Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 75 of the 75 references it cites were found in a registry or an index; 37 of the 59 sources its claims cite were checked for support.

Interpretable abstractions of artificial neural networks predict behavior and neural activity during human information gathering

Open access

Published version of a preprint

Abstract

Humans and other animals are driven to acquire information about opportunities in their environments, yet how they evaluate what is worth learning remains unclear. Here we combine artificial neural networks with symbolic regression to extract an expressive yet interpretable model that specifies how human participants evaluate decision-relevant information during choice. The recovered function depends primarily on the relative evidence accumulated across options rather than absolute uncertainty about each, revealing that participants seek information symmetry across alternatives rather than minimizing uncertainty option by option. This account outperforms standard models of uncertainty-based exploration and generalizes to an independent dataset. Using ultrahigh-field (7T) functional magnetic resonance imaging optimized for midbrain and brainstem, we simultaneously measured activity across five neuromodulatory nuclei and two cortical regions. Ventral tegmental area activity showed opposed coding of information and selection values, a pattern suited to arbitrating between sampling and choosing, and anterior cingulate cortex and anterior insula tracked value-of-information computations.

Overview

CoChat Summary

TL;DR

The paper combines a hybrid artificial neural network with symbolic regression to model how humans value information during sampling, recovering a four-parameter function based on relative evidence across options rather than absolute uncertainty. The symbolic model predicts behavior and 7T fMRI activity, with VTA showing opposed coding of information and selection values and ACC/AI tracking value-of-information computations.

Key findings

  • In 20 participants, sampling increased with higher initial uncertainty (β = −0.069, s.e. = 0.013, z = −5.18, P = 2.27 × 10−7) and with smaller final red-dot discrepancy (β = −0.131, s.e. = 0.019, z = −6.87, P = 6.48 × 10−12); uncertainty in the blocked option did not significantly influence sampling (β = −0.006, s.e. = 0.006, z = −1.01, P = 0.310).
  • The hybrid-ANN model fit behavior better than linear and UCB models for all participants, and the symbolic model performed comparably to the ANN. The symbolic functions generalized to an independent two-armed bandit dataset, outperforming UCB (Wilcoxon W = 15.0, n = 89, median difference = −0.07, P = 4.31 × 10−16) and an offset UCB variant (W = 286.0, n = 89, median difference = −0.03, P = 2.21 × 10−12).
  • Symbolic regression recovered a four-parameter function depending on the ratio Na/Nu, interpreted as attentional inertia (β1), information satiation (β2), undirected exploration (β3) and directed exploration (β4); separate first-visit and subsequent-visit parameters yielded eight total parameters. This supports an information-symmetry principle: participants balance knowledge across options rather than minimizing uncertainty option by option.
  • ANN-derived VoI better fit neural data than linear (Cohen’s d = −1.025, Wilcoxon P = 2.12 × 10−12) and UCB (d = −1.105, P = 5.64 × 10−13); the symbolic function performed almost as well as the ANN (d = −0.129, P = 0.006). VTA showed positive coding of summed VoS (β = 11.3, P = 0.020) and negative coding of summed VoI (β = −46.1, P = 1.93 × 10−5); SN was negatively associated with VoI sum (β = −24.9, P = 6.38 × 10−5) and difference (β = −22.7, P = 0.008); AI and ACC tracked VoI-related signals; VSN, DRN and LC showed no significant associations after Bonferroni correction.

Methods

  • Twenty participants (14 females), aged 19–32, completed four 25-min sessions of an information-sampling task inside a 7T MRI scanner. Trials presented three patches of 100 moving dots; green covers revealed initial information simultaneously, gray covers revealed dots sequentially every 150 ms, and on 80% of trials one patch was blocked, leaving two options for subsequent analyses. Sampling was capped at 24 s for two-option trials and 33 s for three-option trials.
  • The hybrid model computed values for four actions (stay, switch, select attended, select unattended). It compared linear, UCB and ANN-based VoI functions; the ANN was a Lipschitz-bounded deep network with four hidden layers of 32 neurons each and tanh activations, trained with Adam and L-BFGS. Symbolic regression (SymbolicRegression.jl) was applied to the trained ANN’s predictions to recover closed-form equations.
  • fMRI data were acquired at 7T with 1-mm isotropic voxels and TR = 1.378 s. A GLM used VoI as a parametric modulator while controlling for outcome and VoS; model comparisons used per-voxel MSE, Cohen’s d with bootstrap confidence intervals and Wilcoxon signed-rank tests. ROI analyses used mixed-effects models weighted by inverse variance, with Bonferroni correction across five subcortical ROIs.

Limitations

  • The paper states that the specific symbolic equations are tailored to the two-option structure and encode relative comparisons between attended and unattended options; unlike UCB, they do not naturally extend to n-armed bandits. Possible extensions include replacing Nunattended with the mean across unattended options or weighting unattended options by salience or recency.
  • The paper states that the task does not permit dissociation of decision uncertainty, evidence uncertainty and reward uncertainty, which are highly correlated in the design; orthogonal manipulation of these uncertainty types is an important direction for future research.
  • The paper reports that VSN, DRN and LC showed no significant association with the sum or difference of VoI or VoS after Bonferroni correction, and notes that within the sensitivity limits of the measurement and analysis, these nuclei do not strongly encode these decision variables in this task. Evident from the text: the neural analyses are restricted to two-option trials, and the sample size is 20 participants.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41593-026-02342-9

Cites

75 of 75 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Information-sampling task. a, The decision-making process involves two key decisions: whether to gather more information or make a selection (top versus bottom) and, if gathering information, whether to sample from the currently attended option or switch to the alternative (top right versus top left). b, Three approaches to computing the VoI as a function of the number of samples. Top: a linear function where value decreases at a constant rate with each additional sample. Middle: UCB algorithm that captures diminishing returns, with steeper initial decline that flattens as samples accumulate. Bottom: an ANN that learns the mapping between samples and information value; the learned function’s form is not specified a priori. c, Task structure showing the two phases of the information-sampling task. In phase 1, participants are presented with three patches of dots covered by green or gray covers. After revealing the green-covered dots in each patch, one patch is blocked (gray circle). In phase 2, participants can freely sample information by hovering over patches, with gray-covered dots revealing their true colors sequentially, before making a final selection. When participants switched patches, previously revealed dots in the unattended patch returned to their gray-covered state, requiring reliance on memory (as illustrated by the gray patch in phase 2, right panel). d, Brain ROIs: LC, DRN, VSN, SN and VTA, which have been implicated in uncertainty processing and information sampling.
Fig. 1 Information-sampling task. a, The decision-making process involves two key decisions: whether to gather more information or make a selection (top versus bottom) and, if gathering information, whether to sample from the currently attended option or switch to the alternative (top right versus top left). b, Three approaches to computing the VoI as a function of the number of samples. Top: a linear function where value decreases at a constant rate with each additional sample. Middle: UCB algorithm that captures diminishing returns, with steeper initial decline that flattens as samples accumulate. Bottom: an ANN that learns the mapping between samples and information value; the learned function’s form is not specified a priori. c, Task structure showing the two phases of the information-sampling task. In phase 1, participants are presented with three patches of dots covered by green or gray covers. After revealing the green-covered dots in each patch, one patch is blocked (gray circle). In phase 2, participants can freely sample information by hovering over patches, with gray-covered dots revealing their true colors sequentially, before making a final selection. When participants switched patches, previously revealed dots in the unattended patch returned to their gray-covered state, requiring reliance on memory (as illustrated by the gray patch in phase 2, right panel). d, Brain ROIs: LC, DRN, VSN, SN and VTA, which have been implicated in uncertainty processing and information sampling.CC BY · PubMed Central
Fig. 2 Sampling behavior and computational model. a, Relationship between accuracy and average amount of evidence collected. Each colored dot represents an individual participant, with color indicating their final score according to the color scale on the right. The dark green star marks the position of the optimal agent (optimal under the cost assumptions specified in Methods). n = 20 participants. b, Number of samples collected under different task conditions. Left: bar graph showing the number of samples collected across three levels of relative unsigned proportion of red dots between patches (0.1, 0.2, 0.3). Right: bar graph showing the number of samples collected across three levels of initial information (30, 45, 60 green dots). Individual participant data points are shown as colored dots, and black stars indicate the model-derived optimal agent’s behavior. c, Schematic of the computational model. The model transforms objective magnitudes (number and color of dots in each patch) into subjective magnitudes through attentional discounting and memory decay functions. These subjective magnitudes are used to compute the value of selecting each patch (pink) and the value of sampling more information (blue), which together determine the final action. The dotted blue circle highlights the component that computes the VoI, which is implemented using a linear function, UCB algorithm or ANN. Source data
