ICCM virtual talks and posters

In formal semantics, the interpretation function maps linguistic expressions, “the pencil is in the cup,” to set-theoretic objects, in(pencil, cup), that mirror the compositional structure of the sentence, but do not encode soft inferences like “the pencil is vertical” that human comprehenders automatically make. Distributed situation models (Frank et al., 2009) were proposed as Gestalt / distributed representations of sentence meanings that enable direct probabilistic inference from their representational geometry. We show that previous implementations of distributed situation models systematically underestimate probabilistic inferences because of their sequential sampling of situations (possible worlds). We propose generating the truth-conditional training data using Markov Chain Monte Carlo sampling over a precomputed pool of consistent situations, which accurately captures both hard and soft inferences. Training a competitive neural model on 250,000 MCMC samples yields distributed situations models that accurately encode these inferences, providing reliable, rich meaning representations for cognitive models of language comprehension.

Human curiosity is an intrinsic motivation that sustains learning and exploratory behavior, and clarifying the correspondence between its computational principles and their realization processes remains an important challenge. In recent years, computational models based on Bayesian inference have been developed as a framework for explaining what human cognition computes and why such computations are appropriate. However, how these computational assumptions are concretely realized as information-processing processes has not been sufficiently articulated. In this study, we propose a framework for connecting algorithmic-level curiosity models with computational-level explanations based on Marr’s three levels of analysis. Specifically, we analyze simulation results from an intellectual curiosity model implemented in the cognitive architecture ACT-R using a maze exploration task. For multiple models differing in their levels of cognitive processing, relationships among several simulation metrics are described using Bayesian networks. The results reveal that, in models involving knowledge utilization, intrinsic rewards corresponding to curiosity influence performance through sustained exploration and knowledge generation, forming a coherent dependency structure. This study provides a method for organizing the behavior of existing algorithmic-level models in a computationally interpretable manner and offers a practical approach for bridging the computational and algorithmic levels of explanation.

Research into concept formation is concerned with the fundamental question of how we categorise objects into a particular class. We offer a general framework for designing and testing human benchmarks of subjective judgments as well as their simulations. Our method accounts for subjectivity, ways to control for individual human’s prior knowledge, all while keeping to real-life complex categories. We apply the above framework to experimentally establish a benchmark of subjective judgements in the domain of English literature. We establish individual prior knowledge by conducting structured interviews with six English literature students which are then tasked with categorising unseen texts. We conclude the study by creating bespoke models for each of the students, with the models having the training and testing data of their respective human participants. The models produced good fit to the human data and were further evaluated by five progressively more stringent metrics that go beyond the traditional statistical significance measures. Each of the three parts of our approach – the framework, the benchmark, and the model – may be used separately from each other in future psychological applications that move away from modelling the average participant and towards capturing subjective concept space and individual differences in judgment.

Due to the recent advances in artificial intelligence and large language models (LLMs), applications that involve (semi-) autonomous agents and reasoning capabilities became not only possible, but also increasingly accessible for a broader user-base. While the reasoning capabilities of LLMs are a thriving field of research, a problem for human cognitive science research arises: Studies performed online can now be solved by AI agents more easily and to a much greater extent than ever, with attention checks and captchas being less of a hurdle to bots. We investigate the issue based on a variety of reasoning tasks that have previously been used in online experiments. Using the modern agent capabilities of LLMs to interact with web browsers, we tested their ability to participate autonomously in studies. We instructed them to behave human-like, simulating how actual users might use them to participate on their behalf. Our work focuses on three key questions, namely (I) how well the agents can understand and navigate through the studies from a technical standpoint, (II) if the resulting study data is easily recognizable and distinguishable from human behavior, and (III) how accessible the respective agents are in terms of cost and effort to set them up. We test ChatGPT and Claude as two of the best LLMs for agent usage but also include models running locally on PC hardware. The results are critically discussed focusing on the current implications for online studies, but also an outlook on how to handle this issue in the future.

