Modeling individual and item differences in atypicality inferences with ACT-R

Guifu Liu, Vera DembergarXiv 2026 · arXiv

TL;DR

This paper introduces a resource-rational cognitive model built in the ACT-R architecture to explain why people vary in their ability to draw "atypicality inferences." By grounding model parameters in individual cognitive traits (like Raven’s Matrices scores) and item-level properties (like event typicality), the authors successfully predict human pragmatic reasoning patterns that traditional, fully rational models often miss.

What Problem Does This Paper Address?

When someone tells you, "Mary went to a restaurant. She ate there!", your brain likely does a double-take. Eating at a restaurant is so predictable that mentioning it feels redundant. To make sense of this, humans often perform a "pragmatic repair": we infer that the event must be atypical—perhaps Mary usually only goes to restaurants for drinks, making the fact that she actually ate noteworthy.

However, humans aren't consistent. Some people draw these inferences frequently, while others stick to the literal meaning. Furthermore, some scenarios (like picking toppings for a pizza) are less "typical" than others (like eating at a restaurant), changing the likelihood of an inference. Existing frameworks, like the Rational Speech Act (RSA), assume a "perfectly rational" agent and struggle to account for this messy human variability. This paper aims to bridge that gap using a model that accounts for limited cognitive resources and individual differences.

Methods

The authors developed a computational model using ACT-R (Adaptive Control of Thought—Rational), a cognitive architecture that simulates human thought processes.

The Strategy Selection Loop

Instead of calculating every possible meaning, the model follows a resource-rational flow (see Figure 2):

  1. Start Literal: It first tries to interpret the utterance literally.
  2. Check Felicity: If the literal meaning is redundant (violates Gricean maxims), it receives negative feedback ($F_{neg}$).
  3. Explore Alternatives: If there is still "time" (persistence $\tau$), the model selects a new strategy—like finding a successful accommodation—based on its expected utility ($u$).
  4. Terminate: The process ends when a felicitous interpretation is found or the model runs out of persistence.
Figure 2: An overview of our model’s control flow adapted from Duff et al. (2026). 𝑆is a strategy in reaching an interpre- tation, 𝐴is a distinct accommodation for the atypical action, 𝑡is the model runtime, and 𝜏is persistence. Any successful accommodation strategy is felicitous.

Grounding Parameters

To avoid "cheating" by fine-tuning the model to the test data, the authors estimated parameters independently:

  • Individual Differences: Persistence ($\tau$) was mapped from participants' Raven’s Progressive Matrices scores (a measure of fluid intelligence). Higher scores equal higher persistence in searching for pragmatic meanings.
  • Item Differences: The utility of literal interpretation ($u_0$) was derived from human habituality ratings. The utility of accommodations ($u_1$) was estimated using GPT-J-6B surprisal—essentially using a language model to judge how plausible a specific "explanation" for the atypicality would be.

Datasets

The study utilizes data from two primary sources:

  • Ryzhova et al. (2023): Provided the core test set of 1,056 annotated human responses across 24 story items and 193 participants, including their Raven’s Matrices scores.
  • Kravtchenko and Demberg (2022b): Provided the experimental items (brief stories like the restaurant example) and the human habituality ratings used to set item-level priors.

Key Insights and Ideas

  • Pragmatism as Persistence: The model suggests that "being pragmatic" isn't just about being smarter; it's about being more persistent. Individuals with higher fluid intelligence (Raven's scores) are modeled as having higher persistence, allowing them to explore the hypothesis space longer before giving up on a redundant utterance.
  • Individual > Item: While both matter, individual differences in cognitive traits had a much larger impact on inference rates than the specific details of the story items.
  • Resource Rationality: The model captures the "trade-off" humans make. We don't always find the most "rational" interpretation; we find the best one we can within our cognitive limits. As shown in Figure 3, increasing persistence leads to a higher proportion of successful accommodations.
Figure 3: The effect of persistence and initial utilities on inference rates.

Key Results

The authors compared their full model (Indiv + Item) against several baselines. The full model achieved the highest overall accuracy (57.31%) in predicting whether a specific individual would draw an inference on a specific item.

Model Accuracy F1 Score Sensitivity Specificity
Baseline (Fixed) 51.81% 59.25 61.78 38.76
Indiv + Item (Full) 57.31% 64.29 67.77 43.57

Qualitatively, the model showed that adding item-level information specifically helped detect inferences in "pragmatic" participants, though it slightly struggled with "literal" participants who consistently ignored context (see Figure 4).

Figure 4: Per-participant accuracy difference between the full Indiv + Item and Indiv model for least and most pragmatic participants (by their atypicality rates in Ryzhova et al., 2023).

Limitations and Future Work

  • The "Notice Reject" Problem: Currently, the model struggles to account for "notice reject" cases—where a person realizes an inference is intended but rejects the explanation as implausible (e.g., "Mary wouldn't go to a restaurant just for a drink after a long day"). The authors suggest adding an evaluative step to check the plausibility of the accommodation.
  • Static Predispositions: The model doesn't yet account for how a person's personal experience might dynamically change their item-level predispositions.
  • Online vs. Post-hoc: The human data was collected via post-hoc explanations. It remains an open question if these same cognitive loops happen in real-time during natural conversation.

Summary updated

Abstract

Humans engage in pragmatic reasoning when encountering mentions of highly predictable events. Such mentions may appear redundant, leading comprehenders to infer that the event is atypical, and accommodate this inference with an explanation for why it is atypical. However, these inferences are not drawn uniformly across individuals, and experimental items vary in inference rates due to differences in event typicality. We show that these patterns can be predicted with a resource-rational cognitive model in which comprehenders select interpretation strategies based on individual exploration tendencies and item-level predispositions. These parameters are estimated independently from participants' Raven's Matrices score, event typicality rating and language model based accommodation likelihood. Taken together with recent findings from reference games, our results motivate a new perspective on the mechanisms that support pragmatic inferences.
BibTeX
@misc{liu2026modeling,
  author = {Guifu Liu and Vera Demberg},
  title = {{Modeling individual and item differences in atypicality inferences with ACT-R}},
  year = {2026},
  eprint = {2611.11416},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.11416},
}
Figure