Behavioral signatures of post-decisional attention in preferential choice Authors: Ariel Zylberberg, Ian Krajbich, Michael N Shadlen, Emilio Salinas, Ariel Zylberberg, Ian Krajbich, Michael N Shadlen Published: 2026 ABSTRACT Attention plays a key role in decision-making by directing limited cognitive resources to relevant information. It has been proposed that attention also biases the decision process, due to a multiplicative interaction between attention and subjective value (e.g., Krajbich et al., 2010). We tested two predictions of models that posit a causal multiplicative effect of attention on decision formation: (i) the last fixation should be more informative about the choice when the overall value of the alternatives is high, and (ii) more attention should be directed to the chosen option when choices conflict with stated preferences than when they do not. Reanalyzing several datasets from a food-choice task, we found no evidence supporting these predictions. An alternative model where attention reflects choices after the decision has completed explains key observations, including the last-fixation bias, the gaze-cascade effect, and the effect of the overall value of the alternatives on response times. However, this model does not fully account for the association between dwell time and choice. We conclude that gaze behavior prior to the choice report likely reflects both decisional and post-decisional processes. Introduction Attention plays a key role in decision-making, enabling individuals to focus on relevant information while ignoring distractions. It has also been hypothesized that attention biases decision-makers’ preferences and choices. During decisions involving spatially distributed stimuli, gaze progressively shifts toward the option ultimately chosen, a phenomenon termed the gaze cascade effect (Shimojo et al., 2003; Glaholt and Reingold, 2009). Additionally, decision-makers often select the option they fixate on last before reporting their choice (a ‘last-fixation’ bias; Krajbich et al., 2010). Experimental manipulations of gaze, including spatial cues, variations in exposure duration, and salience control, have demonstrated that options receiving more attention are more likely to be chosen (Nittono and Wada, 2009; Bhatnagar and Orquin, 2022; Pleskac et al., 2023). As gaze reflects spatial attention, these findings support the idea that attention biases choice (Shimojo et al., 2003; Armel et al., 2008; Zajonc, 1968). The attentional drift-diffusion model (aDDM) formalizes this hypothesis in a way that makes it suitable to quantitatively explain choice and response time (RT) (Krajbich et al., 2010). The aDDM builds on the drift-diffusion model of decision-making (DDM), in which decisions are made by accumulating noisy samples of momentary evidence over time (Ratcliff, 1978). In preference-based decisions, the momentary evidence depends on the subjective value difference between the options. The decision is thought to terminate when the accumulated evidence crosses an upper or lower bound, simultaneously resolving the choice and the time it took to make it. The DDM and some of its variants, such as random walk, race, and attractor models, have been successful in explaining both choice and RT in a range of perceptual and cognitive decisions in which no role for attention is assumed (Link, 1975; Ratcliff and McKoon, 2008; Usher and McClelland, 2001; Vickers, 1979; Wang, 2002). The aDDM extends the DDM by including a role for attention in the decision process. It proposes that attention can change the subjective value of decision alternatives: specifically, the value of unattended options is discounted by a multiplicative factor. This framework assumes that attention exerts its influence intra-decisionally—that is, during the deliberation process, before a choice is made. Originally developed for preference-based decisions between two options, the aDDM has been extended to multi-alternative (Krajbich and Rangel, 2011; Thomas et al., 2021), perceptual (Smith and Krajbich, 2019; Tavares et al., 2017), purchase (Krajbich et al., 2012), and attribute-based decisions (Fisher, 2021; Yang and Krajbich, 2023). The model provides an explanation for the association between choice, RT, subjective value, and gaze allocation, including the apparent causal influence of gaze on choice (for a review, see Krajbich, 2019). Here, we examine two behavioral predictions of models that posit a multiplicative influence of attention on subjective value. These predictions concern how gaze allocation (i.e., time spent viewing each item) relates to the overall value of the alternatives and to whether choices are consistent with stated preferences. One prediction of the aDDM and related models is that gaze allocation at the end of the decision should be more predictive of the choice when the overall value of the alternatives is higher (Ting and Gluth, 2025). A second prediction is that the difference in dwell time between items should depend strongly on choice consistency—specifically, that the lower-valued item should be looked at for longer for choices that are inconsistent with stated preferences. We tested the predictions by reanalyzing data from several previously published value-based decision tasks (Krajbich et al., 2010; Smith and Krajbich, 2018; Chen and Krajbich, 2016; Gwinn and Krajbich, 2016; Folke et al., 2016; Sepulveda et al., 2020). Contrary to these predictions, the data show systematic deviations from the aDDM and related frameworks (Callaway et al., 2021; Jang et al., 2021). Instead, the results suggest that the association between gaze and choice is at least partly post-decisional, arising after a covert commitment to a decision but before the overt response is executed (Cavanagh et al., 2014; Westbrook et al., 2020). Results We analyzed data from previously published studies that used the food-choice paradigm introduced by Krajbich et al., 2010. In this paradigm, hungry participants performed two stages. In the first stage, participants viewed snack food items one at a time and rated how much they would like to consume each one (Figure 1A); most of the studies we reanalyze used a numerical liking scale (Krajbich et al., 2010; Smith and Krajbich, 2018; Chen and Krajbich, 2016; Gwinn and Krajbich, 2016), while others elicited subjective value through an incentive-compatible Becker–DeGroot–Marschak auction (Folke et al., 2016; Sepulveda et al., 2020). In the second stage, participants were presented with pairs of items and asked to choose which one they would prefer to consume at the end of the experiment (Figure 1B). Throughout the choice stage, participants’ gaze was recorded, allowing identification of the moments when attention was directed to the left versus right item (Figure 1C). Food-choice task. The food-choice paradigm consists of two phases. In the first phase (A), participants were shown pictures of snack food items, one at a time, and were asked to rate how much they would like to consume each item, either on a numerical liking scale (Krajbich et al., 2010; Smith and Krajbich, 2018; Chen and Krajbich, 2016; Gwinn and Krajbich, 2016) or by stating their maximum willingness to pay in an incentive-compatible auction (Folke et al., 2016; Sepulveda et al., 2020). Note that while the original experiments used photographs of real packaged foods, the items depicted here are illustrative mockups. In the second phase (B), participants were presented with pairs of items and asked to choose which one they would prefer to consume at the end of the experiment. Participants responded manually; in most datasets, responses were made using the left and right arrow keys on a keyboard. (C) Gaze allocation between the left and right items shown from the moment both snacks appeared on the screen, for 11 representative trials from Krajbich et al., 2010. Red and blue indicate gaze directed to the left and right items, respectively. Times when the gaze was not directed to either item are left blank. Black dots indicate the time when the left or right key was pressed. The attentional drift-diffusion model The decision-making process in the aDDM is governed by the state of a scalar decision variable, x\begin{document}$x$\end{document}, which takes the value zero at the start of each trial. The decision variable is updated according to(1)xt=xt−1+d(rleft−θrright)+ϵt\begin{document}$$\displaystyle x_{t}=x_{t-1}+d(r_{\mathrm{left}}-\theta r_{\mathrm{right}})+\epsilon_{t}$$\end{document} when the decision-maker is looking at the left item, and according to(2)xt=xt-1+d⁢(θ⁢rleft-rright)+ϵt\begin{document}$$\displaystyle x_{t}=x_{t-1}+d(\theta r_{\mathrm{left}}-r_{\mathrm{right}})+\epsilon_{t}$$\end{document} when looking at the right item. rleft\begin{document}$r_{\mathrm{left}}$\end{document} and rright\begin{document}$r_{\mathrm{right}}$\end{document} represent the values assigned to the left and right items during the rating phase, respectively. If the decision variable reaches a bound at +1, the decision-making process terminates and the left item is selected, and if it reaches –1, the right item is selected. The parameter d\begin{document}$d$\end{document} controls the integration speed, θ\begin{document}$\theta$\end{document} is a parameter between 0 and 1 that determines how much the value of the unattended item is discounted, and ϵt\begin{document}$\epsilon_{t}$\end{document} is white Gaussian noise with variance σ2\begin{document}$\sigma^{2}$\end{document}. The difference between the aDDM and the standard drift-diffusion model is that gaze modulates the drift rate through the parameter θ\begin{document}$\theta$\end{document}. The last-fixation bias on choice does not increase with overall value A central feature of the aDDM is the multiplicative interaction between gaze and value. This multiplicative interaction accounts for faster response times when the overall value of the options is high—a phenomenon we refer to as the magnitude effect on response times (MERT) (Smith and Krajbich, 2019; Ratcliff et al., 2018). Given that attention has a greater influence when both options are highly valued, the last-fixation bias—the tendency to choose the item that was fixated last—should also be amplified under these conditions. A similar observation has recently been made by Ting and Gluth, 2025. We refer to this prediction of the aDDM as the magnitude effect on the last-fixation bias (MELFB). We illustrate the MELFB prediction through simulations of the aDDM using the best-fitting parameters identified by Krajbich et al., 2010 and Smith and Krajbich, 2019. To evaluate the impact of attention on choice, we apply logistic regression with the following model:(3)logit[pleft]=∑s(βΔr,sΔr+β0,s)Is+∑gβLastFix,gIgILastFix,\begin{document}$$\displaystyle \mathrm{logit}[p_{\mathrm{left}}]=\sum_{s}\left(\beta_{\Delta r,s}\Delta r+\beta_{ 0,s}\right)I_{s}+\sum_{g}\beta_{\mathrm{LastFix},g}I_{g}I_{\mathrm{LastFix}},$$\end{document} where Δr\begin{document}$\Delta r$\end{document} represents the difference in value between the left and right items. Is\begin{document}$I_{s}$\end{document} is an indicator variable that equals 1 for trials completed by subject s\begin{document}$s$\end{document} and 0 otherwise. ILastFix\begin{document}$I_{\mathrm{LastFix}}$\end{document} is another indicator variable, set to 1 if the left item was fixated last and 0 if the right item was fixated last. We categorize trials into quintiles of Σr\begin{document}$\Sigma r$\end{document}, the sum of the rating assigned to the left and right items; the variable Ig\begin{document}$I_{g}$\end{document} identifies trials belonging to quintile g∈{1…5}\begin{document}$g\in\{1\dots 5\}$\end{document}. The βs are the regression coefficients. The first set of terms on the right-hand side of the equation (the summation over s\begin{document}$s$\end{document}) captures the effect of Δr\begin{document}$\Delta r$\end{document} (along with a participant-specific bias) on the probability of choosing the left item. The second set of terms (the summation over g\begin{document}$g$\end{document}) reflects the influence of attention, measured by whether the left or right item was fixated last, with this effect estimated for each quintile of Σr\begin{document}$\Sigma r$\end{document}. Figure 2A shows the regression coefficient βLastFix,g\begin{document}$\beta_{\mathrm{LastFix},g}$\end{document} for the different quintiles of overall value, for the simulations of the aDDM model. The dashed black line is not a fit to these data points, but is obtained from a model similar to that of Equation 3 except that the term associated with the quantiles of overall value is replaced by an interaction between ILastFix\begin{document}$I_{\mathrm{LastFix}}$\end{document} and Σr\begin{document}$\Sigma r$\end{document} (Equation 12). The analysis shows that the influence of the item attended last on choice increases with the overall value of the items (Figure 2A). aDDM behavioral predictions. (A) Estimated strength of the association between last-dwell focus and choice (logistic regression coefficient; Equation 3) as a function of overall value (Σr\begin{document}$\Sigma r$\end{document}). The five data points correspond to quintiles of the data, split by Σr\begin{document}$\Sigma r$\end{document}. The data were generated from simulations of the aDDM using the parameters reported in Krajbich et al., 2010 and Smith and Krajbich, 2019. The dashed line is derived from a related regression model that includes an interaction term between Σr\begin{document}$\Sigma r$\end{document} and whether the last fixation before the report was on the ultimately chosen or unchosen item. Error bars indicate s.e. (B) Time spent looking at the chosen item minus time spent looking at the unchosen item, as a function of response time. The data were generated from simulations of the aDDM. For each participant, trials were grouped into 20 categories defined by response-time decile and whether the choice was consistent with the initial ratings. Trials in which the two items received the same rating during the rating phase were excluded because such choices cannot be classified as either consistent or inconsistent. The response times shown on the abscissa correspond to the mean response time across participants for each decile. Error bars indicate s.e. across participants. The difference in dwell time is independent of choice consistency A second unexplored prediction of the aDDM is that the difference in total fixation time between the chosen and unchosen