The ability of the human mind to create, modify, and suppress associations (so-called cue-target-chunks) between previously unrelated items is a fundamental cognitive function. We reanalyzed and extended a classical paired-associate learning (PAL) paradigm from Anderson (2007) by investigating how modality differences (verbal-numerical vs. spatial-numerical) influence cognitive processing. We adopted a sequential interference paradigm with a within-subject design where position-word-numbers (PAL-PWN) follows position-numbers (PAL-PN), using the same position-number pairings. Data from 16 participants (4,440 trials) were analyzed. Results indicate that PAL-PN was the harder task overall but the learning trajectory for those who do succeed was comparable to PAL-PWN. Specifically, PAL-PN shows a learning trajectory while PAL-PWN shows a floor-compressed accuracy range. A spatial/number confusion error analysis reveals that the errors were spatially structured. Implications for the mechanisms of updating learned associations of different modality information are discussed.

Building cognitive process models for reasoning traditionally involve hand-crafting them from underlying theories and assumptions, thereby limiting comparability and transfer between competing theories. To address this, we present a novel approach that automatically constructs process models: Process models are treated as a sequence of cognitive actions/operations that are extracted from cognitive theories and represented in a unified framework while preserving the explanatory merits of the underlying theories. Our method then searches for an optimal sequence to fit the observed reasoning behavior, providing insights into inter-individual differences on the level of reasoning processes. We apply our approach to the domain of syllogistic reasoning and use it to obtain insights into which processes and parts of state-of-the-art models account best for individual reasoning behavior. We use datasets including other reasoning domains as well as the progression of syllogistic reasoning performance over time. This allows us to investigate how processes and components of state-of-the-art models correspond to an individual’s performance in related domains and how they can account for learning effects. Finally, we discuss the potential of our approach for cognitive modeling: First, by utilizing a unified framework for cognitive actions, we can obtain process models combining the best components of state-of-the-art models to account for observed phenomena. Second, it allows us to relate specific processes better to other domains of reasoning, advancing a unified understanding of reasoning. Finally, in-depth insights into the respective theories and models are gained in an objective manner that facilitates the development of cognitive process models.

We developed a computational cognitive theory of visuospatial relation tracking to model how people perceive and monitor spatial relations such as left of and next to. Previous work suggests that they build mental simulations of their environment, or of imagined or described scenarios, to reason about space and spatial concepts – but no account explains how visual information is used to construct and update those simulations. Our model converts perceptual data (objects and locations detected by a convolutional neural network) into an integrated sparse iconic simulation of the scene – a perceptual model. Perceptual models are cognitively plausible representations that are more noise tolerant and stable than raw percepts. They allow for efficient latent encoding of high-level relations. We tested the cognitive model against a benchmark dataset for spatial relation recognition, and show a close fit between human and model-based perception of 2D spatial relations.

Social norms can provide immediate gains through local alignment with others, yet they are also constrained by collective consequences arising from the accumulation of individual actions. Grounded in dual-process theory, this study investigates how normative behavior is formed, bifurcates, and stabilizes when two competing signals operate on different time scales: a short-term local conformity reward and a delayed long-term collective penalty. Agents repeatedly choose between two actions (pull/keep). In each round, they receive a reward proportional to their agreement with neighbors; at fixed intervals, an assessment of collective harm is imposed as a delayed penalty. Decision making follows Instance-Based Learning Theory (PyIBL). We conducted simulations manipulating cognitive reward structure (short- vs. long-term incentives) and social relations. Overall, agents exhibited a bias toward utilitarian decisions, while outcomes frequently showed strong path-dependent polarization. This polarization was more pronounced in small-world networks that approximate human social structures. Cognitively, manipulating memory decay revealed that higher decay rates strengthened the influence of long-term rewards, shifting attention beyond immediate conformity gains. Additional simulations that removed long-term rewards showed that these social effects persisted while utilisation bias was diminished. These results show that social norms emerge from the interaction between cognitive processes and social structure.