Fig. 2 Sampling behavior and computational model. a, Relationship between accuracy and average amount of evidence collected. Each colored dot represents an individual participant, with color indicating their final score according to the color scale on the right. The dark green star marks the position of the optimal agent (optimal under the cost assumptions specified in Methods). n = 20 participants. b, Number of samples collected under different task conditions. Left: bar graph showing the number of samples collected across three levels of relative unsigned proportion of red dots between patches (0.1, 0.2, 0.3). Right: bar graph showing the number of samples collected across three levels of initial information (30, 45, 60 green dots). Individual participant data points are shown as colored dots, and black stars indicate the model-derived optimal agent’s behavior. c, Schematic of the computational model. The model transforms objective magnitudes (number and color of dots in each patch) into subjective magnitudes through attentional discounting and memory decay functions. These subjective magnitudes are used to compute the value of selecting each patch (pink) and the value of sampling more information (blue), which together determine the final action. The dotted blue circle highlights the component that computes the VoI, which is implemented using a linear function, UCB algorithm or ANN. Source dataCC BY · PubMed Central
Fig. 3 Model comparison. a, Schematic of three approaches to computing the VoI as a function of the number of samples (see Fig. 1b for details): linear function (top), UCB algorithm with diminishing returns (middle) and ANN with learned function form (bottom). b, Model comparison across participants showing loss relative to the linear model. Each vertical line represents a participant, with light blue circles (UCB model), dark blue circles (hybrid-ANN model) and open squares (symbolic model derived through symbolic regression) showing the relative loss for each model type. The dotted line shows the performance of an end-to-end ANN (trained to directly predict actions without cognitive structure). Lower values on the y axis indicate better model fit. n = 20 participants; leave-one-out cross-validation. Source data
Fig. 3 Model comparison. a, Schematic of three approaches to computing the VoI as a function of the number of samples (see Fig. 1b for details): linear function (top), UCB algorithm with diminishing returns (middle) and ANN with learned function form (bottom). b, Model comparison across participants showing loss relative to the linear model. Each vertical line represents a participant, with light blue circles (UCB model), dark blue circles (hybrid-ANN model) and open squares (symbolic model derived through symbolic regression) showing the relative loss for each model type. The dotted line shows the performance of an end-to-end ANN (trained to directly predict actions without cognitive structure). Lower values on the y axis indicate better model fit. n = 20 participants; leave-one-out cross-validation. Source dataCC BY · PubMed Central
Fig. 4 Visualization of VoI functions across computational models. a, 3D surface plots showing how the VoI varies with the number of revealed dots in both the attended and unattended patches. Each model (linear, UCB and ANN) is represented by two surfaces: darker color for the value of staying with the attended patch and lighter color for the value of switching to the unattended patch. b, Heatmaps showing the value of staying as a function of evidence collected from both patches. Color intensity represents the magnitude of the value, with the scale shown on the right. c, Heatmaps showing the value of switching as a function of evidence collected from both patches. d, Vector-field plots showing how collecting an additional sample from the attended option affects both the value of staying (x axis) and the value of switching (y axis). Each arrow represents the transition from current values (arrow origin) to updated values (arrow tip) after collecting one new sample. The color of the arrows indicates the difference in evidence between attended and unattended patches according to the color scale on the right, and the length of the arrow indicates the size of the update. Source data
Fig. 4 Visualization of VoI functions across computational models. a, 3D surface plots showing how the VoI varies with the number of revealed dots in both the attended and unattended patches. Each model (linear, UCB and ANN) is represented by two surfaces: darker color for the value of staying with the attended patch and lighter color for the value of switching to the unattended patch. b, Heatmaps showing the value of staying as a function of evidence collected from both patches. Color intensity represents the magnitude of the value, with the scale shown on the right. c, Heatmaps showing the value of switching as a function of evidence collected from both patches. d, Vector-field plots showing how collecting an additional sample from the attended option affects both the value of staying (x axis) and the value of switching (y axis). Each arrow represents the transition from current values (arrow origin) to updated values (arrow tip) after collecting one new sample. The color of the arrows indicates the difference in evidence between attended and unattended patches according to the color scale on the right, and the length of the arrow indicates the size of the update. Source dataCC BY · PubMed Central
Fig. 5 Symbolic representation of the ANN-derived VoI function. a, Process of deriving interpretable mathematical expressions from behavioral data. Left: schematic of the dataset containing patterns of dots across time. Middle: function approximation using a deep neural network. Right: symbolic representation showing the mathematical equations derived through symbolic regression for the value of staying and the value of switching, where Na and Nu represent the number of dots in the attended and unattended patches, respectively. b, Cross-task generalization comparison. Scatter plot showing leave-one-out cross-validation loss on a test set from a two-armed bandit task20 for both the symbolic model (left) and the UCB model (right). Each gray line connects performance of both models for the same participant. Statistical comparison used a two-sided Wilcoxon signed-rank test on per-subject mean losses (W = 15.0, median difference = −0.07, P = 4.31 × 10−16). n = 89 participants. c, Parameter sensitivity analysis. Three-dimensional surface plots showing how the value functions change with different parameter values. Each surface represents the function with a specific parameter value according to the color scales above each plot. Source data
