accepted posters

*presenting author

Poster #181 

Title: Belief Tracing, Explanation, and Counterfactual in Multi-Robot Teamwork

Authors: Sridhar Sola*

Abstract: Existing AI agents are limited by an inability to explain their decision and beliefs to humans. This is rooted in the architecture of the agent. Hybrid knowledge-based agents, however, have shown convincing ability to surpass this limitation. We illustrate how, by exploiting the non-monotonic logic-based knowledge representation in one such architecture, we can extract causal chains that describe how a belief evolved over an episode. We can then construct explanations as answers to questions like What Plan?, Why Action A? Why Not Action A?, and counterfactuals, i.e., What If A Happened? Looking under the hood of the agent's causal reasoning increases interpretability. We then ground an understanding of the concept of explanation in abductive reasoning and causality to demonstrate that our subsystem of explanations is, in fact, explanation.

Poster #182

Title: Caregiver presence influences the explore-exploit tradeoff

Authors: Annya Dahmani*, Dorsa Amir, Alison Gopnik

Abstract: The decision to explore a novel option or exploit a known one — referred to as the explore-exploit tradeoff — has received much attention from diverse fields of research, ranging from computer science to developmental psychology. However, much of the work on this topic has focused exclusively on an individual acting alone, a scenario which does not fully capture the rich social dynamics of human decision-making in which other people can influence the costs of exploration. Specifically, one factor which may affect the explore-exploit tradeoff is the presence of caregivers, who provide added safety in children’s environments, which may promote explorative behavior in early life. In the first test of this hypothesis, we investigate whether children think that other children are more likely to explore when caregivers are present. In this ongoing study, American children, ages 4 to 8, are presented with vignettes of children making decisions to explore novel options or exploit known ones. In the vignettes, the characters either make decisions alone, in the presence of a peer, or in the presence of a parent. We predict that participants will be more likely to make exploration predictions when the characters are in the presence of caregivers compared to when they are alone or with peers. The results of this study can help us better understand how children conceptualize the impact of caregiver presence on decision-making. 

Poster #183

Title: Children prioritize purely exploratory actions in observe-vs.-bet tasks

Authors: Eunice Yiu*, Kai Sandbrink, Eileen Liu, Alison Gopnik

Abstract: In causal learning, agents often need to make decisions between selecting actions that are familiar and have previously caused positive results (exploitation), and seeking new information that could allow them to uncover more effective causal actions (exploration). Understanding how humans learn their sophisticated exploratory strategies over the course of their development remains an open question for both computer and cognitive science. Existing studies typically use classic bandit or gridworld tasks that confound the rewarding with the informative causes of an outcome. In this study, we adopt an observe-vs.-bet task that separates “pure exploration” from “pure exploitation” by giving participants the option to either observe an instance of an outcome and receive no reward, or to bet on one action that is eventually rewarding, but offers no immediate feedback. We collected data from 33 five-to-seven-year-old children who completed the task at one of three different bias levels. We compared how children performed with both approximate solutions to the partially-observable Markov decision process and meta-reinforcement learning models that was meta trained on the same decision making task across different probability levels. We found that the children observe significantly more than the two classes of algorithms and qualitatively more than adults in similar tasks. We then quantified how children’s policies differ between the different efficacy levels by fitting probabilistic programming models and by calculating the likelihood of the children’s actions under the task-driven model. The fitted parameters of the behavioral model as well as the direction of the deviation from neural network policies demonstrate that the primary way children adapt their behavior is by changing the amount of time that they bet on the most-recently-observed arm while maintaining a consistent frequency of observations across bias levels, suggesting both that children model the causal structure of the environment and a “hedging behavior” that would be impossible to detect in standard bandit tasks. The results shed light on how children reason about reward and information, providing an important developmental benchmark that can help shape our understanding of human behavior that we hope to investigate further using recently-developed neural network reinforcement learning models on reasoning about information and reward.

Poster #184

Title: A Combined SFT/MRT Approach to Measuring and Predicting Performance with Multimodal Signals

Authors: Sarah Sinclair-Amend*, Ying-Yu Chen, Mark Parent, Kash Todi, and Joseph Houpt

Abstract: Many technologies leverage multiple modalities to deliver information to a user, especially auditory and visual information. In some cases, dividing up information across modalities can allow more information to be communicated, however the information also has the potential to overwhelm a user.  Multiple resource theory (MRT) is one model of how information sources that vary in terms of modality and content (e.g., semantic versus spatial information) may or may not interfere with each other.  We developed a model that leverages both the qualitative structure of MRT and the quantitative precision of systems factorial technology. This model includes parameters associated with individual workload levels within and between modalities across the perception and action cycle. This allows for post hoc assessment of performance, and quantitative predictions for future performance, particularly for counterfactual predictions.  These findings will support individualized, adaptive information delivery to users while minimizing overwhelm.  We will present the model along with a simulation study and an application to an audio-visual motion discrimination task.