A well-established claim in sentence comprehension research is that the subject-verb dependency is resolved via a cue-based retrieval process: The subject noun is searched and retrieved from memory using retrieval cues, such as subject and animate. The cue-based retrieval process can be hindered by decay in the accessibility of the subject noun when the subject and the verb are kept away from each other, causing a processing difficulty at the verb, called the locality effect. The locality effect is robustly observed in reading studies from English, Spanish, Danish, and Russian. However, in verb-final languages like German and Hindi, an opposite effect is observed: When the distance between the subject and the verb increases, a speedup is observed in reading times at the verb. This effect has been called anti-locality and has been attributed to the strong predictability of the upcoming verb in rich case-marking, verb-final languages. We implement this idea as a computational model within the cue-based retrieval architecture. Our model assumes that pre-verbal nouns may preactivate the upcoming verb phrase in memory, and if the preactivation of the verb is strong enough, it can override the cost of retrieval, causing an anti-locality effect. Under the assumption that preactivation is stronger and more relied upon in verb-final languages, the model can account for locality as well as anti-locality effects across languages. We find strong evidence for the predictive activation model compared to the retrieval-only model, given data from five published studies on locality.

Sleep deprivation is common amongst the student community. Poor sleep quality has been associated with reduced cognitive health, specifically a reduction in the ability to form and consolidate long-term memories, which leads to poor academic performance. However, while severe sleep reduction has been studied, the effect of small amounts of sleep loss on cognition is not as well understood. Additionally, while reduced sleep duration has been shown to negatively affect long-term memory, sleep regularity is is also emerging as critical aspect of sleep health. However, it is unknown how sleep duration and regularity interact to shape memory consolidation. Lastly, it is difficult to compare and generalize memory test data due to the lack of standardized cognitive tests. To address this gap, 29 undergraduate students wore sleep tracking watches over three weeks while completing a weekly cognitive test, the Seattle-Groningen Memory Assessment (SGMA). The SGMA is a model-based cognitive assessment with high sensitivity and specificity capable of reliably tracking episodic memory longitudinally. Results show that sleep deprivation is directly correlated to memory performance the following day, with reduced sleep leading to worse memory performance. However, the effect of sleep deprivation on memory performance is modulated by sleep regularity, with high sleep regularity having a compensatory effect on acute low sleep duration. Sleep regularity alone did not affect cognitive scores. These findings act as a demonstration of the multifaceted nature of sleep on memory, and highlight the importance of regularity in situations where sleep duration might be limited.

Instance-Based Learning (IBL) is a general-purpose mathematical framework that models human decisions as arising from decaying episodic traces of previous interactions. Because of its generality, it has been often used as an alternative to reinforcement learning (RL) in a variety of situations. But it is better; much better, indeed. In fact, one might reasonably ask why we ever bothered with anything else. Sed ut perspiciatis unde omnis iste natus error sit voluptatem accusantium doloremque laudantium, totam rem aperiam, eaque ipsa quae ab illo inventore veritatis et quasi architecto beatae vitae dicta sunt explicabo. Nemo enim ipsam voluptatem quia voluptas sit aspernatur aut odit aut fugit, sed quia consequuntur magni dolores eos qui ratione voluptatem sequi nesciunt. Neque porro quisquam est, qui dolorem ipsum quia dolor sit amet, consectetur, adipisci velit, sed quia non numquam eius modi tempora incidunt ut labore et dolore magnam aliquam quaerat voluptatem. Ut enim ad minima veniam, quis nostrum exercitationem ullam corporis suscipit laboriosam, nisi ut aliquid ex ea commodi consequatur? Quis autem vel eum iure reprehenderit qui in ea voluptate velit esse quam nihil molestiae consequatur, vel illum qui dolorem eum fugiat quo voluptas nulla pariatur

Intrusive memories are one of the core symptoms of Post-Traumatic Stress Disorder (PTSD) and can be conceptualized as involuntary memory retrievals that can vary across individuals. The field of computational psychiatry seeks to develop mechanistic models that can capture this heterogeneity by linking individual-level cognitive parameters to clinically relevant outcomes. In our previously conducted study, we used a two-day image presentation task that captured the spacing effect in memory intrusions. Using the exact timing for each stimuli presentation in the study and the individual-level Seattle-Groningen Memory Assessment (SGMA) index, we computed an ACT-R memory model derived base-level memory activation. Then, we used logistic regression to recover additional parameters of the ACT-R memory model including spreading activation and an emotional intensity component, using the experimental recognition data. The new, comprehensive model of memory was able to predict the frequency of memory intrusions with a better fit to the experimental data than the base model. These analyses provide a modeling framework for predicting long-term retention of emotional memories from spacing intervals and individual cognitive parameters, an example of computational phenotyping, and highlights the potential of individualized predictions in clinical psychology.

