Code accompanying Ngiam & Lee (2026)

Ngiam, W. X. Q., & Lee, M. D. (2026). Model-based evidence for task-dependent representations in psychological similarity and working memory.
Preprint: https://osf.io/preprints/psyarxiv/fm9vz_v2

This repository contains the MATLAB and JAGS code used to fit and compare generative psychological models of orientation representations across perceptual reproduction, visual working memory, and similarity comparison. The analyses test whether a single shared psychological representation can account for all three tasks, or whether memory and similarity require task-specific distortions relative to perception.

Empirical data are from Tomic & Bays (2024), packaged as data/tomicBays.mat. Each task is implemented with a wrap-copy generative model and, for several tasks, an alternative von Mises generative model—each with its own JAGS specification (.txt) and MATLAB driver (.m).

Models write MCMC output under per-task storage/, task figures under figures/, and paper-level analyses under analysis/.

Requirements

• MATLAB (scripts tested with standard toolboxes; parallel chains use the Parallel Computing Toolbox when doParallel = 1)
• JAGS 4.x
• trinity (callbayes, codatable, grtable, gelmanrubin, etc.)
• Lab plotting utilities in supportingFiles/ (setFigure, Raxes, findKeepChains, pantoneColors.mat)

Add the repo root and supportingFiles/ to the MATLAB path, or run drivers from their task folders as documented below.

Data

Location	Contents
data/tomicBays.mat	Preprocessed perceptual (dp), memory (dm), and similarity (ds) structs used by all drivers
data/*.csv	Raw task exports from the authors
data/parse_1.m	Script that builds similarity trial indices from CSV (see data/README.md)

All task scripts load ../data/tomicBays relative to their folder.

Repository layout

orientationModeling/
├── data/                    # tomicBays.mat and raw CSVs
├── supportingFiles/         # Shared plotting and chain-filter helpers
├── perceptualReproduction/  # Single-stimulus reproduction
├── memoryReproduction/      # Working-memory reproduction (swap + no-swap)
├── similarityComparison/    # Pairwise similarity choice (AB vs CD)
├── commonRepresentation/    # Joint representation across three tasks
├── clusterRepresentation/   # Clustered deviations from perceptual μ
└── analysis/                # Paper figures and trial-level diagnostics

Each task folder typically contains:

• *_jags.txt — JAGS model definition
• *.m — MATLAB driver (preLoad loads cached storage/*.mat when available)
• storage/ — Saved chains (gitignored; produced by fitting)
• figures/ — Model-specific output figures

Wrap-copy generative models

Orientations are modeled on a half-circle, [0, π), because line stimuli are equivalent under 180° rotation. Distances, reproduction errors, and similarity comparisons must therefore respect modulus arithmetic on the half-circle (wrap at 0 and π), not ordinary arithmetic on the real line.

JAGS does not provide a half-circle Gaussian. The wrap-copy construction implements the required geometry for the Gaussian (normal-error) generative model: on each trial a latent index selects among three copies of each mental orientation, μ, μ − π, and μ + π. Mental samples are drawn from a Gaussian centered on the chosen copy; distances use min(|x₁ − x₂|, π − |x₁ − x₂|). That discrete wrap choice is what makes normal errors behave correctly on [0, π).

Task folder	Driver	JAGS model	What it models
perceptualReproduction/	perceptualReproduction.m	perceptualReproduction_jags.txt	Reproduce one orientation
memoryReproduction/	memoryReproduction.m	memoryReproduction_jags.txt	Recall target among foils (with swap process)
memoryReproduction/	memoryReproductionNoSwap.m	memoryReproductionNoSwap_jags.txt	Same without swap errors
similarityComparison/	similarityComparison.m	similarityComparison_jags.txt	Choose more similar pair (censored decision)
commonRepresentation/	commonRepresentation.m	commonRepresentation_jags.txt	One μ map for all three tasks
clusterRepresentation/	clusterRepresentation.m	clusterRepresentation_jags.txt	Perceptual μ plus task-specific shifts (γ)

similarityComparison/similarityWrapChainInits.m supplies censored trial initial values for the similarity driver.

Von Mises generative models (alternative specification)

These are separate generative models, not reparameterizations of the wrap-copy models. Mental variables are von Mises on the doubled circle ([0, 2π)), mapped to half-circle distances without latent ±π wraps. Each task has its own *_vonMises_jags.txt paired with a matching driver:

• perceptualReproduction/perceptualReproduction_vonMises.m
• memoryReproduction/memoryReproduction_vonMises.m, memoryReproductionNoSwap_vonMises.m
• similarityComparison/similarityComparison_vonMises.m
• commonRepresentation/commonRepresentation_vonMises.m

The repository includes the modified von Mises JAGS module under jags-vonMises/ (fork of yeagle/jags-vonmises; see that folder’s README). After clone:

cd jags-vonMises && make

Then set JAGS_LIBS to jags-vonMises/ (or make install system-wide) and pass 'modules', {'vonmises'} through callbayes (see comments in each von Mises driver).