items should depend on the consistency of the choice with the item ratings. We define inconsistent choices as those in which the lower-rated item is selected over the higher-rated one. The difference in total fixation time between the chosen and unchosen items is referred to as ΔDwell. According to the aDDM, if individual dwells are not influenced by item value, ΔDwell should be greater for inconsistent choices than for consistent ones. This prediction is confirmed in our simulations of the aDDM (Figure 2B; p < 10-7, Equation 13, H0:βc=0\begin{document}$H_{0}\mbox{:}\,\beta_{c}=0$\end{document}). This prediction can be understood as follows. If, in a given trial, the decision-maker had primarily attended to the higher-rated item, the likelihood of making an inconsistent choice would be lower than if attention had been equally divided between the options, because attention would have increased the extent to which the drift rate favored the higher-rated item. Thus, when a choice is known to be inconsistent, it is more likely that gaze was predominantly directed to the lower-rated item. This logic also applies to consistent choices, but to a lesser extent. Even if attention is primarily allocated to the lower-valued item during a trial, it is still highly likely that the decision-maker will ultimately choose the higher-valued item. This is because, while attending to the lower-valued item reduces the effective drift rate in favor of the higher-valued item, it does not usually result in a change in the sign of the drift rate. Therefore, knowing that the decision was consistent does not provide as much information about the allocation of attention as knowing that the decision was inconsistent. A similar logic explains why ΔDwell should depend on response time. The aDDM predicts that ΔDwell increases with RT for inconsistent choices (Figure 2B). This is because the response time sets an upper limit on ΔDwell, so the longer the RT, the greater the value that ΔDwell can be. In contrast, ΔDwell decreases—and even becomes negative—with RT for consistent choices (Figure 2B). This is because if the response time is longer than expected—given the values of the items being compared—then it is likely that attention was mostly focused on the lower-valued item. The effect is not as strong for inconsistent choices, because even if attention were primarily focused on the lower-valued item, the drift rate would still usually favor the unselected, higher-valued item. Testing the predictions in data from the food-choice task We tested these predictions of the aDDM with data from the food-choice task. To test the prediction regarding the MELFB, we fit the logistic regression model (Equation 3) to the data of Krajbich et al., 2010 and other datasets of the food-choice task (Smith and Krajbich, 2018; Chen and Krajbich, 2016; Gwinn and Krajbich, 2016). In contrast with the prediction of the aDDM, the estimated strength of the association between last-dwell focus and choice was either non-significant or significantly negative (Figure 3; p-values and βs indicated in the figure). A similar result was recently reported by Ting and Gluth, 2025. That is, in contrast with the model’s prediction, the last-fixation bias does not increase with the overall value of the alternatives. Association strength between last dwell and choice. Estimated strength of the association between last-dwell focus and choice as a function of Σr\begin{document}$\Sigma r$\end{document}. Same analysis and conventions as in Figure 2A. Each panel shows data from a different dataset of the food-choice task: (A) Krajbich et al., 2010, (B) Smith and Krajbich, 2018, (C) Chen and Krajbich, 2016, and (D) Gwinn and Krajbich, 2016. We also tested the prediction regarding the difference in ΔDwell between consistent and inconsistent choices. Across several datasets, we found no significant differences in ΔDwell for consistent and inconsistent choices (Figure 4; Equation 13, H0:βc=0\begin{document}$H_{0}:\beta_{c}=0$\end{document}; p>0.13\begin{document}$p > 0.13$\end{document} for every dataset). This is incompatible with the prediction of the aDDM. This incompatibility is not specific to the aDDM, but as we show below, applies to other instantiations of models, including those in which gaze is optimally allocated. (Note that some of these datasets were excluded from Figure 3 because last-dwell focus was unavailable in the public data.) Difference in looking time for consistent and inconsistent choices. Time spent looking at the chosen item minus time spent looking at the unchosen item, as a function of response time. Same analysis and conventions as in Figure 2. The seven panels correspond to behavioral data from (A) Krajbich et al., 2010, (B) Smith and Krajbich, 2018, (C) Chen and Krajbich, 2016, (D) Gwinn and Krajbich, 2016, (E) Folke et al., 2016, and (F, G) Sepulveda et al., 2020. In Sepulveda et al., 2020, participants either selected the item they preferred or the one they did not prefer. We analyzed the two variants separately. Is the gaze–choice association purely post-decisional? The aspects of the data we analyzed in Figures 3 and 4 are suggestive of a non-multiplicative interaction between attention and value (Cavanagh et al., 2014; Westbrook et al., 2020). We explore a model in which the link between gaze and choice arises only after a choice has been covertly made. Attention has no effect on the choice itself or on the time taken to make the choice. We refer to this model as the post-decision-gaze (PDG) model. In the PDG model, the decision is made by accumulating momentary evidence over time. The momentary evidence is represented by samples from a Gaussian distribution. The mean of the sampling distribution is a linear function of Δr\begin{document}$\Delta r$\end{document}, such that the drift rate of the drift-diffusion process (µ) is given by(4)μ=κ(rleft−rright),\begin{document}$$\displaystyle \mu=\kappa(r_{\mathrm{left}}-r_{\mathrm{right}}),$$\end{document} where κ\begin{document}$\kappa$\end{document} is a signal-to-noise parameter. Unlike the aDDM, the drift rate is not modulated by the focus of attention. Unlike most implementations of the DDM, in the PDG model, the variance of the momentary evidence depends on Σr\begin{document}$\Sigma r$\end{document}. This assumption is needed to explain why response times depend on both Δr\begin{document}$\Delta r$\end{document} and Σr\begin{document}$\Sigma r$\end{document} (Ratcliff et al., 2018). We parameterize the variance as(5)σ2=1+γ⁢(rleft+rright),\begin{document}$$\displaystyle \sigma^{2}=1+\gamma\left(r_{\mathrm{left}}+r_{\mathrm{right}}\right),$$\end{document} where γ\begin{document}$\gamma$\end{document} is fit to the data. The assumption has empirical support in neurobiology (‘Discussion’). The evidence accumulation process begins after a short sensory delay, τs\begin{document}$\tau_{\mathrm{s}}$\end{document} (Figure 5A). For the purposes of our analysis, we fixed τs\begin{document}$\tau_{\mathrm{s}}$\end{document} at 0.3 s for all participants. Note that while neurophysiological studies in monkeys have estimated that τs\begin{document}$\tau_{\mathrm{s}}$\end{document} is on the order of 0.2 s (Roitman and Shadlen, 2002; Steinemann et al., 2022), τs\begin{document}$\tau_{\mathrm{s}}$\end{document} is likely to be longer in the food-choice task since participants start each trial fixating on a central spot before directing their gaze to one of the choice alternatives (Krajbich et al., 2010). Sketch of the PDG model. (A) The decision is generated by accumulating momentary evidence over time until the process reaches either an upper or lower bound. Evidence accumulation begins after a sensory delay, τs\begin{document}$\tau_{\mathrm{s}}$\end{document}. Once a bound is crossed, an additional motor delay, τm\begin{document}$\tau_{\mathrm{m}}$\end{document}, elapses before the manual response is executed. Thus, the response time equals the decision time plus the non-decision delays τs\begin{document}$\tau_{\mathrm{s}}$\end{document} and τm\begin{document}$\tau_{\mathrm{m}}$\end{document}. Gaze does not affect the decision process. Instead, at a random latency τe\begin{document}$\tau_{\mathrm{e}}$\end{document} after the bound crossing, the gaze is directed toward the chosen option and remains there until the manual response. Because τe\begin{document}$\tau_{\mathrm{e}}$\end{document} is typically shorter than τm\begin{document}$\tau_{\mathrm{m}}$\end{document}, the chosen item is usually the last fixated item before the response. (B) Example simulation in which the gaze is already on the chosen item at time τe\begin{document}$\tau_{\mathrm{e}}$\end{document} following bound crossing; therefore, no gaze shift occurs. (C) Due to stochastic variability, τe>τm\begin{document}$\tau_{\mathrm{e}}{ > }\tau_{\mathrm{m}}$\end{document} in a subset of trials, causing the post-commitment gaze shift to the chosen item to occur only after the manual response. This explains why, in some trials—including the example shown—the non-chosen item is the last one fixated before the response. The model also includes a non-decision latency, τm\begin{document}$\tau_{\mathrm{m}}$\end{document}, between the time when the decision-maker commits to a choice, signaled by crossing the decision threshold, and the time when a key is pressed to report the choice (Figure 5A). τm\begin{document}$\tau_{\mathrm{m}}$\end{document} accounts for latencies related to motor preparation and is independent of decision difficulty. The response time is given by the sum of τm\begin{document}$\tau_{\mathrm{m}}$\end{document}, τs\begin{document}$\tau_{\mathrm{s}}$\end{document}, and the decision time (Figure 5A). In the model, attention has no causal effect on the decision process. Therefore, there is a 50% chance that the gaze will be directed to either item at the time a threshold is crossed. The key assumption of the model is that once the decision variable crosses the decision threshold, the gaze is directed to the chosen item. To account for eye-movement-related latencies, we assume that directing gaze to the chosen item occurs with a latency of τe\begin{document}$\tau_{\mathrm{e}}$\end{document} from the time of bound crossing (Figure 5A), after which the gaze is held on the chosen item until the response. If the decision-maker was already looking at the chosen item after τe\begin{document}$\tau_{\mathrm{e}}$\end{document} has elapsed from the time of bound crossing, no additional gaze shift occurs (Figure 5B). Because the time it takes to make an eye movement is usually less than the time it takes to complete the manual response, τm\begin{document}$\tau_{\mathrm{m}}$\end{document}, the decision-maker is more likely to be looking at the chosen item when a key is pressed (Figure 5A and B). However, due to variability in τm\begin{document}$\tau_{\mathrm{m}}$\end{document} and in τe\begin{document}$\tau_{\mathrm{e}}$\end{document}, the time at which gaze is directed to the chosen item after choice commitment varies from trial to trial, and may even occur after the key press, as in Figure 5C. Crucially, in contrast with the aDDM and related models (Krajbich et al., 2010; Thomas et al., 2019; Krajbich and Rangel, 2011), it is the choice that affects gaze allocation, not the other way around. Fits of the PDG model to the behavioral data We fit the PDG model to the choice and response time data from Krajbich et al., 2010. Figure 6A and B show the proportion of trials in which participants chose the item on the left and the average RT as a function of Δr\begin{document}$\Delta r$\end{document}. The PDG model provides a good fit to the choice and RT data. Data from Krajbich et al., 2010 and fits of the PDG model. (A) Proportion of trials in which the left item was selected as a function of the difference in value between the left and right items (Δr\begin{document}$\Delta r$\end{document}). Points represent behavioral data and shading represents model fits. (B) Mean response time as a function of the value difference between the two items. (C) Residual response time (after subtracting the contribution of Δr\begin{document}$\Delta r$\end{document}) as a function of the sum of the value of the two items presented in the trial (Σr\begin{document}$\Sigma r$\end{document}). Error bars indicate standard error of the mean (s.e.m.) across trials. (D) Proportion of trials in which the left item was selected, separated by whether the last fixation before the report was on the right (purple) or the left (green) item. (E) Probability that the decision-maker is looking at the item that was ultimately chosen, plotted relative to the time since the two items were presented on the screen (left) and the response time (right). Error bands indicate 95% confidence intervals for the mean across participants. In the stimulus-aligned plot, the data are shown from the first moment that one of the two items is fixated on in at least 50% of the trials, which is ~0.25 s. (F) MELFB for data (black) and model (orange). Same conventions as in Figure 3. (G) Gaze bias for consistent and inconsistent choices. Same conventions as in Figure 4. (H) Proportion of trials in which the left item was selected as a function of the difference in dwell time between the left and right items. Data (black) and model (orange) were grouped in deciles of the dwell difference, separately for each participant, and then averaged across participants. (I) Proportion of trials in which the item looked at first was selected, as a function of duration of the first dwell. Data (black) and model (orange) were grouped in deciles of the first dwell duration, separately for each participant, and then averaged across participants. In all panels except panels (D) and (G), data are shown in black and model simulations are shown in orange. Error bars and error bands, unless otherwise noted, show the standard error of the mean (s.e.m.) across participants (N=39 for both model and data). Figure 6—figure supplement 1.Removing the contribution of Δr\begin{document}$\Delta r$\end{document} from response times.Illustration of the method used to remove the contribution of Δr\begin{document}$\Delta r$\end{document} from the RT. The gray markers indicate the RT for each trial of a representative participant. The abscissa are the values of Δr\begin{document}$\Delta r$\end{document} for the corresponding trial. These data points were fitted with the bell-shaped function shown in the figure, with parameters a\begin{document}$a$\end{document},b\begin{document}$b$\end{document},µ and σ. The best-fitting function captures the general trend in the data, as can be seen by comparing the model fits (black solid line) with the average response time per value of Δr\begin{document}$\Delta r$\end{document} (red, mean plus s.e.m.). To compute the RT residuals, we subtract from each trial the value of RT^\begin{document}$\hat{RT}$\end{document} corresponding to the corresponding value of Δr\begin{document}$\Delta r$\end{document}. Fits were performed independently for each participant. Removing the contribution of Δr\begin{document}$\Delta r$\end{document} from response times. Illustration of the method used to remove the contribution of Δr\begin{document}$\Delta r$\end{document} from the RT. The gray markers indicate the RT for each trial of a representative participant. The abscissa are the values of Δr\begin{document}$\Delta r$\end{document} for the corresponding trial. These data points were fitted with the bell-shaped function shown in the figure, with parameters a\begin{document}$a$\end{document},b\begin{document}$b$\end{document},µ and σ. The best-fitting function captures the general trend in the data, as can be seen by comparing the model fits (black solid line) with the average response time per value of Δr\begin{document}$\Delta r$\end{document} (red, mean plus s.e.m.). To compute the RT residuals, we subtract from each trial the value of RT^\begin{document}$\hat{RT}$\end{document} corresponding to the corresponding value of Δr\begin{document}$\Delta r$\end{document}. Fits were performed independently for each participant. Figure 6—figure supplement 2.Fits of the duration of the dwells.Distribution of the durations of the first dwell (left) and middle dwells (right). Middle dwells include all dwells except the first and last. The durations were fitted with a log-normal distribution (red), independently for the first and middle dwells. The best-fitting log-normal parameters were used to simulate the aDDM and PDG. On each trial, the first dwell is sampled from the corresponding distribution, and the subsequent dwell durations up to the bound crossing are sampled from the distribution of middle dwells. Fits of the duration of the dwells. Distribution of the durations of the first dwell (left) and middle dwells (right). Middle dwells include all dwells except the first and last. The durations were fitted with a log-normal distribution (red), independently for the first and middle dwells. The best-fitting log-normal parameters were used to simulate the aDDM and PDG. On each trial, the first dwell is sampled from the corresponding distribution, and the subsequent dwell durations up to the bound crossing are sampled from the distribution of middle dwells. The model also accounts for MERT—the tendency to make faster decisions when the items being compared are overall more desirable, even when the value difference between them is the same (Smith and Krajbich, 2019; Sepulveda et al., 2020; Ratcliff et al., 2018). To illustrate this effect in Krajbich’s data, we fit a bell-shaped curve to the relation between RT and Δr\begin{document}$\Delta r$\end{document}, and computed residuals of RT by subtracting from each trial’s RT the expectation given by the best-fitting bell-shaped curve (see Figure 6—figure supplement 1 for an illustration of the method). We then analyzed how the residuals of RT correlated with Σr\begin{document}$\Sigma r$\end{document}. This correlation was negative, indicating that responses were faster when Σr\begin{document}$\Sigma r$\end{document} was higher. That is, the data show a magnitude effect on RT (Figure 6C). Note that the analysis of RT residuals rather than raw RT is necessary because of the positive correlation between Δr\begin{document}$\Delta r$\end{document} and Σr\begin{document}$\Sigma r$\end{document} present in the data. The DDM has been considered incapable of explaining the magnitude effect on RT because in the most common version of the DDM the drift rate depends only on the difference in value between options and ignores their absolute value, leading to the erroneous prediction that the response time for a given Δr\begin{document}$\Delta r$\end{document} would be independent of Σr\begin{document}$\Sigma r$\end{document}. The PDG model captures the magnitude effect because the variance of the momentary evidence is allowed to change with Σr\begin{document}$\Sigma r$\end{document} (Equation 5; Ratcliff et al., 2018). For the best-fitting model, the variance of the momentary evidence increased with Σr\begin{document}$\Sigma r$\end{document} (Appendix 1—table 1). Because higher variance leads to faster responses (Zylberberg et al., 2016), the model displays a magnitude effect on RT, similar to the data (Figure 6C). In the PDG model, attention has no causal influence on choice. Yet, it successfully accounts for several features of the observed association between gaze and choice. We compared the probability of choosing the right item—as a function of Δr\begin{document}$\Delta r$\end{document}—on trials in which the last fixation was on the right versus the left item. The model predicts a systematic relation between the last fixation and choice that closely mirrors what is observed in the data (Figure 6D). In the PDG model, a gaze bias toward the chosen item can only occur during the non-decision time between choice commitment and report, giving rise to a gaze cascade effect—the observation that the probability of looking at the ultimately chosen item aligned to the response increases gradually over time (Figure 6E). This gradual increase is explained as the average of step-like events (saccades to the chosen item) that occur at different times with respect to the response time. The variation in timing is due to trial-to-trial variability in non-decision time (captured by the model parameter σnd\begin{document}$\sigma_{\mathrm{nd}}$\end{document}, see ‘Methods’) and in τe\begin{document}$\tau_{\mathrm{e}}$\end{document}. Since the model was fit to maximize the likelihood of its parameters based on the choice and RT data, without using gaze information, the gaze cascade effect can be considered a prediction of the model. The PDG model also correctly predicts the two novel behavioral observations shown in Figures 3 and 4. In the PDG model, there is no significant change in the strength of the last-fixation bias as a function of the overall value of the alternatives, consistent with the experimental data (Figure 6F). This is because, in the model, the probability of directing gaze to the covertly chosen item is independent of variables affecting the decision process, like Δr\begin{document}$\Delta r$\end{document} or Σr\begin{document}$\Sigma r$\end{document}. The model also correctly predicts that ΔDwell does not depend on the consistency of the choice. In the model, ΔDwell is often positive because gaze is directed to the covertly chosen item and because τe\begin{document}$\tau_{\mathrm{e}}$\end{document} is usually smaller than τm\begin{document}$\tau_{\mathrm{m}}$\end{document}. These factors do not depend on the consistency of the choice, resulting in similar values of ΔDwell for consistent and inconsistent choices (p=0.27, Equation 13, H0:βc=0\begin{document}$H_{0}:\beta_{c}=0$\end{document}; Figure 6G). The gaze cascade effect continues after the choice report In the PDG model, the bound crossing may have occurred hundreds of milliseconds before the choice is reported, and the two events are not time-locked due to variability in τm\begin{document}$\tau_{\mathrm{m}}$\end{document} and τe\begin{document}$\tau_{\mathrm{e}}$\end{document}. The model posits that the gaze cascade arises because shifting gaze to the chosen item takes time, and the likelihood of having completed that shift increases with time elapsed since boundary crossing. By this logic, the model predicts an even further increase in the probability of looking at the chosen item immediately after the choice report. To test this prediction, we examined the gaze allocation after the choice report. Figure 7A shows the probability of looking at the chosen item, as a function of time, aligned to the response. The probability of looking at the chosen item continues to increase after the choice report. This can be seen more clearly in Figure 7B, which shows the probability of looking at the chosen item during the 200 ms immediately before and after the choice report. For most participants, the probability is greater after the choice report (p<10−7\begin{document}$\mathrm{p} < 10^{-7}$\end{document}, Wilcoxon signed-rank test). Gaze allocation after the choice report. (A) Probability that the decision-maker is looking at the item that was ultimately chosen, aligned to RT. This probability increases even after the choice report. Error bands indicate 95% confidence intervals for the mean across participants. (B) Proportion of time that the decision-maker is looking at the item that was ultimately chosen, calculated for the last 200 ms before the response (abscissa) and for the first 200 ms after the response (ordinate). Each data point represents one participant. Proportions were calculated as the sum of the time spent looking at the chosen item divided by the time spent looking at either one of the items (i.e., we exclude the times when the gaze was not directed at one of the two items). That is, a common process—directing the gaze to the chosen item—may explain the allocation of gaze both immediately before and after the choice report. In contrast, models that posit an exclusively intra-decision effect of attention on choice must then explain the allocation of gaze after the response as a separate process (e.g., directing gaze to the chosen item), or by assuming that individuals continue to evaluate the decision alternatives even after the choice report. Parameter-sensitivity analysis In the PDG model, the gaze–choice association arises from the temporal gap between manual and oculomotor response latencies, that is, from the difference τm\begin{document}$\tau_{\mathrm{m}}$\end{document}- τe\begin{document}$\tau_{\mathrm{e}}$\end{document}. In our simulations, τe\begin{document}$\tau_{\mathrm{e}}$\end{document} is modeled as a normal distribution with mean μe=0.35\begin{document}$\mu_{e}=0.35$\end{document} s and standard deviation σe=μe/3\begin{document}$\sigma_{e}=\mu_{e}/3$\end{document}, truncated at zero to exclude negative values. To assess the sensitivity of the model to this assumption, we varied μe\begin{document}$\mu_{e}$\end{document} across a range and recomputed the key behavioral signatures. Figure 8 shows the resulting gaze-cascade effect, the association between last-fixation focus and choice, and the interaction between ΔDwell and consistency, paralleling the analyses reported in Figure 6E–G. Sensitivity of the gaze–choice association in the PDG model. Simulations of the PDG model across a range of eye-movement latency values (μe\begin{document}$\mu_{e}$\end{document}), assessing the robustness of the model’s predictions. (A) Predicted gaze-cascade effect. Similar analysis to that in Figure 6E. (B) Predicted association between last-dwell focus and choice. Similar analysis to that in Figure 6F. (C) Predicted ΔDwell (chosen minus unchosen) for consistent and inconsistent choices as a function of response time. Similar to the analysis in Figure 6G. As μe\begin{document}$\mu_{e}$\end{document} increases, the probability that gaze is already directed toward the chosen item at the time of response decreases (Figure 8A). Because both τe\begin{document}$\tau_{\mathrm{e}}$\end{document} and τm\begin{document}$\tau_{\mathrm{m}}$\end{document} are variable, longer mean eye-movement latencies reduce the likelihood that the eye movement is completed before the manual response. Accordingly, the predicted association between gaze and choice weakens as μe\begin{document}$\mu_{e}$\end{document} increases (Figure 8B). The magnitude of ΔDwell also declines with increasing μe\begin{document}$\mu_{e}$\end{document} (Figure 8C). However, for all values of μe\begin{document}$\mu_{e}$\end{document}, ΔDwell remains positive and does not differentiate between consistent and inconsistent choices. Thus, the qualitative predictions of the PDG model are robust to the precise value assumed for τe\begin{document}$\tau_{\mathrm{e}}$\end{document}. In the ‘Discussion’, we explain why the comparatively large best-fitting values of τe\begin{document}$\tau_{\mathrm{e}}$\end{document} do not necessarily conflict with shorter post-decisional eye-movement latencies reported in neurophysiological studies of monkeys (Roitman and Shadlen, 2002). Limitations of the PDG model Despite the ability of the PDG model to explain many aspects of the data, other aspects are not as well captured. In the PDG model, choice accuracy is unaffected by fixation patterns prior to crossing the decision threshold. That is, whether decision-makers focus more on the higher-value item or distribute their attention evenly between the options, the model’s predictions for choice and RT remain unchanged. This invariance arises from the lack of a causal relation between gaze and the decision-making process. As a result, the PDG model makes no concrete predictions for the gaze pattern before choice commitment. In simulated data—where dwell durations are independent of the value of the alternatives—the PDG model fails to explain certain experimental findings that demonstrate an association between dwell duration and choice (Krajbich et al., 2010; Callaway et al., 2021). We highlight this limitation with two analyses. The first examines the likelihood of choosing the item on the left as a function of the difference in dwell time between the left and right items. Both empirical data and PDG model simulations reveal a positive association, but this relation is steeper in the data (Figure 6H). That is, ΔDwell is more predictive of choice in the experimental data than in the model, suggesting that decision-makers may fixate longer on the eventually chosen item even before covert choice commitment. The PDG model also fails to capture the relation between the duration of the first dwell and choice. Figure 6I shows the proportion of trials in which decision-makers selected the initially fixated option as a function of the first dwell duration. This relation is steeper in the data than in the model. The PDG model does predict a positive relation between the duration of the first dwell and choice, but this is only because the bound is crossed during the first dwell on some trials. On trials with