Fig. 5 Symbolic representation of the ANN-derived VoI function. a, Process of deriving interpretable mathematical expressions from behavioral data. Left: schematic of the dataset containing patterns of dots across time. Middle: function approximation using a deep neural network. Right: symbolic representation showing the mathematical equations derived through symbolic regression for the value of staying and the value of switching, where Na and Nu represent the number of dots in the attended and unattended patches, respectively. b, Cross-task generalization comparison. Scatter plot showing leave-one-out cross-validation loss on a test set from a two-armed bandit task20 for both the symbolic model (left) and the UCB model (right). Each gray line connects performance of both models for the same participant. Statistical comparison used a two-sided Wilcoxon signed-rank test on per-subject mean losses (W = 15.0, median difference = −0.07, P = 4.31 × 10−16). n = 89 participants. c, Parameter sensitivity analysis. Three-dimensional surface plots showing how the value functions change with different parameter values. Each surface represents the function with a specific parameter value according to the color scales above each plot. Source dataCC BY · PubMed Central
Fig. 6 Neural correlates of ANN-derived value signals. a, Whole-brain analysis showing regions where BOLD activity correlates with the difference between the value of staying and the value of switching (VoI stay − VoI switch). b, Whole-brain analysis showing regions correlating with the sum of the value of staying and switching (VoI stay + VoI switch). Color bars indicate Z-statistic values; thresholded at Z > 3.1, cluster-corrected P < 0.001. c, ROI analysis results for neuromodulatory nuclei. Left: sagittal view showing locations of VTA (orange), SN (green), VSN (pink), DRN (purple) and LC (blue). Right: coefficient estimates (effect sizes from weighted mixed-effects models on beta values) for regressors representing the sum (Σ) and difference (Δ) of the VoS and VoI within each ROI. Each small dot represents an individual participant’s random-effect estimate plus the fixed effect; shaded bars indicate the 95% confidence interval of the fixed effect (group mean). n = 20 participants (×4 sessions per participant). Source data
Fig. 6 Neural correlates of ANN-derived value signals. a, Whole-brain analysis showing regions where BOLD activity correlates with the difference between the value of staying and the value of switching (VoI stay − VoI switch). b, Whole-brain analysis showing regions correlating with the sum of the value of staying and switching (VoI stay + VoI switch). Color bars indicate Z-statistic values; thresholded at Z > 3.1, cluster-corrected P < 0.001. c, ROI analysis results for neuromodulatory nuclei. Left: sagittal view showing locations of VTA (orange), SN (green), VSN (pink), DRN (purple) and LC (blue). Right: coefficient estimates (effect sizes from weighted mixed-effects models on beta values) for regressors representing the sum (Σ) and difference (Δ) of the VoS and VoI within each ROI. Each small dot represents an individual participant’s random-effect estimate plus the fixed effect; shaded bars indicate the 95% confidence interval of the fixed effect (group mean). n = 20 participants (×4 sessions per participant). Source dataCC BY · PubMed Central
Extended Data Fig. 1 Softmax competition balances sampling and selection. Model predictions when one option becomes clearly better. Starting from equal evidence (20 red, 20 black dots in both options), we incrementally added red dots to the attended option. As the attended option became clearly better, the probability of selecting it increased (pink line), while the probability of continuing to sample (staying, dark blue line) decreased. This demonstrates that although the value of information in the hybrid model depends only on sample counts (N) rather than evidence quality (red/black composition), the model correctly predicts reduced sampling when evidence clearly favors one option: softmax competition between sampling and selection action values shifts probability mass from sampling to selection as the value of selection grows with evidence divergence. Source data
Extended Data Fig. 1 Softmax competition balances sampling and selection. Model predictions when one option becomes clearly better. Starting from equal evidence (20 red, 20 black dots in both options), we incrementally added red dots to the attended option. As the attended option became clearly better, the probability of selecting it increased (pink line), while the probability of continuing to sample (staying, dark blue line) decreased. This demonstrates that although the value of information in the hybrid model depends only on sample counts (N) rather than evidence quality (red/black composition), the model correctly predicts reduced sampling when evidence clearly favors one option: softmax competition between sampling and selection action values shifts probability mass from sampling to selection as the value of selection grows with evidence divergence. Source dataCC BY · PubMed Central
Extended Data Fig. 2 Simulation-recovery validation of the hybrid modelling approach. Data were simulated from four agents, each using a distinct non-linear value-of-information function (colour-coded blue, green, pink, and orange). Applying the hybrid modelling approach to each simulated dataset recovered the underlying generating function. Left, three-dimensional scatter showing the value of information as a function of the number of dots revealed in the attended (Nattended) and unattended (Nunattended) patches. Colour indicates the simulated generating function. Right, correlation between true and recovered functions across three sample sizes (4,020, 8,040, and 12,060 observations). Each dot represents a single recovery attempt; colour indicates the simulated function. Pearson r > 0.95 between recovered and true functions across sample sizes, demonstrating reliable recovery across the range of actual participant data. See Methods for simulation details. Source data