Poster #185

Title: Comparing Machines and Children: Using Developmental Psychology Experiments to Assess the Strengths and Weaknesses of LaMDA Responses

Authors: Eliza Kosoy, Emily Rose Reagan*, Leslie Lai, Alison Gopnik, Danielle Krettek Cobb

Abstract: Developmental psychologists have spent decades devising experiments to test the intelligence and knowledge of infants and children, tracing the origin of crucial concepts and capacities. Moreover, experimental techniques in developmental psychology have been carefully designed to discriminate the cognitive capacities that underlie particular behaviors. We propose this metric as a tool to aid in investigating LLMs’ capabilities in the context of ethics and morality. Results from key developmental psychology experiments have historically been applied to discussions of children’s emerging moral abilities, making this work a pertinent benchmark for exploring such concepts in LLMs. We propose that using classical experiments from child development is a particularly effective way to probe the computational abilities of AI models in general and LLMs in particular. First, the methodological techniques of developmental psychology, such as the use of novel stimuli to control for past experience or control conditions to determine whether children are using simple associations, can be equally helpful for assessing the capacities of LLMs. In parallel, testing LLMs in this way can tell us whether the information that is encoded in text is sufficient to enable particular responses, or whether those responses depend on other kinds of information, such as information from exploration of the physical world. In this work we adapt classical developmental experiments to evaluate the capabilities of LaMDA, a large language model from Google. We propose a novel LLM Response Score (LRS) metric which can be used to evaluate other language models, such as GPT. We find that LaMDA generates appropriate responses that are similar to those of children in experiments involving social and proto-moral understanding, perhaps providing evidence that knowledge of these domains is discovered through language. On the other hand, LaMDA’s responses in early object and action understanding, theory of mind, and especially causal reasoning tasks are very different from those of young children, perhaps showing that these domains require more real-world, self-initiated exploration and cannot simply be learned from patterns in language input. 

Poster #186

Title: Exploring Causal Representations with the Method of Virtual World Cognitive Science

Authors: Iris Oved*

Abstract: I will explore a method I call Virtual World Cognitive Science, and the light it sheds on old and new questions about the role of causation and causal models in cognition. Advances in artificial intelligence and 3D simulation allow researchers to construct embodied virtual agents that can perceive, interact with, and learn in virtual environments in ways that align with some of our best theories of human cognition. These synthetic, embodied agents open up new possibilities for conducting controlled experiments into core cognitive abilities such as perception, concept acquisition, construction of folk theories, abstract reasoning, planning, joint attention, and language acquisition. I will illustrate the method of Virtual World Cognitive Science with the VoxWorld simulated environment of James Pustejovsky and Nikhil Krishnaswamy (https://voxml.github.io/voxicon/), which is built on the game engine, Unity, and some of their virtual toddlers who learn to stack blocks and communicate about them with language, gesture, and gaze. Besides revealing the role in imbuing representations with aboutness, such causal connections between minds and worlds allow systems to acquire concepts as abstract/amodal representations (as Concept Atomists want) from observed regularities in perception and language (as Concept Empiricists want). 

Poster #187

Title: From Child's Play to AI: Insights into Automated Causal Curriculum Learning

Authors: Annya Dahmani*, Eunice Yiu*, Tabitha E. Lee, Nan Rosemary Ke, Oliver J. Kroemer, Alison Gopnik

Abstract: We study how reinforcement learning algorithms and children develop their causal curriculum to achieve a challenging goal that is not solvable at first. Adopting the Procgen environmaents that comprise various tasks as challenging goals, we found that 5- to 7-year-old children actively used their current level progress to determine their next step in the curriculum and made improvements to solving the goal during this process. This suggests that children treat their level progress as an intrinsic reward, and are motivated to master easier levels in order to do better at the more difficult one, even without explicit reward.  To evaluate RL agents, we exposed them to the same demanding Procgen environments as children and employed several curriculum learning methodologies. Our results demonstrate that RL agents that emulate children by incorporating level progress as an intrinsic reward signal exhibit greater stability and are more likely to converge during training, compared to RL agents solely reliant on extrinsic reward signals for game-solving. Curriculum learning may also offer a significant reduction in the number of frames needed to solve a target environment. Taken together, our human-inspired findings suggest a potential path forward for addressing catastrophic forgetting or domain shift during curriculum learning in RL agents. 