more than one dwell, the PDG model predicted no significant association between first-dwell duration and choice probability (all p>0.05, likelihood-ratio tests; Equation 14, Figure 9A). In contrast, the empirical data revealed a significant positive association for shorter dwell-count conditions. Specifically, longer first-dwell durations significantly increased the probability of choosing the first-fixated item in two-dwell (p=0.013) and three-dwell trials (p=0.026), whereas this relationship was not significant for four-dwell (p=0.72) or five-dwell trials (p=0.96; Figure 9B). Association between first-dwell duration and choice probability. (A) Model-predicted probability of choosing the option that was fixated first, as a function of the duration of the first dwell. Trials are grouped by the total number of dwells (2–5), shown in separate colors. Data were binned into quartiles of first-dwell duration and then averaged across participants. Error bars indicate s.e.m. (B) Same analysis as (A), for the behavioral data. Predictions of other decision-making models For comparison with the PDG model, we analyze different decision-making models fit to the data of Krajbich et al., 2010. These models include variants of the aDDM (additive attention, inter-trial drift-rate variability, and post-decisional latencies) as well as pseudo-optimal models of decision-making. Additive intra-decision attention We fit an additive variant of the aDDM to the data from Krajbich et al., 2010, using identical fitting procedures. In this additive intra-decisional attention model, the decision variable evolves according to(6)xt=xt-1+d⁢(rleft-rright+s⁢ω)+ϵt,\begin{document}$$\displaystyle x_{t}=x_{t-1}+d(r_{\mathrm{left}}-r_{\mathrm{right}}+s\omega)+\epsilon_{t},$$\end{document} where s=+1\begin{document}$s=+1$\end{document} when attention is directed to the left item and s=−1\begin{document}$s=-1$\end{document} when it is directed to the right item. In this formulation, the drift rate shifts by +ω\begin{document}$+\omega$\end{document} or −ω\begin{document}$-\omega$\end{document} depending on which item is attended, in contrast to the aDDM, where the unattended item’s value is discounted multiplicatively. The additive model, just as the multiplicative variant (Figure 10A), fails to capture qualitative aspects of the data, most notably the interaction between choice consistency and ΔDwell (Figure 10B). Fits of different models to the data of Krajbich et al., 2010. Same analyses and conventions as in Figure 6. (A) Data and fits of the aDDM model. (B) aDDM with additive instead of multiplicative attention. (C) aDDM with post-decision attention. (D) aDDM with inter-trial variability in the drift rates. (E) aDDM with inter-trial variability in the drift rates and post-decision attention. (F) Model from Callaway et al., 2021. (G) Model from Jang et al., 2021. Figure 10—figure supplement 1.Model with intra- and post-decisional attention, and inter-trial variability in the items’ value.Unlike the model illustrated in Figure 10D, here the items’ values–instead of the drift rates–are corrupted with additive Gaussian noise. This noise is fixed within each trial but varies randomly across trials. Model parameters were fit to individual participants’ choice and RT. Same conventions as in Figure 6. Model with intra- and post-decisional attention, and inter-trial variability in the items’ value. Unlike the model illustrated in Figure 10D, here the items’ values–instead of the drift rates–are corrupted with additive Gaussian noise. This noise is fixed within each trial but varies randomly across trials. Model parameters were fit to individual participants’ choice and RT. Same conventions as in Figure 6. Figure 10—figure supplement 2.aDDM model simulations generated using the aDDM with the best-fitting parameters reported by Krajbich et al., 2010 and Smith and Krajbich, 2019.This version of the model assumes no inter-trial variability in drift rate. Unlike the models presented in the main text, this version of the aDDM was fit by Krajbich et al., 2010 to data pooled across participants. Same conventions as in Figure 6. aDDM model simulations generated using the aDDM with the best-fitting parameters reported by Krajbich et al., 2010 and Smith and Krajbich, 2019. This version of the model assumes no inter-trial variability in drift rate. Unlike the models presented in the main text, this version of the aDDM was fit by Krajbich et al., 2010 to data pooled across participants. Same conventions as in Figure 6. Combined intra- and post-decision attention Because neither a purely intra-decision nor a post-decision account of the gaze–choice association fully captures the behavioral data, we considered the possibility that a combination of the two accounts provides a better explanation (Westbrook et al., 2020). To this end, we combined elements from the aDDM and PDG models. In this variant of the model, prior to committing to a choice, decision dynamics follow the aDDM framework—the value of the unattended option is multiplicatively discounted and evidence accumulation terminates upon reaching a threshold at ±B\begin{document}$\pm B$\end{document}. Following threshold crossing, the gaze shifts to the covertly chosen item after a delay τe\begin{document}$\tau_{\mathrm{e}}$\end{document}, as described in the PDG model. Incorporating these post-decisional gaze shifts into the aDDM (Figure 10C) improves the model’s ability to explain the dependency between the last dwell, overall value, and choice (compare Figure 10A and C, columns I and IV). However, this hybrid approach still (i) inaccurately predicts that the gaze bias is stronger for inconsistent than for consistent choices (Figure 10C, column V), and (ii) displays a weaker association between the duration of the first fixation and choice, compared to what is observed in the data (Figure 10C, column VII). From this we conclude that the combination of intra- and post-decision attentional mechanisms does not fully account for the behavioral data. Drift-rate variability across trials None of the models that treat attention as an intra-decisional process were able to account for the difference in gaze bias between choices consistent versus inconsistent with stated preferences. We considered the possibility that inter-trial variability in the drift rate (e.g., Ratcliff and McKoon, 2008) could explain this discrepancy. To test this, we fit a variant of the aDDM in which additive Gaussian noise corrupts the drift rate on each trial. The behavioral data and model fits are shown in Figure 10D. The model still incorrectly predicts a larger gaze bias for inconsistent choices than for consistent choices. It also failed to capture the relation between first-dwell duration and choice. To assess whether post-decisional processes might improve the model’s match to the data, we incorporated a post-decision gaze shift mechanism, similar to that in the PDG model. Even with this post-decisional mechanism, the model predicts a larger gaze bias for inconsistent than for consistent choices, unlike what is observed in the data, and still fails to capture the observed association between first-dwell duration and choice (Figure 10E). We also explored an alternative implementation of inter-trial drift-rate variability, in which the item values—rather than the drift rate—are perturbed by additive Gaussian noise that remains constant within a trial. This approach is intended to capture the possibility that the values reported during the rating phase may differ from those used in the choice phase (Zylberberg et al., 2024). However, this variant did not provide a better fit to the behavioral data than the previous implementation (Figure 10—figure supplement 1). Optimal gaze allocation The allocation of attention in the aDDM is exogenous to the decision process—that is, attention shifts between items independently of the internal dynamics of decision-making. In contrast, more recent studies model the control of attention as endogenous, arising from an optimization that balances the cost of delaying the decision and collecting additional evidence, with the expected benefit of improved accuracy (Callaway et al., 2021; Jang et al., 2021; Song et al., 2019; Hébert and Woodford, 2018; Gluth et al., 2026; Gluth et al., 2020; Zhu, 2022). For instance, Callaway et al., 2021 formalized the decision process as a partially observable Markov decision process (POMDP), and approximated its solution by assuming that the value of additional deliberation can be expressed as a linear combination of factors such as the expected reward from acquiring an extra evidence sample and the expected reward if the item values were perfectly known. We asked whether the optimal attentional allocation model of Callaway et al., 2021 would show the same discrepancies with the behavioral data as the aDDM. To this end, we reanalyzed the simulations of the optimal model performed by Callaway and colleagues. Like the aDDM, the model of Callaway et al., 2021 incorrectly predicts a strong effect of overall value on the strength of the association between attention and choice (Figure 10F). This is clearly counter to what is observed in the data (Figure 3). We also analyzed the association between ΔDwell and choice consistency (Figure 10F). The model of Callaway et al., 2021 predicts that ΔDwell depends on choice consistency, with greater ΔDwell for inconsistent choices. The explanation is the same as for the aDDM: if attention has a causal influence on choice, then for inconsistent choices it is highly likely that attention was directed to the item of lower value for a substantial fraction of the trial. Again, this is in clear contrast to what we observe in the data (Figure 4). We also performed simulations of the ‘optimal’ model of Jang et al., 2021, using the best-fitting parameters reported in their study (Figure 10G). The model predicts that ΔDwell changes sign as a function of response time, contrary to what is observed in the data. For example, at response times of ~1.5 s, it incorrectly predicts that people will tend to choose the option they looked at the least. The model of Jang et al., 2021 also predicts that the influence of the last fixation on choice depends on Σr\begin{document}$\Sigma r$\end{document}, unlike what we observe in the data. We conclude that neither the aDDM nor the POMDP-based models account for the aspects of the data analyzed in Figures 3 and 4. Discussion We found post-decision signatures in the choice-RT-gaze data, suggesting that the association between gaze and choice, especially late in the trial, is not exclusively formative. Models positing an intra-decision (i.e., formative or constructive) multiplicative effect of attention on value—such as the attentional drift-diffusion model (aDDM)—predict that (i) the last-fixation bias should be stronger when the items under consideration are overall more desirable, and (ii) ΔDwell—the difference in time spent looking at the chosen versus unchosen item—should be greater for inconsistent than for consistent choices. This is because inconsistent choices, in such models, benefit disproportionately from the attentional amplification of value. The data do not support these predictions (Figures 3 and 4). Instead, these observations are better explained by a post-decision account of the gaze–choice association—that is, one in which gaze shifts to the selected item after a covert commitment to a choice. We argue that directing gaze to the chosen item after a covert choice commitment is sensible, as the benefits of attending to a stimulus do not end with the decision itself. In naturalistic settings, for instance, selecting a food item is typically followed by the action of reaching toward it, where visual attention supports spatial localization and motor planning for the upcoming action. Although participants in our computerized task did not physically act on their choices, these sensorimotor processes are likely highly automatized and may still be engaged by default, even when not strictly required. Beyond motor preparation, post-decisional attention may also serve additional functions, such as facilitating sensory anticipation of the reward, supporting metacognitive evaluation of the decision, and contributing to value updating for future choices. From this perspective, a degree of attentional ‘stickiness’—whereby the chosen item remains preferentially attended after commitment—could emerge as an effectively optimal policy once these post-decisional processes are taken into account. Moreover, a specific feature of the task design may further reinforce this tendency: in the snacks paradigm, the unchosen item typically disappears from the screen immediately after a response is registered. It is therefore plausible that directing gaze to the chosen item after commitment partly reflects anticipation of the imminent disappearance of the unchosen option. To disentangle these mechanisms, it would be interesting for future work to test whether this attentional bias persists when the chosen item, rather than the unchosen one, is the stimulus that disappears upon response. The PDG model captures many empirical observations, including the last-fixation bias and the gaze cascade effect. To account for the effect of overall value (Σr\begin{document}$\Sigma r$\end{document}) on RT (Smith and Krajbich, 2019), the PDG model relies on the assumption that the variability of value representations increases with the values themselves (Ratcliff et al., 2018). This assumption is supported by neurobiological evidence. In the cortex, more desirable options tend to evoke higher firing rates (Platt and Glimcher, 1999; Padoa-Schioppa and Assad, 2006), and neural signal variance typically scales with its mean (Tolhurst et al., 1981). If the two value representations are independent, the variance of the momentary evidence, Δr\begin{document}$\Delta r$\end{document}, should therefore increase in proportion to Σr\begin{document}$\Sigma r$\end{document}. In contrast, the aDDM explains the same relation between overall value and RT more directly—as a consequence of the multiplicative effect of attention on value—and thus does so more parsimoniously (Krajbich et al., 2010). Moreover, empirical results show that higher overall value leads to both faster and more accurate choices, arguing against the notion that sensitivity decreases with value. Instead, participants appear to invest additional effort in high-value decisions, which may offset any increase in variability (Shevlin et al., 2022; Ting and Gluth, 2025). A limitation of the PDG model is that it underestimates the observed association between both early dwell-time and choice probability, as well as the association between ΔDwell and choice probability. Inspired by Westbrook et al., 2020, we therefore examined a hybrid model in which attention initially exerts a causal influence on the choice process and subsequently reflects the chosen option. This mixed model reintroduced some of the same limitations seen in the aDDM—most notably, its inability to account for the null effect of overall value on the last-fixation bias (MELFB) and the similarity in gaze bias across consistent and inconsistent decisions. We conclude that neither of these models is fully able to account for the data from the food-choice task. The key assumption of the PDG model is that there is a delay between the moment a choice is internally committed and the moment it is externally reported with a key press. Because eye movements are typically faster than manual responses (τe\begin{document}$\tau_{\mathrm{e}}$\end{document} < τm\begin{document}$\tau_{\mathrm{m}}$\end{document} in our simulations), this delay creates a window during which gaze can already be directed toward the covertly chosen item before the response is formally registered. We do not interpret these non-decision latencies as irreducible physiological minima for moving the eyes or pressing a button (Bompas et al., 2025). Rather, they are inferred indirectly by fitting an additive non-decision-time parameter to the behavioral data, which we decompose into a sensory delay (τs\begin{document}$\tau_{\mathrm{s}}$\end{document}) and a manual execution delay (τm\begin{document}$\tau_{\mathrm{m}}$\end{document}). Values of τe\begin{document}$\tau_{\mathrm{e}}$\end{document} are then chosen so that the model reproduces the observed magnitude of the behavioral effects. This estimation procedure has important limitations. Some participants show relatively ‘flat’ chronometric functions: response times vary little with value despite otherwise normal psychometric performance. Such patterns likely reflect processes not explicitly represented in the model, including procrastination, reduced motivation, task-unrelated thought, or noise in item ratings. Within a drift-diffusion framework, however, these cases are accommodated by assigning a long non-decision time together with a short evidence-accumulation period (Appendix 1—table 1). Consequently, some estimated non-decision times are substantially longer than would be expected if they represented only sensory and motor delays. A further limitation is conceptual. We model non-decision time as occurring either before or after evidence accumulation, whereas in reality decisional and non-decisional components are likely temporally interleaved (Graziano et al., 2011). This simplification may also inflate the recovered latency estimates. With these caveats in mind, sensory and oculomotor delays on the order of 300 ms remain broadly plausible, although they likely lie near the upper end of a realistic range. The estimated eye-movement latency is especially long. For instance, in monkeys trained to report simple perceptual decisions with a saccade, roughly 100 ms elapses between the threshold-crossing signal in parietal cortex (or the superior colliculus) and the executed eye movement (Roitman and Shadlen, 2002; Stine et al., 2023). Crucially, however, varying the assumed non-decision latencies across a reasonable range does not alter the qualitative predictions of the model (Figure 8). Similar shortcomings to those observed in the aDDM were observed in models that derive attention-choice associations from optimal or near-optimal policies. Specifically, Callaway et al., 2021 and Jang et al., 2021 formalized the decision process as a partially observable Markov decision process (POMDP), in which attention enhances either the quantity (Callaway et al., 2021) or quality (Jang et al., 2021) of evidence about item value. To explain the gaze–choice association, both models assume that priors over item values are miscalibrated, such that decision-makers underestimate the true values. As a result, less-attended items are more biased toward zero, since less-attended items are more influenced by the miscalibrated prior. However, simulations based on these models fail to reproduce key empirical findings, including the association between ΔDwell, decision consistency, and RT (Figure 10F and G). The possibility that the gaze–choice association is partially post-decisional may help reconcile discrepancies between studies using free-viewing paradigms and those employing causal manipulations of attention. Common approaches to manipulating attention include limiting exposure duration (Frömer et al., 2022), interrupting trials based on gaze duration (Pärnamets et al., 2015; Newell and Le Pelley, 2018; Tavares et al., 2017; Pleskac et al., 2023), and cueing spatial attention (Störmer and Alvarez, 2016; Gwinn et al., 2019). A meta-analysis of the effects of visual attention on binary consumer choice (Bhatnagar and Orquin, 2022) found that these manipulations typically shift choice probabilities by ~2–4% from a 50% baseline (cf. Tavares et al., 2017). These small effects contrast sharply with the large effects estimated by fits of the aDDM, which often posit a 70% discount of unattended items (Krajbich et al., 2010). The discrepancy between model predictions and the results of the attention-manipulation studies may arise from attempting to account for post-decision gaze–choice correlations using models that assume that the gaze–choice association is exclusively intra-decisional. A post-decision gaze–choice association may offer a parsimonious explanation for differences in gaze behavior between tasks in which items are either chosen or rejected. In ‘choose’ tasks, participants select the preferred item; in ‘reject’ tasks, they exclude the less preferred item. Although logically equivalent in binary choice, the gaze is directed more to the preferred item in ‘choose’ tasks and to the non-preferred item in ‘reject’ tasks (Sepulveda et al., 2020; van der Laan et al., 2015; Mitsuda and Glaholt, 2014; Nittono and Wada, 2009). This difference has been explained by a variant of the aDDM in which attention modulates the integration of goal-relevant evidence (Sepulveda et al., 2020). Without ruling out this possibility, our results suggest that the difference in gaze allocation between ‘choose’ and ‘reject’ tasks arises because toward the end of the trial the gaze is directed to the selected option, regardless of whether it is to be accepted or rejected. In this work, we focused on a simple yet widely used class of models in which decisions are based on comparing noisy value signals assigned to each item, and where explicit ratings are assumed to reflect the true underlying value of those items. An alternative class of models proposes that decisions arise from comparisons along individual feature dimensions, rather than at the level of the item as a whole (Tversky, 1972; Roe et al., 2001; Usher and McClelland, 2004; Busemeyer and Townsend, 1993; Summerfield and Tsetsos, 2015; Shadlen and Shohamy, 2016; Lichtenstein and Slovic, 2006; Johnson et al., 2007; Lee and Hare, 2023; Yang and Krajbich, 2023; Fisher, 2021). In food-choice tasks, for example, relevant features might include expected satiety, caloric content, tastiness, saltiness, and so on (Rangel and Hare, 2010; Sullivan et al., 2015; Rramani et al., 2020; Suzuki et al., 2017). Attention is thought to fluctuate across these features, updating a decision variable for each alternative as different dimensions are sampled. Which dimensions are evaluated—and the weight assigned to each—may depend on the items being compared, past experience, or the broader decision context (Noguchi and Stewart, 2018; Roe et al., 2001; Juechems and Summerfield, 2019; Zylberberg et al., 2024; Lee and Pezzulo, 2022; Trueblood et al., 2014; Bhatia, 2013). As a result, the desirability of an item during the decision process may differ from that reported during the rating phase. A choice may appear inconsistent relative to initial ratings, but not relative to the specific features that were actually attended during deliberation. Similarly, if the ratings themselves are noisy and do not perfectly reflect subjective value, then some choices labeled as inconsistent may, in fact, be consistent with the decision-maker’s true preferences. Incorporating noise into the rating process would blur the distinction between consistent and inconsistent choices, potentially improving the alignment between model and data. These speculative ideas remain to be tested in future work. Overall, our findings suggest that the association between attention and choice is not fully captured by models in which attention plays a purely causal role during decision formation. The data suggest an additional post-decision association, in which attention reflects, rather than shapes, the covert choice. While this does not preclude a formative influence of attention at earlier stages, it highlights the importance of considering the temporal dynamics of commitment when interpreting gaze patterns. Methods Food-choice task We reanalyzed previously published data from six food-choice studies (Krajbich et al., 2010; Smith and Krajbich, 2018; Chen and Krajbich, 2016; Gwinn and Krajbich, 2016; Folke et al., 2016; Sepulveda et al., 2020). Across studies, the task followed the same two-phase paradigm. In the first (rating) phase, participants evaluated snack food items individually, either by rating how much they would like to consume each item on a numerical scale (Krajbich et al., 2010; Smith and Krajbich, 2018; Chen and Krajbich, 2016; Gwinn and Krajbich, 2016) or by stating their maximum willingness to pay using an incentive-compatible Becker–DeGroot–Marschak auction (Folke et al., 2016; Sepulveda et al., 2020). In the second (choice) phase, participants were presented with pairs of previously rated items and asked to choose which one they would prefer to consume at the end of the experiment, indicating their response with a key (or button) press. The studies differed in several design details. The number of items rated ranged from 70 (Krajbich et al., 2010) to 147 (Smith and Krajbich, 2018; Gwinn and Krajbich, 2016). Items with negative ratings were excluded from the choice phase in studies using a liking scale. The maximum permitted difference in value between the two items in a pair varied across studies: five rating points in Krajbich et al., 2010 and Smith and Krajbich, 2018, three in Chen and Krajbich, 2016, and one in Gwinn and Krajbich, 2016; no such constraint was applied in Folke et al., 2016 and Sepulveda et al., 2020. The number of choice trials per participant ranged from 100 (Krajbich et al., 2010) to 240 (Sepulveda et al., 2020). In most studies, participants had to fixate on a central marker before each trial began. Further details of each experiment can be found in the original publications and in Appendix 1—table 5. PDG model In the PDG model, choice and decision time are determined by the state of a scalar decision variable, x\begin{document}$x$\end{document}. The decision variable is the cumulative sum of samples from a normal distribution with mean μdt\begin{document}$\mu dt$\end{document} and variance σ2dt\begin{document}$\sigma^{2}dt$\end{document}, and thus its evolution is described by the stochastic differential equation,(7)dx=μdt+σdW,\begin{document}$$\displaystyle dx = \mu dt + \sigma dW,$$\end{document} where W\begin{document}$W$\end{document} is the standard Wiener process and x(t=0)=0\begin{document}$x(t=0)=0$\end{document}. The drift rate µ is given by(8)μ=κ(rleft−rright),\begin{document}$$\displaystyle \mu=\kappa(r_{\mathrm{left}}-r_{\mathrm{right}}),$$\end{document} where rleft\begin{document}$r_{\mathrm{left}}$\end{document} and rright\begin{document}$r_{\mathrm{right}}$\end{document} are the ratings assigned to the items presented on the left and right of the screen. The variance of the momentary evidence, σ2\begin{document}$\sigma^{2}$\end{document}, scales linearly with the sum of the ratings (Equation 5). The accumulation process ends when the decision variable, x(t)\begin{document}$x(t)$\end{document}, reaches one of two bounds positioned symmetrically around zero, at ±B\begin{document}$\pm B$\end{document}. The left (right) item is chosen when the decision variable reaches the upper (lower) bound. This first passage time establishes the decision time. The RT is the sum of the decision time plus the non-decision latencies, τnd\begin{document}$\tau_{\mathrm{nd}}$\end{document}. We assume that τnd\begin{document}$\tau_{\mathrm{nd}}$\end{document} is normally distributed with mean μnd\begin{document}$\mu_{\mathrm{nd}}$\end{document} and standard deviation σnd\begin{document}$\sigma_{\mathrm{nd}}$\end{document}. The accumulation process starts after a short sensory delay, τs\begin{document}$\tau_{\mathrm{s}}$\end{document} (Figure 5), which is part of the non-decision latency. We used a fixed value of τs=0.3\begin{document}$\tau_{\mathrm{s}}=0.3$\end{document} s for all participants. The remaining non-decision time (τm=τnd−τs\begin{document}$\tau_{\mathrm{m}}=\tau_{\mathrm{nd}}-\tau_{\mathrm{s}}$\end{document}) is assigned to the time between crossing a decision bound and reporting the choice. After crossing a bound, gaze is directed to the selected item. The time taken to switch gaze to the chosen item has an associated non-decision latency of τe\begin{document}$\tau_{\mathrm{e}}$\end{document}. This latency is assumed to follow a truncated normal distribution with a mean of μe=0.35\begin{document}$\mu_{e}=0.35$\end{document} seconds and a standard deviation σe\begin{document}$\sigma_{e}$\end{document} of one-third of the mean, truncated to ensure non-negative values. Since τe\begin{document}$\tau_{\mathrm{e}}$\end{document} is usually smaller than τm\begin{document}$\tau_{\mathrm{m}}$\end{document}, the gaze is informative about the choice. We fit the model parameters λ={κ,B,γ,μnd,σnd}\begin{document}$\lambda=\{\kappa,B,\gamma,\mu_{\mathrm{nd}},\sigma_{\mathrm{nd}}\}$\end{document} using maximum likelihood to the choice and RT data from each trial. Fits were performed independently for each participant. The log-likelihood of the parameters is given by(9)log⁡Lλ=∑i=1nlog⁡p⁢(choice(i),RT(i)|Δ⁢r(i),Σ⁢r(i),λ),\begin{document}$$\displaystyle \log L_{\lambda}=\sum_{i=1}^{n}\log