Extended Data Fig. 2 Simulation-recovery validation of the hybrid modelling approach. Data were simulated from four agents, each using a distinct non-linear value-of-information function (colour-coded blue, green, pink, and orange). Applying the hybrid modelling approach to each simulated dataset recovered the underlying generating function. Left, three-dimensional scatter showing the value of information as a function of the number of dots revealed in the attended (Nattended) and unattended (Nunattended) patches. Colour indicates the simulated generating function. Right, correlation between true and recovered functions across three sample sizes (4,020, 8,040, and 12,060 observations). Each dot represents a single recovery attempt; colour indicates the simulated function. Pearson r > 0.95 between recovered and true functions across sample sizes, demonstrating reliable recovery across the range of actual participant data. See Methods for simulation details. Source dataCC BY · PubMed Central
Extended Data Fig. 3 Feature selection identifies the ANN inputs essential for computing the value of information. Dot plot showing the change in model loss (ΔLoss) when each candidate feature is removed from the ANN input vector. Each blue dot indicates the mean loss difference across cross-validation folds; vertical blue lines indicate 95% confidence intervals. Features tested: N-same (non-gray dots in the attended patch), N-other (non-gray dots in the unattended patch), N-blocked (non-gray dots in the blocked patch), ρ-same (subjective proportion of red dots in the attended patch), ρ-other (subjective proportion of red dots in the unattended patch), cursor (current cursor position, that is which patch is attended), and visit (first-visit indicator: true while the participant continues sampling the first-attended patch; false after the first switch). Four features are critical—their removal substantially increases loss: cursor, visit, N-same, and N-other. In contrast, removing ρ-same, ρ-other, or N-blocked does not impair model performance. The cursor and visit variables structure how the ANN computes the value of information: cursor determines which function applies (value of staying when attending a patch, value of switching when attending the other patch), while visit modulates the parameters of these functions. n = 20 participants; leave-one-out cross-validation. Source data
Extended Data Fig. 3 Feature selection identifies the ANN inputs essential for computing the value of information. Dot plot showing the change in model loss (ΔLoss) when each candidate feature is removed from the ANN input vector. Each blue dot indicates the mean loss difference across cross-validation folds; vertical blue lines indicate 95% confidence intervals. Features tested: N-same (non-gray dots in the attended patch), N-other (non-gray dots in the unattended patch), N-blocked (non-gray dots in the blocked patch), ρ-same (subjective proportion of red dots in the attended patch), ρ-other (subjective proportion of red dots in the unattended patch), cursor (current cursor position, that is which patch is attended), and visit (first-visit indicator: true while the participant continues sampling the first-attended patch; false after the first switch). Four features are critical—their removal substantially increases loss: cursor, visit, N-same, and N-other. In contrast, removing ρ-same, ρ-other, or N-blocked does not impair model performance. The cursor and visit variables structure how the ANN computes the value of information: cursor determines which function applies (value of staying when attending a patch, value of switching when attending the other patch), while visit modulates the parameters of these functions. n = 20 participants; leave-one-out cross-validation. Source dataCC BY · PubMed Central
Extended Data Fig. 4 Mixed-effects logistic regressions linking the ANN-derived value of information to three nested decisions. Each panel shows subject-specific random-effect estimates (dots) together with ± 95% confidence intervals of the fixed effects (rectangles). All P-values are from two-sided Wald tests on regression coefficients. n = 20 participants. a, Probability of sampling versus selecting a patch. Predictors: sum of the value of information across options (VoI sum; β = 2.860, s.e. = 0.167, z = 17.16, P = 5.25 × 10−66) and unsigned difference in VoI between options (∣VoI difference∣; β = 1.292, s.e. = 0.085, z = 15.16, P = 6.14 × 10−52). Value-of-selection predictors (VoS sum, ∣VoS difference∣) are included as controls. b, Probability of staying with the currently attended patch versus switching. Predictors: VoI for the attended patch (β = 3.971, s.e. = 0.434, z = 9.14, P = 6.03 × 10−20) and VoI for the unattended patch (β = − 5.880, s.e. = 0.456, z = − 12.89, P = 5.13 × 10−38). Value-of-selection differences are included as controls. c, Probability of selecting the attended versus unattended patch. Predictors: difference in red-dot proportions between attended and unattended patches (ρ difference; β = 4.572, s.e. = 0.336, z = 13.60, P = 3.77 × 10−42); sum of red-dot proportions across patches (ρ sum; β = 0.631, s.e. = 0.104, z = 6.08, P = 1.19 × 10−9); and VoI(attended) - VoI(unattended) (β = − 0.782, s.e. = 0.159, z = − 4.91, P = 9.16 × 10−7). Participants were less likely to select a patch whose remaining value of information was higher than the alternative’s, a pattern consistent with an account in which participants seek to minimize uncertainty about the option they will ultimately choose. Note: the predictors used in these regressions (sums, unsigned differences, per-option VoIs) were not directly optimized during ANN training, so these regression analyses provide insight into how the learned value-of-information representation decomposes into interpretable decision variables.