Sampling limitations. Von Mises directions in JAGS can suffer from poor MCMC exploration when the posterior wraps the circle—scalar samplers treat [0, 2π) as an interval with endpoints, so chains may fail to move between modes near 0 and 2π. That issue is discussed for the yeagle module in this Cross Validated thread. In our hands this limits reliable inference most clearly for similarity comparison, where each trial conditions on four mental samples under censoring (dinterval), multiplying the effect of poor circular mixing. The wrap-copy Gaussian models in this repository are the primary fits used in the paper; von Mises drivers are included as an alternative specification, not as a drop-in replacement with equal performance.

Representation plots in both model families use boundary-aware credible intervals on the half-circle (error bars split across 0 and π when the interval wraps).

Stan

This repository does not include Stan code. We attempted Stan implementations but did not obtain workable inference for the full paper models:

• Similarity comparison: the wrap-copy model uses censoring (dinterval on the similarity decision). Stan does not handle this structure in the same way as JAGS. We also tried a trembling-hand variant without censoring, but Stan’s need to represent intermediate mental samples made that approach impractical at scale.
• Memory reproduction: the swap model relies on a discrete latent mixture over recall identity (ξ). A standard Stan implementation with mixture marginalization did not converge in long runs with heavy thinning.
• Perceptual reproduction: a Stan port with the built-in von Mises would likely be feasible, but we did not pursue it once similarity and memory proved problematic.

We would welcome Stan implementations from others who can address these constraints.

Quick start

1. Fit a single task model

cd /path/to/orientationModeling/perceptualReproduction
perceptualReproduction          % set preLoad = false in script to refit

Equivalent entry points exist in each task folder. After a successful run, expect storage/<modelName>_tomicBays_jags.mat and figures under figures/.

2. Fit joint-task models (optional standalone)

commonRepresentation and clusterRepresentation fit all three tasks in one JAGS run. They load only data/tomicBays.mat and write their own storage/*.mat files—they do not read posteriors from the single-task fits.

cd ../commonRepresentation
commonRepresentation

cd ../clusterRepresentation
clusterRepresentation

Fit the single-task models first only if you need their chains for analysis/drawFigures.m (e.g. threeScatter, fourScatter) or for side-by-side comparison with the joint models.

3. Generate cross-task figures

cd ../analysis
drawFigures   % edit analysisList at top of script to enable analyses

Analysis scripts (analysis/)

Script	Purpose
drawFigures.m	Multi-panel paper figures; enable cases via analysisList

drawFigures.m analysis options (uncomment in analysisList):

Name	Description
threeScatter	Perceptual, memory, similarity representation scatters
fourScatter	Adds common representation
twoRepresentations	Polar comparison of perceptual vs similarity μ
clusterSavageDickey	Savage–Dickey plots for cluster γ parameters
statisticalSummaries	Text summaries written under analysis/results/
descriptiveAdequacyPerceptual	Posterior predictive check, perception
descriptiveAdequacyMemory	Posterior predictive check, memory (swap)
descriptiveAdequacyMemoryNoSwap	Posterior predictive check, memory (no swap)
descriptiveAdequacySimilarity	Posterior predictive agreement histogram
swapExamples	Example trials with strong swap inference (memoryReproduction chains)

Outputs go to analysis/figures/ (and analysis/results/ for summaries). Most analyses load precomputed chains from ../<task>/storage/.

Script	Purpose
findDiscriminatingSimilarityTrials.m	Example trials where compressed vs physical μ maps disagree; requires similarityComparison/storage/similarityComparison_mixture_tomicBays_jags.mat

Gitignored paths

.gitignore excludes **/storage/ and **/tmp/. Clone the repo and run drivers to regenerate MCMC files locally, or copy storage/*.mat from a completed analysis environment.

Citation

If you use this code, please cite:

Ngiam, W. X. Q., & Lee, M. D. (2026). Model-based evidence for task-dependent representations in psychological similarity and working memory. PsyArXiv. https://osf.io/preprints/psyarxiv/fm9vz_v3

The empirical dataset should also be cited:

Tomić, I., & Bays, P. M. (2024). Perceptual similarity judgments do not predict the distribution of errors in working memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 50(4), 535–549.