p(\mathrm{choice}^{(i)},\mathrm{RT}^{(i)}| \Delta r^{(i)},\Sigma r^{(i)},\lambda),$$\end{document} where the summation is over trials. The joint probability of choice and decision time was obtained by numerically solving the Fokker–Planck (FP) equation associated with the drift-diffusion process using the Chang–Cooper fully implicit method (Chang and Cooper, 1970; Zylberberg et al., 2016; Kiani and Shadlen, 2009). To obtain the joint probability distribution over choice and RT, we convolve the probability distribution of decision times with the distribution of non-decision times, given by a truncated normal distribution with parameters μnd\begin{document}$\mu_{\mathrm{nd}}$\end{document} and σnd\begin{document}$\sigma_{\mathrm{nd}}$\end{document} (the truncation constrains the non-decision times to be positive). Attentional drift-diffusion model We simulated the aDDM (Equations 1; 2) using the best-fitting parameters reported by Krajbich et al., 2010: d\begin{document}$d$\end{document} = 0.0002 ms-1, σ\begin{document}$\sigma$\end{document} = 0.02, θ\begin{document}$\theta$\end{document} = 0.3 and bounds B=±1\begin{document}$B=\pm 1$\end{document}. The response time was calculated as the sum of the decision time from the drift-diffusion process and a fixed non-decision time of tnd=0.355\begin{document}$t_{\mathrm{nd}}=0.355$\end{document} s, as in Smith and Krajbich, 2019. In the aDDM (as well as in the DDM), it is necessary to set one of the parameters d\begin{document}$d$\end{document}, σ\begin{document}$\sigma$\end{document}, or B\begin{document}$B$\end{document} to a fixed value for the model parameters to be identifiable. In the original description of the aDDM (Krajbich et al., 2010), the upper and lower bounds were set to a fixed value of ±1. For consistency with the PDG model, we reformulated the aDDM so that the variance of the noise accumulated during one second of unbounded accumulation is equal to 1. Then the decision variable of the aDDM, x(t)\begin{document}$x(t)$\end{document}, evolves as in Equation 7 with σ=1\begin{document}$\sigma=1$\end{document}. Time is discretized in steps of dt=0.001\begin{document}$\mathrm{dt}=0.001$\end{document} s. The drift rate µ is equal to κ(rleft−θrright)\begin{document}$\kappa(r_{\mathrm{left}}-\theta r_{\mathrm{right}})$\end{document} when looking at the item on the left, and to κ(θrleft−rright)\begin{document}$\kappa(\theta r_{\mathrm{left}}-r_{\mathrm{right}})$\end{document} when looking at the item on the right, where κ\begin{document}$\kappa$\end{document} is the signal-to-noise. A model equivalent to the original aDDM is obtained with the following parameters:(10)κ=dσ1000=0.3162B=±1σ11000=±1.5811θ=0.3.\begin{document}$$\displaystyle \begin{array}{ll}\kappa &= \frac{d}{\sigma}\sqrt{1000}= 0.3162 \\ B&=\pm \frac{1}{\sigma}\frac{1}{\sqrt{1000}}= \pm 1.5811 \\ \theta&= 0.3.\end{array}$$\end{document} Simulating the aDDM requires modeling how attention alternates between the two items. We fit the empirically observed dwell durations with log-normal distributions. Separate fits were conducted for the first and middle dwells. Middle dwells are those that were neither the first nor the last of the stimulus-viewing epoch. The fits provide a good match to the experimental data (Figure 6—figure supplement 2). Each simulated trial of the aDDM begins by sampling from the first dwell distribution, followed by sampling from the middle dwell distribution until a decision threshold is reached. Similar to the experimental data, the first fixation has a 0.74 probability of being directed to the left item, and attention alternates between the two items thereafter. We simulated the same trials (same value pairs) as those completed by the participants, repeating each trial 10 times. We fit three variants of the aDDM to individual participant data. In the first variant, attention has an additive effect on value, rather than a multiplicative effect (Figure 10B; see equations for the drift rate in the main text). In the second variant (Figure 10D), the drift rate exhibits inter-trial variability. Specifically, on each trial, the drift rate µ is given by κ(rleft−θrright)+ν\begin{document}$\kappa(r_{\mathrm{left}}-\theta r_{\mathrm{right}})+\nu$\end{document} when fixating the left item, and by κ(θrleft−rright)+ν\begin{document}$\kappa(\theta r_{\mathrm{left}}-r_{\mathrm{right}})+\nu$\end{document} when fixating the right item. Here, ν\begin{document}$\nu$\end{document} is trial-specific noise drawn from a normal distribution with mean 0 and standard deviation σd\begin{document}$\sigma_{d}$\end{document}, which is estimated from the data. Importantly, ν\begin{document}$\nu$\end{document} is constant within a trial but varies across trials. The third aDDM variant (Figure 10—figure supplement 1) can be interpreted as an alternative implementation of the one with inter-trial drift-rate variability. The noise does not directly affect the drift rate, but the items’ value. Specifically, the drift rate µ is given by κ((rleft+ν1)−θ(rright+ν2))\begin{document}$\kappa((r_{\mathrm{left}}+\nu_{1})-\theta(r_{\mathrm{right}}+\nu_{2}))$\end{document} when fixating the left item, and by κ(θ(rleft+ν1)−(rright+ν2))\begin{document}$\kappa(\theta(r_{\mathrm{left}}+\nu_{1})-(r_{\mathrm{right}}+\nu_{2}))$\end{document} when fixating the right item. Here, νx\begin{document}$\nu_{x}$\end{document} represents trial-specific noise, drawn from a normal distribution with mean 0 and standard deviation σd\begin{document}$\sigma_{d}$\end{document}. This formulation reflects the idea that the value reported in the rating phase may deviate from the item’s ‘true’ underlying value (Zylberberg et al., 2024; Polanía et al., 2019). Additionally, we simulated the aDDM using the best-fitting parameters reported in previous studies by Krajbich et al., 2010 and Smith and Krajbich, 2019 (Figure 10—figure supplement 2). The core parameters of the aDDM are λ={κ,B,θ}\begin{document}$\lambda=\{\kappa,B,\theta\}$\end{document}, corresponding respectively to the signal-to-noise ratio, the decision bound height, and the value-scaling factor applied to the unattended option. In the model with inter-trial drift-rate variability, σd\begin{document}$\sigma_{d}$\end{document} is added to capture the standard deviation of the drift-rate noise. The models are fit to maximize the likelihood of the parameters given the choice and decision time (DT\begin{document}$\mathrm{DT}$\end{document}), and given the sequence and duration of the dwells observed on each trial:(11)log⁡Lλ=∑i=1nlog⁡p(choice(i),DT(i)|rleft(i),rright(i),D(i),λ),\begin{document}$$\displaystyle \log L_{\lambda} = \sum_{i=1}^{n} \log p(\mathrm{choice}^{(i)}, \mathrm{DT}^{(i)}|r_{\mathrm{left}}^{(i)},r_{\mathrm{right}}^{(i)},\mathrm{D}^{(i)},\lambda), $$\end{document} where D(i)\begin{document}$\mathrm{D}^{(i)}$\end{document} is the sequence and duration of dwells observed on trial i\begin{document}$i$\end{document}, and DT(i)\begin{document}$\mathrm{DT}^{(i)}$\end{document} is the decision time on trial i\begin{document}$i$\end{document} that is assumed to be equal to the sum of the dwells on either of the two items. The joint probability of choice and decision time was computed via numerical approximations of the corresponding FP equation. Unlike the special case where θ=1\begin{document}$\theta=1$\end{document}, here the drift rate varies with the focus of attention. As a result, solving the FP equation becomes more computationally demanding, since the probability density of the decision variable must be computed separately for each trial, given the trial-specific stochastic fluctuations in gaze. Numerical solutions were obtained using a fully implicit method (Chang and Cooper, 1970), propagating the probability density of the decision variable over time adjusting the drift rate depending on the focus of gaze. For the model with inter-trial variability in the drift rate, we discretized the distribution of drift perturbations into a finite number of bins. Specifically, we drew nbins=11\begin{document}$n_{\mathrm{bins}}{=}11$\end{document} quantile-based samples from a zero-mean Gaussian distribution with standard deviation σdrift\begin{document}$\sigma_{\mathrm{drift}}$\end{document}, using the midpoint of each quantile as a representative value. We numerically solved the FP equation independently for each of these bins. The final probability distribution over choice and decision times was computed as the average (uniform-weighted) across the solutions for each bin, effectively marginalizing over the distribution of drift perturbations. We use the best-fitting parameters to simulate the aDDM independently for each participant. From the simulations we obtain a choice and decision time for each trial. Response times are estimated as the decision time plus mean non-decision time, μnd\begin{document}$\mu_{nd}$\end{document}, defined as the trial-average RT of each participant minus the trial-average decision time obtained from the model simulations. Model simulations Simulations of the PDG model (Figure 6) and aDDM (Figure 10A–E) were made using the same trials (same value pairs) that participants completed, with each trial repeated 10 times. Dwell durations were randomly sampled from log-normal distributions fit to the duration of the dwells (Figure 6—figure supplement 2). First and subsequent dwells were fit separately. The probability of sampling the left item first was set to 0.74 to match the value obtained from the experimental data. Combined aDDM-PDG model The combined aDDM-PDG model (Figure 10C) builds on the aDDM fit to single-participant data. We add to the data simulated with the aDDM (Figure 10A) two aspects of the PDG model: (i) a sensory delay between the onset of the food items and the start of the evidence accumulation process of τs\begin{document}$\tau_{\mathrm{s}}$\end{document}, and (ii) another delay τe\begin{document}$\tau_{\mathrm{e}}$\end{document} between the time at which a bound is crossed and the time at which the gaze is directed to the chosen item (truncated such that τe\begin{document}$\tau_{\mathrm{e}}$\end{document} is non-negative). Parameters τs\begin{document}$\tau_{\mathrm{s}}$\end{document}, μe\begin{document}$\mu_{e}$\end{document} and σe\begin{document}$\sigma_{e}$\end{document} were set to 0.25 s, 0.2 s and 0.05 s, respectively. The same approach and parameter values were used in the aDDM model with inter-trial variability in the drift rate (Figure 10E). Model fitting Parameter optimization was performed using the Bayesian Adaptive Direct Search method (Acerbi and Ma, 2017). Appendix 1—tables 1–4 show the best-fitting parameters for the PDG model, the aDDM, the model with intra-decisional additive attention, and the aDDM with inter-trial drift-rate variability, respectively. Optimal decision models The optimal model of Jang et al., 2021 has five free parameters: the cost of switching attention between items (cs\begin{document}$c_{s}$\end{document}), the cost, per second, of accumulating evidence (c\begin{document}$c$\end{document}), the variance of the evidence sampling distribution (σX2\begin{document}$\sigma_{X}^{2}$\end{document}), the variance of the prior distribution (σz2\begin{document}$\sigma^{2}_{z}$\end{document}), and the relative information gain for attended vs. unattended items (κ\begin{document}$\kappa$\end{document}). See Jang et al., 2021 for a detailed explanation of the model. We simulated 1280 trials per participant with the parameters that Jang et al., 2021 reported best replicated the human behavioral data: cs=0.0065\begin{document}$c_{s}=0.0065$\end{document}, c=0.23\begin{document}$c=0.23$\end{document}, σx2=27\begin{document}$\sigma_{x}^{2}=27$\end{document}, σz2=18\begin{document}$\sigma_{z}^{2}=18$\end{document}, κ=0.004\begin{document}$\kappa=0.004$\end{document}. As in Jang et al., the prior mean over the items’ values, z¯\begin{document}$\bar{z}$\end{document}, was set to zero, which is lower than the true mean value of the items. The model requires this feature to produce a gaze bias (Jang et al., 2021). We also analyzed the simulations of the optimal model developed by Callaway et al., 2021. The model has five free parameters: the standard deviation of the evidence sampling distribution, the cost of obtaining a sample, the cost of switching attention between items, the degree to which the prior over the item values is biased toward zero, and a ‘temperature’ parameter of a Boltzmann distribution, which controls the degree of stochasticity in the selection of the optimal policy. The simulations (N=4,550,400 trials) with the parameters that best fit the human behavioral data were kindly provided by Frederick Callaway and are available at https://github.com/fredcallaway/optimal-fixations-simple-choice, Callaway, 2023. A detailed explanation of the model and the fitting procedure can be found in the original publication (Callaway et al., 2021). Data analysis The dashed lines in Figure 3 (top) were derived from the following logistic regression model:(12)logit⁢[pleft]=∑s(βΔ⁢r,s⁢Δ⁢r+β0,s)⁢Is+βLastFix⁢ILastFix+βLastFix,Σ⁢r⁢ILastFix⁢Σ⁢r\begin{document}$$\displaystyle \mathrm{logit}[p_{\mathrm{left}}]=\sum_{s}\left(\beta_{\Delta r,s}\Delta r+\beta_{ 0,s}\right)I_{s}+\beta_{\mathrm{LastFix}}I_{\mathrm{LastFix}}+\beta_{\mathrm{ LastFix},\Sigma r}I_{\mathrm{LastFix}}\Sigma r$$\end{document} ILastFix\begin{document}$I_{\mathrm{LastFix}}$\end{document} takes a value of 1 if the left item was fixated on last and 0 if the right item was fixated last. The rightmost term captures the interaction of ILastFix\begin{document}$I_{\mathrm{LastFix}}$\end{document} with Σr\begin{document}$\Sigma r$\end{document}; the associated β is the slope of the dashed line in Figure 3 (top). We fit the following linear regression model to test for an association between choice consistency (c\begin{document}$c$\end{document}) and the difference in looking time between the chosen and unchosen items (ΔDwell):(13)ΔDwell=∑i=1NβiIi+∑i=1NβN+iIiRT+βcc,\begin{document}$$\displaystyle \Delta\mathrm{Dwell}= \sum_{i=1}^{N} \beta_{i} I_{i} + \sum_{i=1}^{N} \beta_{N+i}I_{i} RT + \beta_{c} c, $$\end{document} where N\begin{document}$N$\end{document} is the number of participants, Ii\begin{document}$I_{i}$\end{document} is an indicator variable that takes the value 1 if the trial was completed by subject i\begin{document}$i$\end{document} and 0 otherwise, and c\begin{document}$c$\end{document} is equal to 1 for trials in which the higher-rated item was chosen, defined as a consistent choice, and 0 for trials in which