Extended Data Fig. 4 Mixed-effects logistic regressions linking the ANN-derived value of information to three nested decisions. Each panel shows subject-specific random-effect estimates (dots) together with ± 95% confidence intervals of the fixed effects (rectangles). All P-values are from two-sided Wald tests on regression coefficients. n = 20 participants. a, Probability of sampling versus selecting a patch. Predictors: sum of the value of information across options (VoI sum; β = 2.860, s.e. = 0.167, z = 17.16, P = 5.25 × 10−66) and unsigned difference in VoI between options (∣VoI difference∣; β = 1.292, s.e. = 0.085, z = 15.16, P = 6.14 × 10−52). Value-of-selection predictors (VoS sum, ∣VoS difference∣) are included as controls. b, Probability of staying with the currently attended patch versus switching. Predictors: VoI for the attended patch (β = 3.971, s.e. = 0.434, z = 9.14, P = 6.03 × 10−20) and VoI for the unattended patch (β = − 5.880, s.e. = 0.456, z = − 12.89, P = 5.13 × 10−38). Value-of-selection differences are included as controls. c, Probability of selecting the attended versus unattended patch. Predictors: difference in red-dot proportions between attended and unattended patches (ρ difference; β = 4.572, s.e. = 0.336, z = 13.60, P = 3.77 × 10−42); sum of red-dot proportions across patches (ρ sum; β = 0.631, s.e. = 0.104, z = 6.08, P = 1.19 × 10−9); and VoI(attended) - VoI(unattended) (β = − 0.782, s.e. = 0.159, z = − 4.91, P = 9.16 × 10−7). Participants were less likely to select a patch whose remaining value of information was higher than the alternative’s, a pattern consistent with an account in which participants seek to minimize uncertainty about the option they will ultimately choose. Note: the predictors used in these regressions (sums, unsigned differences, per-option VoIs) were not directly optimized during ANN training, so these regression analyses provide insight into how the learned value-of-information representation decomposes into interpretable decision variables.CC BY · PubMed Central
Extended Data Fig. 5 Model comparison for evidence weighting. Standardised regression coefficients from regressing each model’s output on the number of dots revealed in the attended and unattended patches. Circles represent weights for the attended evidence, diamonds represent weights for the unattended evidence; colours transition from light green to dark blue across models, and dotted lines connect the weights across models a, Value of staying. The linear model depends exclusively on attended evidence (βattended = 1.0, βunattended = 0.0). The UCB model is predominantly attended-driven with modest unattended modulation (βattended = 0.8, βunattended = 0.2). The ANN shows balanced weights on both dimensions (βattended = 0.7, βunattended = 0.6), quantitatively confirming the diagonal gradient visible in Fig. 4bb, Value of switching. A mirror pattern emerges, using the same visual conventions as in panel a: the linear model depends exclusively on unattended evidence, UCB predominantly on unattended with modest attended modulation, and the ANN shows balanced weighting across both dimensions. Source data
Extended Data Fig. 5 Model comparison for evidence weighting. Standardised regression coefficients from regressing each model’s output on the number of dots revealed in the attended and unattended patches. Circles represent weights for the attended evidence, diamonds represent weights for the unattended evidence; colours transition from light green to dark blue across models, and dotted lines connect the weights across models a, Value of staying. The linear model depends exclusively on attended evidence (βattended = 1.0, βunattended = 0.0). The UCB model is predominantly attended-driven with modest unattended modulation (βattended = 0.8, βunattended = 0.2). The ANN shows balanced weights on both dimensions (βattended = 0.7, βunattended = 0.6), quantitatively confirming the diagonal gradient visible in Fig. 4bb, Value of switching. A mirror pattern emerges, using the same visual conventions as in panel a: the linear model depends exclusively on unattended evidence, UCB predominantly on unattended with modest attended modulation, and the ANN shows balanced weighting across both dimensions. Source dataCC BY · PubMed Central
Extended Data Fig. 6 Symbolic regression approximation of ANN-derived value-of-information functions. a, Correlation between ANN-generated value of information and the symbolic-function approximations across all participants (n = 20). Four correlations are shown: value of staying and value of switching, each for first visits and for subsequent visits to a patch. Each point represents one participant; the highlighted participant (white circle) is shown in detail in panel b. b, Example participant showing 3D visualizations of the value-of-information functions. Left column: target functions from the trained ANN. Right column: estimates from the optimized symbolic functions. Top row: first-visit functions; bottom row: subsequent-visit functions. Dark points represent value of staying; light points represent value of switching. c, Individual parameter estimates across all participants. Each parameter corresponds to a specific component of the value-of-information computation: β1(1), β1( > 1) (stay-VoI offsets, ’attentional inertia’); β2(1), β2( > 1) (stay-VoI scaling, ’information satiation’); β3(1), β3( > 1) (switch-VoI offsets, ’undirected exploration’); β4(1), β4( > 1) (switch-VoI scaling, ’directed exploration’), for first (superscript 1) and subsequent (superscript >1) visits respectively. The highlighted participant (white circle) corresponds to the example shown in panel b. Source data
