BGPT: Paper Review: A Distributional Response Time Analysis of the Perceptual Disfluency Effect

Explore by Goal

Quick Answer Copied

Core finding

Across three experiments, the paper argues that perceptual blurring during encoding changes *processing-stage timing* (distribution shape in RTs) in a way that can translate into better later recognition—especially for high-blur stimuli. It connects this stage-specific view to both distributional modeling (ex-Gaussian) and process-level decision modeling (drift diffusion). However, cognitive-stage interpretations of ex-Gaussian parameters remain debated, so model-based mechanistic claims should be treated as hypothesis-shaped rather than definitive.

Most decision-relevant outputs (from the paper text)

RT distributions: high-blur increases distribution shift (μ) and slow-tail/skew (β/τ) relative to clear & low-blur.
Memory: high-blur yields higher recognition sensitivity than clear & low-blur.
Boundary conditions: in Experiment 2, memory benefits depend on word frequency (not all “harder/longer tail” situations produce a memory benefit).

Primary source: the paper itself .

Long Answer

Paper Review (Visual + Skeptical): Distributional Response Time Analysis of the Perceptual Disfluency Effect

Primary manuscript: .

Figure 1. Ex-Gaussian directional summary (high-blur vs clear+low-blur)

The paper reports that high-blur increases both typical-speed shift (μ) and slow-tail/skew (β/τ) relative to the combined clear+low-blur comparison in Experiments 1A and 1B. Note: the authors also acknowledge ongoing debate about mapping ex-Gaussian parameters onto cognitive “stages.”

Figure 2. Recognition sensitivity and blur

The paper reports higher recognition sensitivity (d′) for high-blur than for clear and low-blur, and provides Bayesian model estimates.

Figure 3. Boundary condition: blur × word frequency (Experiment 2)

The Experiment 2 results emphasize that not all “more late-stage processing” conditions yield a memory advantage: the paper reports an interaction where the memory benefit of blur is selective with respect to word frequency.

Methods & Modeling (what was done, and what each method can/can’t tell you)

Experimental manipulations

Blur during encoding: clear vs low-blur vs high-blur via Gaussian blur parameters (as described in the methods section of the manuscript).
Tasks: Experiments 1A/1B used a lexical decision task during encoding, followed by surprise old/new recognition; Experiment 2 used semantic categorization (animal vs not) plus word frequency manipulation.

Distributional RT logic (ex-Gaussian)

The paper uses ex-Gaussian RT decomposition to characterize shift and slow-tail properties, consistent with standard ex-Gaussian usage in word recognition where mean/variance/tail effects can diverge.
They explicitly note that mapping ex-Gaussian parameters onto cognitive stages remains debated; i.e., the parameterization can be statistically meaningful while cognitive interpretation can be underdetermined.

Process modeling (drift diffusion model)

The drift diffusion model (DDM) is used as a process model for two-choice speed/accuracy tradeoffs, with drift rate (v) and non-decision time (Ter) offered as plausible handles for “evidence accumulation” vs other components.
The paper also fits a hierarchical Bayesian Wiener diffusion model for two-choice response times, consistent with established hierarchical diffusion modeling approaches.

Bayesian inference & diagnostics

They implement Bayesian multilevel modeling using brms/Stan and report posterior predictive checks, ESS, and convergence diagnostics including R̂ targets.
They mention evidence ratios/generalized Bayes factors and interpret ER thresholds; Bayes-factor-like measures are known to be sensitive to priors and modeling choices.

Critical review (what’s strong, what’s uncertain, what could mislead)

Strengths

Distributional + process-model triangulation: combining ex-Gaussian distributional analysis with DDM provides converging descriptive/procedural constraints—useful because mean-only analyses often miss differential distribution shape changes.
Pre-registration and open materials: Experiments 1A and 2 are preregistered, and the paper provides OSF/GitHub links.
Bayesian diagnostics are reported: the manuscript reports posterior predictive checks and MCMC convergence diagnostics.

Key uncertainties / interpretational risks

Ex-Gaussian “stage” mapping is underdetermined: The paper argues stage-specificity (early vs late) using parameters (μ vs β/τ), but ex-Gaussian parameters do not have a unique cognitive meaning across all tasks/architectures; interpretation depends on the modeling assumptions used to relate distributional parameters to latent processes.
Evidence ratio thresholds: ER/Bayes-factor-like metrics are sensitive to priors and model specification; this can change whether “evidence” crosses reported thresholds.
Ecological generalization: The experimental tasks are laboratory-style word recognition tasks; applied learning contexts may involve different attention, expectancy, and cognitive load dynamics. The disfluency literature itself shows boundary conditions such as test expectancy and task demands.

What would most likely disconfirm the paper’s “stage-specific” story?

If blur manipulations that match the RT distributional shape (μ/β/τ patterns) do not produce the predicted recognition sensitivity changes, then the link between those distributional components and mnemonic benefit would be weak.
If alternative distributional models (or different parameterizations) fit the same data but yield different “early vs late” conclusions, then the cognitive-stage mapping would be underdetermined.
If frequency-dependent interactions disappear under preregistered replications using different stimulus sets, then the boundary condition (that late-stage processing is necessary but not sufficient) would be less stable.

Relation to prior disfluency work (and why this paper’s modeling choice matters)

The paper targets a specific blur-based desirable difficulty effect reported by Rosner et al., emphasizing that not all disfluency manipulations are equivalent.
It also contributes to a disfluency literature with mixed replication outcomes, where boundary conditions such as test expectancy and task demands have repeatedly mattered.
Thus, the modeling contribution is that it tries to explain *when* effects should appear by linking blur to specific distributional properties that can, in principle, be tested by manipulating processing stages more directly.

Judgement (skeptical, evidence-weighted)

Main claim confidence

Moderate: clear blur→RT distribution changes + blur→d′ sensitivity increase in Exp 1; stronger boundary-condition structure in Exp 2.

Mechanism confidence

Limited: ex-Gaussian parameter-to-stage mapping is debated; DDM helps but still depends on model assumptions.

Reproducibility risk

Lower than average due to preregistration + OSF/GitHub, but interpretive reproducibility (model choices, priors, parameterizations) still matters.

Link to bespoke author-by-author reviews

Feedback:

Updated: April 18, 2026