the lower-rated item was chosen, defined as an inconsistent choice. We used a one-tailed t-test to test whether βc\begin{document}$\beta_{c}$\end{document} is negative, that is, whether ΔDwell is larger for inconsistent than for consistent choices. When the use of a one-tailed test is not explicitly mentioned, we used two-tailed t-tests to test whether βc\begin{document}$\beta_{c}$\end{document} differed significantly from zero. To test the hypothesis that there was no difference in the probability of looking at the chosen item before and after the choice report (Figure 7), we used a Wilcoxon signed-rank test. We define two time epochs, one from –200 ms to 0 ms, and the other from 0 ms to 200 ms, relative to the time of the choice report. For each time epoch, we calculated the time each participant spent looking at the chosen item and divided it by the time spent looking at either item. We obtain two proportions per participant (Figure 7B), which we subjected to a two-tailed Wilcoxon signed-rank test. For the plots showing the probability of looking at the chosen item aligned to stimulus onset (e.g., Figure 6E, left), we eliminated the gaze information from the 500 ms prior to the choice report. This step was taken to eliminate, from the stimulus-aligned plots, gaze effects that might be related to the response. For the plots showing the psychometric function split by whether the last fixated item was the one on the left or right (e.g., Figure 6D), we classified trials as ‘left item fixated on last’ and ‘right item fixated on last’ depending on which item was being looked at at the time of the choice report. We excluded trials in which the participant either was not fixating on one of the two relevant items at the time of the choice report or the direction of gaze could not be resolved (e.g., eye blinks). We repeated the analyses reclassifying trials according to which item was last looked at before the choice report regardless of when it occurred during the trial, and obtained nearly identical results. To test the association between first-dwell duration and the probability of choosing the option that was fixated first (Figure 9), we fit the following logistic regression model, separately for each total–dwell–count condition (2–5 total dwells):(14)logit[pmatch]=β0+β1D+β2rfirst+β3rother\begin{document}$$\displaystyle \mathrm{logit}[p_{\mathrm{match}}] = \beta_{0} + \beta_{1} D + \beta_{2} r_{\mathrm{first}}+ \beta_{3} r_{\mathrm{other}}$$\end{document} where pmatch\begin{document}$p_{\mathrm{match}}$\end{document} is the probability of choosing the item fixated first, D\begin{document}$D$\end{document} is the standardized first-dwell duration, and rfirst\begin{document}$r_{\mathrm{first}}$\end{document} and rother\begin{document}$r_{\mathrm{other}}$\end{document} represent the value ratings of the first-fixated and alternative items, respectively. Statistical significance of the first-dwell-duration coefficient was assessed using a likelihood-ratio test, comparing the full model to a reduced model that excluded the β1\begin{document}$\beta_{1}$\end{document} term. To illustrate the gaze cascade effect (e.g., Figure 6E), we filled a matrix of dimensions NumberOfTrials × NumberOfTimeSteps, with a 1 when participants were looking at the right item, a 0 when they were looking at the left item, and a NaN otherwise. The time step was 1 ms. We averaged this matrix across trials, ignoring the NaNs, first within participants and then across participants. For the RT aligned plots, we followed the same procedure after aligning each trial to the response time. For the aDDM and Callaway’s optimal model, we followed a slightly different procedure because in these models the drift rate is undefined when participants are not looking at either of the two items. Therefore, we removed from each trial the times when the participant’s gaze was not directed to either snack item and aligned the responses to the total fixation time rather than to the response time. This study uses convincing modeling methods and analyses of rich behavioral datasets to investigate the role of attention in value-based decision-making; for instance, as when choosing between two snacks. The results are important, as they challenge existing theories that assume that paying attention to an available option biases the eventual choice toward that option. The results suggest that the correlation between attention and decision-making is formed largely after rather than before the (internal) choice process has terminated, a finding that offers an intuitively appealing rethinking of how attention and decision-making processes interact during value-based choices. [Editors' note: this version has been assessed by the Reviewing Editor without further input from the original reviewers. The authors have addressed the weaknesses raised in the previous round of review.] Summary: This study examines whether gaze direction actively shapes choice during food preference decisions or whether gaze and choice evolve largely independently until the moment of commitment. The established framework in this context, the aDDM, assumes that gaze causally biases the accumulation of evidence in favour of the fixated item. The authors show convincingly that this model fails to fit key behavioural patterns across several datasets, as do other published models that make the same assumption. The authors propose an alternative model (Post-Decision-Gaze or PDG) in which gaze and decision formation are decoupled: gaze does not influence the decision process, nor is it drawn toward the ultimately chosen item, until after the decision threshold is reached. Only during the motor execution period (after commitment) is gaze directed to the chosen option. They demonstrate that this model fits several observed patterns better than the aDDM and related variants. Strengths: The work thoroughly considers multiple models and datasets. It advances an interesting alternative perspective on gaze-decision interactions and highlights meaningful shortcomings in existing models. The authors take the time to explain how modelling assumptions produce specific patterns in the data, which is certainly insightful to readers interested in the modelling of value-based decision making. Weaknesses: It is unclear to what extent the model's success relies on the way non-decision time is formalised in the model. In the proposed PDG model, non-decision time is decomposed into separate visual encoding, saccadic execution, and manual execution components. Several values (assumed or recovered) do not match known physiological or behavioural ranges. This is a common issue in the literature, and the authors may want to address it in light of broader work discussing what non-decision time consists of in both manual and saccadic actions (e.g., Bompas et al., 2024, Non decision time: the Higgs boson of decision, Psychological Review). Summary: Zylberberg et al. reanalyze eye-tracking and behavioral data to test two predictions of the attentional Drift Diffusion Model, finding that these predictions are not met. Similarly, predictions of normative models (inspired by rational inattention) are not in line with the data, and the authors propose a post-choice model of attention. This model better accounts for the two effects but also does not account for all patterns, so the authors conclude that eye movements most likely reflect both pre- and post-decisional processes. Strengths: A clear strength is the systematic falsification-based approach of the paper, establishing (partially) new predictions and testing to what extent these are met by extant models and by a newly developed theory. The authors do a good job in providing intuitions behind the effects and the reasons why models such as the aDDM predict them. The paper is of substantial relevance for the field, as it shows that effects pertaining to the last fixation(s) should be interpreted with caution. Another strength is the paper's transparency as the authors clearly acknowledge that their new model does not do a perfect job either. Weaknesses: The paper focuses on analyzing the Krajbich 2010 data, but shows that the second effect replicates in many other datasets. A more principled approach, in which both effects are analyzed and presented for all datasets, would be more convincing. The results should then be shown together for clarity/readability. Similarly, it would be nice to show to what extent the models' predictions depend (not depend) on using the best-fitting parameter values (are there any parameter settings under which the two effects are not predicted?) Summary: In this study, the authors reanalyzed choice, RT and gaze datasets collected from human subjects performing a food-choice task. They show that models that posit a causal role for attention in shaping the decision-making process fail to account for empirical observations in the data. These include the attentional drift diffusion model (aDDM) and models that derive attention-choice associations from an optimal policy. The authors show that a model that assumes that gazes are directed towards the chosen option after decision commitment captures more (but not all) empirical findings, suggesting that attention may reflect decisions once they are made instead of contributing to their formation. However, this post-decision-gaze (PDG) model failed to capture all aspects of the data, suggesting that gaze may reflect both decisional and post-decisional operations, and existing models are still missing some features of the gaze-directing process. The authors provide convincing evidence that post-decision gaze explains a number of empirical findings in this task. Strengths: (1) The analyses are generally appropriate, and the conclusions are supported by the data. (2) The study was rigorous, as the authors considered a number of alternative possible models for behavior, and evaluated their performance based on a wide range of qualitative predictions (as opposed to exclusively relying on model comparison). (3) The proposal that gaze may largely reflect post-decisional processes is interesting, and as far as I am aware, novel. Weaknesses: There was limited discussion about why one might allocate attention post-decision. I would have appreciated more discussion on the potential functional consequences or implications of post-decision gaze. The following is the authors’ response to the original reviews. Public Reviews: Reviewer #1 (Public review): It is unclear to what extent the model's success relies on the way non-decision time is formalised in the model. In the proposed PDG model, non-decision time is decomposed into separate visual encoding, saccadic execution, and manual execution components. Several values (assumed or recovered) do not match known physiological or behavioural ranges. This is a common issue in the literature, and the authors may want to address it in light of broader work discussing what non-decision time consists of in both manual and saccadic actions (e.g., Bompas et al., 2024, Non decision time: the Higgs boson of decision, Psychological Review). In particular, the "saccadic execution" parameter appears far too long and too variable to reflect merely execution; instead, it likely includes decisional components. This would make more sense since manual and saccadic planning essentially rely on distinct brain areas, hence it seems unrealistic that crossing a single threshold would trigger both manual and saccadic execution. Similarly, recovered manual non-decision times are substantially longer (though not more variable) than expected motor execution durations for button presses. These patterns suggest that parts of what the model treats as non-decision time are likely decisional in nature, although perhaps related to "action decision" rather than the "value-based decision" of interest to the authors. To what extent these two processes neatly follow each other or overlap could be usefully considered. We have added a paragraph to the Discussion explaining how our model’s estimates of sensory and motor latencies relate to corresponding values inferred from physiology or behavioral manipulations (e.g., Bompas et al., 2024). Specifically, we write: “The key assumption of the PDG model is that there is a delay between the moment a choice is internally committed and the moment it is externally reported with a key press. Because eye movements are typically faster than manual responses (𝜏e < 𝜏m in our simulations), this delay creates a window during which gaze can already be directed toward the covertly chosen item before the response is formally registered. We do not interpret these non-decision latencies as irreducible physiological minima for moving the eyes or pressing a button (Bompas et al., 2025). Rather, they are inferred indirectly by fitting an additive non-decision-time parameter to the behavioral data, which we decompose into a sensory delay (𝜏s) and a manual execution delay (𝜏m). Values of 𝜏e are then chosen so that the model reproduces the observed magnitude of the behavioral effects. This estimation procedure has important limitations. Some participants show relatively “flat” chronometric functions: response times vary little with value despite otherwise normal psychometric performance. Such patterns likely reflect processes not explicitly represented in the model, including procrastination, reduced motivation, task-unrelated thought, or noise in item ratings. Within a drift-diffusion framework, however, these cases are accommodated by assigning a long non-decision time together with a short evidence-accumulation period (Table S1). Consequently, some estimated non-decision times are substantially longer than would be expected if they represented only sensory and motor delays. A further limitation is conceptual. We model non-decision time as occurring either before or after evidence accumulation, whereas in reality decisional and non-decisional components are likely temporally interleaved (Graziano et al., 2011). This simplification may also inflate the recovered latency estimates. With these caveats in mind, sensory and oculomotor delays on the order of 300 ms remain broadly plausible, although they likely lie near the upper end of a realistic range. The estimated eye-movement latency is especially long. For instance, in monkeys trained to report simple perceptual decisions with a saccade, roughly 100 ms elapses between the threshold-crossing signal in parietal cortex (or the superior colliculus) and the executed eye