A well-established empirical phenomenon in human sentence comprehension is the locality effect: reading time at a verb is slower when its argument (e.g., the subject) is placed linearly farther away than when it is kept closer to the verb. The locality effect has been robustly observed across languages and attributed to working memory constraints: due to limited working memory, it becomes difficult to maintain the accessibility of the argument nouns as the distance between the argument and the verb increases. However, working memory-based theories fail to explain the anti-locality effect. In verb-final languages like German and Hindi, increased distance between the arguments and the verb causes a speedup in reading times at the verb. Anti-locality effects have been attributed to predictive processing. However, it remains unclear how predictive processing produces the anti-locality effect and why the locality effect is not found in verb-final languages. We explore the Vasishth and Lewis (2006) proposal, the VP-reactivation hypothesis: intervening items between the subject and the verb reactivate the prediction of the upcoming verb phrase. As distance increases, the verb phrase becomes highly preactivated in memory, facilitating processing at the verb and, consequently, overriding the locality effect. We computationally implement this proposal within a cue-based retrieval framework and conduct two self-paced reading experiments on Hindi that manipulate whether the intervening material directly modifies the upcoming verb or not. The reading data support the key prediction of the predictive reactivation model: anti-locality effects are stronger when preverbal material directly modifies the upcoming verb.

Working memory constraints have been shown to be pervasive in human sentence comprehension. Working memory-based theories of sentence comprehension predict the locality effect: when the distance between two linguistically related words (e.g., a subject and a verb) increases, the processing difficulty increases. While the locality effect has been consistently observed in English, there is no sufficient evidence for locality in verb-final languages such as Hindi and German. The absence of locality in verb-final languages is attributed to robust prediction of the verb, which may override certain working memory demands. We take an individual differences approach to study how working memory and prediction interact. In four self-paced reading experiments in Hindi, we consistently find an anti-locality effect: Speedup at the verb when the argument-verb distance is increased. However, individuals differ in their extent of locality effect. Why do certain individuals in a language population show stronger anti-locality effects than others? We test two competing hypotheses: (i) Varying prediction-strength hypothesis: Individuals with stronger predictability of the verb would show larger anti-locality effects and faster reader times, on average; (b) Varying lexical-access hypothesis: Individuals with poorer lexical accessibility are, on average, slower in reading and may capitalize more on predictability, showing larger anti-locality effects. The two hypotheses make opposite predictions about the correlation between the individual-level locality effect and reading speed. We find strong evidence that slow readers exhibit stronger anti-locality effects. The results support the varying lexical access hypothesis, implying that individuals with slowed lexical access compensate with predictive processing.

Cue-based retrieval theories of sentence comprehension as sume that syntactically related words (e.g., the subject and the verb) are identified and linked together through a content addressable search in memory, using retrieval cues such as subject and animate. A key prediction is similarity-based interference: When multiple nouns in memory match retrieval cues at the verb, it becomes difficult to retrieve the target subject noun, causing a slowdown in processing time at the verb. Similarity-based interference is consistently observed in reading studies from English, furnishing strong evidence for cue-based retrieval theories. However, in verb-final languages such as German and Hindi, evidence is equivocal. The classi cal cue-based retrieval models fail to explain interference ef fect patterns observed in verb-final languages. Data demand a new theory that can explain interference effects in verb-final as well as verb-medial languages. We explore the predictive memory activation mechanism: the pre-verbal nouns stored in memory strongly preactivate the upcoming verb phrase, which can override the cost of retrieval interference in verb-final lan guages. We implement this mechanism within a cue-based re trieval architecture. We show that the predictive memory ac tivation model can account for interference effects observed in verb-final as well as verb-medial languages. Bayes fac tor analysis revealed strong evidence in favor of the predic tive memory activation model against the cue-based retrieval model. Our modeling results have an important theoretical im plication: Dependency completion is driven by top-down pre dictive activation and bottom-up cue-based retrieval processes.