Extended Data Fig. 6 Symbolic regression approximation of ANN-derived value-of-information functions. a, Correlation between ANN-generated value of information and the symbolic-function approximations across all participants (n = 20). Four correlations are shown: value of staying and value of switching, each for first visits and for subsequent visits to a patch. Each point represents one participant; the highlighted participant (white circle) is shown in detail in panel b. b, Example participant showing 3D visualizations of the value-of-information functions. Left column: target functions from the trained ANN. Right column: estimates from the optimized symbolic functions. Top row: first-visit functions; bottom row: subsequent-visit functions. Dark points represent value of staying; light points represent value of switching. c, Individual parameter estimates across all participants. Each parameter corresponds to a specific component of the value-of-information computation: β1(1), β1( > 1) (stay-VoI offsets, ’attentional inertia’); β2(1), β2( > 1) (stay-VoI scaling, ’information satiation’); β3(1), β3( > 1) (switch-VoI offsets, ’undirected exploration’); β4(1), β4( > 1) (switch-VoI scaling, ’directed exploration’), for first (superscript 1) and subsequent (superscript >1) visits respectively. The highlighted participant (white circle) corresponds to the example shown in panel b. Source dataCC BY · PubMed Central
Extended Data Fig. 7 Neural model comparison across whole brain. Ridge plots showing the bootstrap distributions of Cohen’s d effect sizes for pairwise model comparisons in predicting neural activity (n = 80 sessions, two-sided Wilcoxon signed-rank tests). The ANN vs. Symbolic comparison (bottom, black) shows a small effect size (d = − 0.129, 95% CI [ − 0.494, 0.078], Wilcoxon P = 0.006), with a confidence interval overlapping zero, indicating equivalent performance. In contrast, ANN vs. Linear (dark green, d = − 1.025, 95% CI [ − 1.264, − 0.849], Wilcoxon P = 2.12 × 10−12) and ANN vs. UCB (dark green, d = − 1.105, 95% CI [ − 1.366, − 0.918], Wilcoxon P = 5.64 × 10−13) show large negative effects, indicating substantially better ANN performance. Symbolic vs. Linear (dark red, d = − 0.882) and Symbolic vs. UCB (dark red, d = − 0.953) also show large effects, demonstrating that the symbolic function outperforms standard heuristics while matching ANN performance. Darker shaded regions within each distribution represent 95% confidence intervals. Vertical reference lines indicate conventional effect-size thresholds: no effect (0), small (0.2), medium (0.5), and large (0.8). Negative Cohen’s d values indicate the first model has lower MSE (better fit). Source data
Extended Data Fig. 7 Neural model comparison across whole brain. Ridge plots showing the bootstrap distributions of Cohen’s d effect sizes for pairwise model comparisons in predicting neural activity (n = 80 sessions, two-sided Wilcoxon signed-rank tests). The ANN vs. Symbolic comparison (bottom, black) shows a small effect size (d = − 0.129, 95% CI [ − 0.494, 0.078], Wilcoxon P = 0.006), with a confidence interval overlapping zero, indicating equivalent performance. In contrast, ANN vs. Linear (dark green, d = − 1.025, 95% CI [ − 1.264, − 0.849], Wilcoxon P = 2.12 × 10−12) and ANN vs. UCB (dark green, d = − 1.105, 95% CI [ − 1.366, − 0.918], Wilcoxon P = 5.64 × 10−13) show large negative effects, indicating substantially better ANN performance. Symbolic vs. Linear (dark red, d = − 0.882) and Symbolic vs. UCB (dark red, d = − 0.953) also show large effects, demonstrating that the symbolic function outperforms standard heuristics while matching ANN performance. Darker shaded regions within each distribution represent 95% confidence intervals. Vertical reference lines indicate conventional effect-size thresholds: no effect (0), small (0.2), medium (0.5), and large (0.8). Negative Cohen’s d values indicate the first model has lower MSE (better fit). Source dataCC BY · PubMed Central
Extended Data Fig. 8 Representational Similarity Analysis links neural patterns to sampling behaviour. a, Schematic illustrating the relationship between brain activity and behavior. b, Example Neural RDM (left, based on multi-voxel BOLD activity patterns) and Behavioral RDM (right, based on the number of samples taken per trial) for a participant session. Each cell represents the dissimilarity between a pair of trials; higher values (yellow) indicate greater dissimilarity. c, Kendall’s τ correlation coefficients quantifying the relationship between neural and behavioral RDMs for VTA, AI, ACC, and SN. Neural RDMs for AI, ACC, and SN were derived using masks matched in shape and size to the VTA ROI to control for region size. Each small dot represents the correlation calculated for one participant session; larger dots indicate the mean correlation across participants. Asterisks denote significant positive correlations across participants (two-sided Wilcoxon signed-rank test; ** P < 0.01, Bonferroni corrected). AI: τ = 0.024, P = 0.0006, Bonferroni corrected P = 0.0024; ACC: τ = 0.019, P = 0.0004, Bonferroni corrected P = 0.0017; VTA: τ = 0.008, P = 0.072; SN: τ = 0.004, P = 0.46, Bonferroni corrected P = 1.0. n = 20 participants ( × 4 sessions per participant). Source data
Extended Data Fig. 8 Representational Similarity Analysis links neural patterns to sampling behaviour. a, Schematic illustrating the relationship between brain activity and behavior. b, Example Neural RDM (left, based on multi-voxel BOLD activity patterns) and Behavioral RDM (right, based on the number of samples taken per trial) for a participant session. Each cell represents the dissimilarity between a pair of trials; higher values (yellow) indicate greater dissimilarity. c, Kendall’s τ correlation coefficients quantifying the relationship between neural and behavioral RDMs for VTA, AI, ACC, and SN. Neural RDMs for AI, ACC, and SN were derived using masks matched in shape and size to the VTA ROI to control for region size. Each small dot represents the correlation calculated for one participant session; larger dots indicate the mean correlation across participants. Asterisks denote significant positive correlations across participants (two-sided Wilcoxon signed-rank test; ** P < 0.01, Bonferroni corrected). AI: τ = 0.024, P = 0.0006, Bonferroni corrected P = 0.0024; ACC: τ = 0.019, P = 0.0004, Bonferroni corrected P = 0.0017; VTA: τ = 0.008, P = 0.072; SN: τ = 0.004, P = 0.46, Bonferroni corrected P = 1.0. n = 20 participants ( × 4 sessions per participant). Source dataCC BY · PubMed Central

