or 1200 ms before free viewing, while total trial duration is held fixed. This makes dwell on a given item an experimental variable rather than a choice. The three hypotheses separate cleanly here. The clock model predicts that prior influence depends only on total time. The dwell model predicts that it depends on forced dwell on the cued item. The surprise model predicts an interaction between forced dwell and cue validity: steep decay on misleading trials, gentle decay on valid ones.
Baselines
The naive reference point is the standard drift diffusion model with no gaze term and no prior. In perceptual line-orientation choice, Tavares et al. report 68.4% held-out choice prediction accuracy for it, against 72.3% for the attentional model. Any proposed model must clear both numbers in data of that type.
The strongest existing gaze-dependent accumulator for multi-option choice is the gaze-dependent leaky accumulator of Molter et al., which reports 74.3% choice prediction accuracy on three-lottery risky choice, against 57.9% for multialternative decision field theory. On that dataset, beating it means exceeding 74.3% held-out accuracy and improving held-out log-likelihood, not merely matching it with extra parameters.
The strongest competitor on the actual question is the Bayesian expectation- and precision-weighted model of Frömer et al., extended to free gaze. That extension is our dwell model with fixed $\tau_0$. No reported result was found on Lacuna for its fit indices, so it will be refit here. The optimal-sampling policy of Callaway et al. is the strongest model for fixation durations, and no reported numeric fit was found on Lacuna for it either.
The key ablations are the static and clock versions of $w$, and the surprise model with $\rho$ frozen at 1. The frozen-$\rho$ version isolates whether reliability updating adds anything beyond ordinary Bayesian shrinkage.
For a fair comparison, every model is fit hierarchically with the same sampler and the same posterior-draw budget, using the GLAMbox/PyMC hierarchical machinery. Every model has the same boundary and urgency options. All models are scored on the same held-out trials: every fifth trial per participant, and in the new experiment, entire held-out cue-validity blocks. We report WAIC, as Thomas et al. did, and also held-out log-likelihood, since parameter-count penalties can mislead when models are nested.
Why not just…?
Why not just shift the starting point of the accumulator? A starting-point offset dilutes on its own as evidence accumulates, so it looks like free decay. Urai et al. found, however, that history biases act on the rate of accumulation and persist into slow responses, which a starting-point offset cannot produce. A starting-point model is also blind to where the eyes go, so it predicts no effect of forced dwell once total time is fixed. We include it as a baseline, and the forced-exposure block tests it directly.
Why not just use the Bayesian model with gaze-scaled precision, since it already implies dwell-indexed decay? This is the real threat. If fixed-precision Bayes fits as well as the surprise model, the honest conclusion is that human prior decay is ideal-observer shrinkage, with no evidence for a separate safeguard. That would still be new, because it establishes dwell rather than clock as the index under free gaze, which Frömer et al. could not test with imposed exposure. The decisive contrast is misleading-cue trials: the surprise model predicts that prior influence after one contradicting fixation falls well below the fixed-precision prediction. The misleading-cue rate is set high enough to estimate this per participant.
Why not just add an urgency signal or collapsing bound, which also makes early and late evidence count differently? Urgency is a global, clock-indexed effect. It is included in every model, so it cannot be what separates them. The dwell-versus-clock contrast relies on trials matched in total time that differ in gaze split, which urgency cannot mimic.
Why not just fit the attentional discount separately for early and late fixations? That describes a time-varying gaze effect without saying why it varies. It predicts no dependence on cue validity and no sign reversal for worse-than-expected items. We fit it as a descriptive check. If it matches the surprise model's held-out likelihood, the mechanistic claim is weakened.
Decisive Checks
The central test is a fixed-total-time split in the forced-exposure block. We regress choice of the cued item on its cue value, interacted with forced dwell and cue validity. The clock model predicts a flat slope across dwell levels. The dwell model predicts a declining slope. The surprise model predicts a slope that declines much faster on misleading trials.
The second test is fixation dynamics. With the Callaway et al. policy generating gaze, we ask whether a decaying prior, particularly the surprise version, corrects the underpredicted first-fixation durations and the overpredicted single-fixation trials. The prediction is that a strongly weighted early prior lengthens the first look until its reliability has been checked.
The third test is the Frömer et al. sign inversion under free gaze. The dwell and surprise models predict that the inversion grows with dwell on the item itself. The clock model predicts that it grows with total response time regardless of where the eyes were.
The claim fails if the clock or static model wins on held-out likelihood in the forced-exposure data. It is reduced to plain Bayesian shrinkage if frozen reliability matches the full surprise model.
References
- Dynamic Priors in Bayesian Optimization for Hyperparameter Optimization
- Considering What We Know and What We Don’t Know: Expectations and Confidence Guide Value Integration in Value-Based Decision-Making
- Fixation patterns in simple choice reflect optimal information sampling
- The Attentional Drift Diffusion Model of Simple Perceptual Decision-Making
- Incorporating Expert Priors into Bayesian Optimization via Dynamic Mean Decay
- Choice history biases subsequent evidence accumulation
- $\pi$BO: Augmenting Acquisition Functions with User Beliefs for Bayesian Optimization
- GLAMbox: A Python toolbox for investigating the association between gaze allocation and decision behaviour
- Uncovering the computational mechanisms underlying many-alternative choice
- Gaze-dependent evidence accumulation predicts multi-alternative risky choice behaviour