movement (Roitman and Shadlen, 2002; Stine et al., 2023). Crucially, however, varying the assumed non-decision latencies across a reasonable range does not alter the qualitative predictions of the model (Fig. 8).” Further, we have added a parameter sensitivity analysis. Importantly, although the magnitude of the predicted effects depend on the non-decision latencies, the qualitative aspect of these predictions do not (new Figure 8). Specifically, (i) the increasing tendency to look at the ultimately chosen item as time elapses (new Fig. 8A), (ii) the lack of an interaction between the last-fixation bias and overall value (Fig. 8B), and (iii) the absence of an effect of choice consistency on Δdwell (Fig. 8C) are all findings that are independent of 𝜏e. Reviewer #2 (Public review): The paper focuses on analyzing the Krajbich 2010 data, but shows that the second effect replicates in many other datasets. A more principled approach, in which both effects are analyzed and presented for all datasets, would be more convincing. The results should then be shown together for clarity/readability. Following this suggestion (and the reviewer’s elaboration below), we have substantially restructured the manuscript. Both aDDM predictions are now presented together (new Fig. 2), and Figs. 3–4 test these predictions across multiple food-choice datasets. In doing so, we no longer treat the data from Krajbich et al. (2010) separately, and we extend the analysis of the last-fixation–choice association (MELFB) to additional datasets. We note that the same datasets could not be used in both Figs. 3 and 4, as some lack information on the final fixation required for the MELFB analysis. Nevertheless, results are highly consistent across datasets and align with findings from a recent study by Ting & Gluth (2025), which independently identified and examined one of our key predictions; this work is now cited in the revised manuscript. Finally, to reduce redundancy, we have consolidated all aDDM variants and optimal models into a single figure (new Fig. 10). Similarly, it would be nice to show to what extent the models' predictions depend (not depend) on using the best-fitting parameter values (are there any parameter settings under which the two effects are not predicted?) The key predictions of the model depend on the difference between the manual (𝜏m) and eye-movement-related (𝜏e) latencies. We have now added a parameter-sensitivity analysis to show how the model predictions depend on this difference. The new analysis shows that while the quantitative predictions do depend on the precise latency values, the results are qualitatively similar across values of 𝜏e (new Figure 8). Reviewer #3 (Public review): There was limited discussion about why one might allocate attention post-decision. I would have appreciated more discussion on the potential functional consequences or implications of post-decision gaze. Thank you for this suggestion. We added a new paragraph to the discussion (paragraph #2), where we argue that it is sensible for a decision maker to direct the gaze to the chosen item once a covert choice commitment has been made, as the benefits of attending to a stimulus do not end with the decision itself. Specifically we now write: “Instead, these observations are better explained by a post-decision account of the gaze-choice association that is, one in which gaze shifts to the selected item after a covert commitment to a choice. We argue that directing gaze to the chosen item after a covert choice commitment is sensible, as the benefits of attending to a stimulus do not end with the decision itself. In naturalistic settings, for instance, selecting a food item is typically followed by the action of reaching toward it, where visual attention supports spatial localization and motor planning for the upcoming action. Although participants in our computerized task did not physically act on their choices, these sensorimotor processes are likely highly automatized and may still be engaged by default, even when not strictly required. Beyond motor preparation, post-decisional attention may also serve additional functions, such as facilitating sensory anticipation of the reward, supporting metacognitive evaluation of the decision, and contributing to value updating for future choices. From this perspective, a degree of attentional “stickiness” whereby the chosen item remains preferentially attended after commitment could emerge as an effectively optimal policy once these post-decisional processes are taken into account. Moreover, a specific feature of the task design may further reinforce this tendency: in the snacks paradigm, the unchosen item typically disappears from the screen immediately after a response is registered. It is therefore plausible that directing gaze to the chosen item after commitment partly reflects anticipation of the imminent disappearance of the unchosen option. To disentangle these mechanisms, it would be interesting for future work to test whether this attentional bias persists when the chosen item, rather than the unchosen one, is the stimulus that disappears upon response.” Recommendations for the authors: Reviewer #1 (Recommendations for the authors): Major Comments: (1) Framing of the modelling approach The manuscript would benefit from acknowledging the known limitations of DDM-based frameworks, especially given that the entire study is conducted within these constraints. The introduction highlights successes of the DDM, but the manuscript does not mention any of its conceptual or empirical limitations. We are unsure about what specific limitations the reviewer has in mind, but we have added a paragraph to discussion mentioning some limitations, like the inflation of the non-decision times and the difficulty of interpreting the fit parameters (Paragraph #5 of Discussion: “The key assumption of the PDG model is that there is...”). (2) Dependence on non-decision time assumptions The alternative model's explanatory power appears to rely heavily on assumptions regarding the decomposition of non-decision time: fixed visual encoding (𝜏s = 0.3 s), manual non-decision time (𝜏m; two free parameters), and saccadic execution (𝜏e; fixed parameters μe = 0.35, σe = 0.11). - 𝜏e is substantially longer and more variable than typical saccadic execution times, suggesting it likely incorporates decisional components. - Estimated 𝜏m values are approximately twice as long as known manual execution durations. - σnd is more plausible, implying that variability is captured correctly but mean durations are not. Together, these points raise the possibility that portions of what the model treats as non-decision time are in fact part of a (action) decision process. Only then does it make sense to assume that Tm is usually larger than Te. If Tm and Te were truly execution delays, then Tm would always be larger than Te. You may find it helpful to consider the framework in Bompas et al. Psych Review (2024), which discusses in detail what non-decision time is likely to comprise across effectors. Thank you we have added (i) a sensitivity analysis showing that our results are robust to changes in the specific value used for the eye movement related latencies (new Fig. 8), and (ii) a new paragraph in Discussion addressing the issue of the mismatch between our parameter estimates and the manual and saccadic execution times (Paragraph #5 of Discussion: “The key assumption of the PDG model is that there is...”). (3) Code availability. The authors should consider sharing all relevant code and data publicly. We agree, we now share the code and data on GitHub and indicate so in the revised manuscript. Minor Comments: (1) Lines 74-77. These are not worded as predictions but as questions; one tests predictions, but answers questions. I feel it would be clearer to stick to predictions (like in the abstract), and the introduction could benefit from explaining these predictions in a bit more detail (I found it difficult to get my head around these predictions from the intro text only). We rewrote the section in the introduction where we provide a gist of the model predictions (last paragraph of Introduction). We agree with the reviewer that the previous explanation was not clear. (2) It is confusing that panel B appears to the left of panel A in Figure 2. We agree. We have restructured the manuscript (following the suggestion of another reviewer), and now Figure 2 has changed and the panels follow a more logical order. (3) Figure 3C - remove MATLAB toggles. Yes, thanks. (4) Figure 5A shows the proportion of left choices, but the text and legend refer to right choices. Good catch, thank you. Reviewer #2 (Recommendations for the authors): This may appear self-serving, but the authors seem to be unaware of some highly relevant work from our group. Most importantly, in a recent publication (Ting & Gluth, 2024, JEP General), we have already looked at the dependency of the last- (or final-) fixation bias on overall value in value-based (VB) and perceptual (P) decisions. In VB, we found a negative effect; in P we did not find a significant effect. This is largely consistent with the current results, showing a negative but not significant trend. Another relevant work is Gluth et al. (2020, Nat Hum Behav), where we extended the aDDM by assuming that the probability to fixate on an option is a function of the accumulated evidence for that option. It would be interesting to know whether this assumption changes the predictions of the aDDM. Finally, we just published a new theory on how people search for information to make efficient value-based decisions (Gluth et al., in press, Psychol Rev; https://osf.io/preprints/psyarxiv/3qzak_v2). Although this theory focuses on multi-attribute choices, it can be applied to "simple" choices, too (by assuming that there is only one attribute = value). Interestingly, while the model also mispredicts a (slight) increase of the last-fixation bias with overall value, it correctly predicts the independency of the dwell-time advantage effect on choice consistency as well as the small increase of the effect with RT (attached here is a figure to show this: [https://elife-rp.msubmit.net/elife-rp_files/2026/01/22/00149589/00/149589_0_attach_9_477122. pdf], and the match with the empirical data shown in Figure 3B and 12 is striking). In general, the model shares many features of the Callaway and Jang models, but does not need to assume a biased value prior, which the authors suggest is responsible for the misprediction of the second effect. I leave it up to the authors to discuss this new theory, but I wanted to point this out. Thank you for pointing this out; these are all relevant points and studies. We now note that the first of our predictions has recently been identified and tested by Ting and Gluth (2025). We also considered extending the manuscript with a variant of the model proposed by Gluth et al. (Psychological Review, 2026). In fact, we attempted to fit this model to the Krajbich et al. (2010) dataset under the assumption that the duration of each sampling epoch is a free parameter. We find this model very interesting. However, in our current implementation it appears to make the same qualitative prediction as the aDDM, namely that ΔDwell depends on choice consistency (see Author response image 1). Given this, we have decided not to include these results in the manuscript. It remains possible that with further development particularly with a more realistic specification of fixation durations (e.g., allowing them to depend on value) the model could account for the full set of observed effects. We think this would be best addressed in a separate study. That said, we do find the model promising, as it provides a better account than most of the alternative models we explored for the patterns shown in panels D, H, and I. Fits of a variant of the MACS model (Gluth et al. 2026) to the data of Krajbich et al. (2010). The paper would benefit substantially from restructuring. The aDDM's predictions are provided first, together with the empirical data, and then the optimal models are discussed. But Figure 2 shows all of this together. Later, the new (PDG) model is elaborated, and its predictions are shown. Towards the end of the results, variations of the aDDM and combinations of aDDM and PDG are shown in a series of figures (8-11), followed by a last figure showing one of the tested effects in other datasets. All of this feels pretty much thrown together without a clear structure. For instance, the aDDM and the optimal models could be described together (or the optimal models get a separate figure). The additive variants could be described earlier. And some figures could be put into the supplement. And the empirical results of the different studies could be shown together. We fully agree with this suggestion. We have now restructured the manuscript along the lines proposed by the reviewer (see the more detailed explanation of the restructuring in our response to the public comments). I strongly suggest avoiding the term "influence" in the y-axis of Figure 2, upper row, as it implies causality. Similarly, in line 182, the term "causal influence" is used in the context of the Callaway model, but as far as I know, this is not what the model assumes. We replaced the y-axis label with “Association of last dwell with choice (β)” Reviewer #3 (Recommendations for the authors): (1) Figure 2 - Panel labels for A and B are reversed? We have restructured the manuscript (following the suggestion of another reviewer), and now Figure 2 has changed. (2) Does 3C include a .pdf screenshot? Thank you, it’s a Matlab bug on Mac. I guess they want us to switch to Python -:) (3) Figure 4 - It would be helpful if the green line were defined in the figure legend. Added (4) The effect size in 5B looks much more dramatic than in 2B(A?) - Is this for one example subject as opposed to all subjects? Please clarify what is different about the data. We are no longer showing the psychometric functions in Figure 2. (5) Line 252 - they say they compared the probability of choosing the right item (Fig. 5B) by the y-labels of that figure, which are all p(choose left). Yes, corrected now. (6) In general, they reference the subpanels of Figure 5 out of order, which causes the reader to jump around. They might consider reordering the panels of the figure so they follow the ordering of descriptions in the text. We agree, we have rearranged the figure panels to follow the ordering of the descriptions in the text.