Verification

Claims the paper makes

10 Valid · 2 Not valid · 5 Not checked · 134 Own results and interpretations

73 more claims

Create a free account to see all 151 claims and to verify the 22 references still open.

See all claims — free account
  • Probably not a substantial claim. 37
  • It may state two separate claims as one. 46
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • Consistentz = -1.01reported p = .31 · recomputed p = .312

    “z = −1.01, p = .310”

  • Consistentz = 1.41reported p = .158 · recomputed p = .159

    “z = 1.41, p = .158”

2 of 2

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 69 references

Research works printed but not linked: 5

  1. Green, D. M. & Swets, J. A. Signal Detection Theory and Psychophysics (Wiley, 1966).
  2. Zupko, J. John Buridan: Portrait of a Fourteenth-Century Arts Master (Univ. Notre Dame Press, 2003).
  3. Sutton, R. S. & Barto, A. G. Reinforcement Learning: An Introduction, Vol. 1 (MIT Press, 1998).
  4. Weinhardt, D., Eckstein, M. K. & Musslick, S. Computational discovery of human reinforcement learning dynamics from choice behavior. OpenReview.net https://openreview.net/forum?id=x2WDZrpgmB (2024).
  5. Wang, R. & Manchester, I. Direct parameterization of Lipschitz-bounded deep networks. In Proc. 40th International Conference on Machine Learning (eds Krause, A. et al.) 36093–36110 (ACM, 2023).

Cited by

See all 1

Metadata

Authors
D’Ambrogio, S., Grohn, J., Khalighinejad, N., Mattar, M. G., Hunt, L., & Rushworth, M. F. S.
Affiliations
  • University of Oxford (GB)
  • New York University (US)
Corresponding author
D’Ambrogio, Simone
Venue
Nature Neuroscience
Published
2026-06-26
Volume
29
Issue
8
Pages
2036-2047
ISSN
  • 1097-6256 (ISSN-L)
  • 1546-1726
OpenAlex
W4411641310
Cited by
1
References
69
Citations by year
2026: 1
FWCI
4.12
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Neuroscience (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by)
  • Research Square (acceptedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
Funders
  • RCUK | Biotechnology and Biological Sciences Research Council (BB/W003392/1)
  • Wellcome Trust (221794/Z/20/Z)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural and Behavioral Psychology Studies
  • Social Sciences › Decision Sciences › General Decision Sciences › Decision-Making and Behavioral Economics
  • Life Sciences › Neuroscience › Cellular and Molecular Neuroscience › Neurotransmitter Receptor Influence on Behavior
Keywords
value of information, information seeking, decision-making, symbolic regression, ventral tegmental area, anterior cingulate cortex, anterior insula, evidence accumulation
MeSH terms
Soft Computing, Adult, Brain, Brain Mapping, Choice Behavior, Decision Making, Female, Humans, Magnetic Resonance Imaging, Male, Neural Networks, Computer, Ventral Tegmental Area, Young Adult

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-28, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-28 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Interpretable abstractions of artificial neural networks predict behavior and neural activity during human information gathering. https://app.cochat.ai/papers/10.1038/s41593-026-02342-9
The paper
D’Ambrogio, S., Grohn, J., Khalighinejad, N., Mattar, M. G., Hunt, L., & Rushworth, M. F. S. (2026). Interpretable abstractions of artificial neural networks predict behavior and neural activity during human information gathering. Nature Neuroscience. https://doi.org/10.1038/s41593-026-02342-9

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.