Poster #188

Title: How do language models bind entities in context?

Authors: Jiahai Feng*, Jacob Steinhardt

Abstract: Language models (LMs) can recall facts mentioned in context, for example when solving reading comprehension tasks. When the context describes facts about multiple entities, the LM has to correctly bind attributes to their corresponding entity. We show, via rigorous causal experiments, that LMs' internal activations represent binding information by exhibiting appropriate binding ID vectors at the entity and attribute positions. We further show that binding ID vectors form a metric subspace and often transfer across tasks. Our results demonstrate that LMs learn interpretable strategies for representing symbolic knowledge in context, which provides a foundational step towards interpreting LM reasoning processes.

Poster #189

Title: Learning Causally-Aware Representations of Multi-Agent Interactions 

Authors: Yuejiang Liu*, Ahmad Rahimi, Po-Chien Luan, Frano Rajič, Alexandre Alahi

Abstract: Modeling spatial-temporal interactions between neighboring agents is at the heart of multi-agent problems such as motion forecasting and crowd navigation. Despite notable progress, it remains unclear to which extent modern representations can capture the causal relationships behind agent interactions. In this work, we take an in-depth look at the causal awareness of the learned representations, from computational formalism to controlled simulations to real-world practice. First, we cast doubt on the notion of non-causal robustness studied in the recent CausalAgents benchmark. We show that recent representations are already partially resilient to perturbations of non-causal agents, and yet modeling indirect causal effects involving mediator agents remains challenging. Further, we introduce a simple but effective regularization approach leveraging causal annotations of varying granularity. Through controlled experiments, we find that incorporating finer-grained causal annotations not only leads to higher degrees of causal awareness but also yields stronger out-of-distribution robustness. Finally, we extend our method to a sim-to-real causal transfer framework by means of cross-domain multi-task learning, which boosts generalization in practical settings even without real-world annotations. We hope our work provides more clarity to the challenges and opportunities of learning causally-aware representations in the multi-agent context while making a first step towards a practical solution.

Poster #190

Title: Meta-learning environment-specific learning rates

Authors: Jonas Simoens, Tom Verguts, Senne Braem*

Abstract: People often have to switch back and forth between different environments that come with different volatilities. While some environments require fast learning (i.e., high learning rates), others call for lower learning rates. Previous reinforcement learning studies have shown that people adapt their learning rates to their environment when differences in these statistics are clustered in time. However, these differences in learning rates could still reflect emergent properties of participants’ (non-environment-specific) responses to locally experienced reward prediction errors. As such, it remains unclear whether people can actually learn about environment-specific learning rates, associate them to relevant contextual features, and instantaneously retrieve them when revisiting environments (i.e., meta-learn the learning rate). Across three experiments (n = 273), we demonstrate that people can alternate using two different learning rates, on a trial-by-trial basis, when switching back and forth between two two-armed bandit tasks in two different environments (i.e., casinos) that differ in volatility. Results from a test phase suggest that participants also learned to attribute these different learning rates to their respective environments. However, this difference was small because participants rapidly adapt to newly experienced volatilities. Therefore, we also ran a study (n = 50) that allowed us to estimate learning rates on a trial-by-trial basis to test whether differences in learning rates were already present on the first trial. The model that fitted the data best had location-specific learning rates (especially in early trials), which differed across locations. We conclude that humans can learn to associate different learning rates to different environments, offering important insights for developing theories of meta-learning.

Poster #191

Title: MoCa: Measuring Human-Language Model Alignment on Causal and Moral Judgment Tasks

Authors: Allen Nie*, Yuhui Zhang, Atharva Amdekar, Chris Piech, Tatsunori Hashimoto, Tobias Gerstenberg

Abstract: Human commonsense understanding of the physical and social world is organized around intuitive theories. These theories support making causal and moral judgments. When something bad happens, we naturally ask: who did what, and why? A rich literature in cognitive science has studied people's causal and moral intuitions. This work has revealed a number of factors that systematically influence people's judgments, such as the violation of norms and whether the harm is avoidable or inevitable. We collected a dataset of stories from 24 cognitive science papers and developed a system to annotate each story with the factors they investigated. Using this dataset, we test whether large language models (LLMs) make causal and moral judgments about text-based scenarios that align with those of human participants. On the aggregate level, alignment has improved with more recent LLMs. However, using statistical analyses, we find that LLMs weigh the different factors quite differently from human participants. These results show how curated, challenge datasets combined with insights from cognitive science can help us go beyond comparisons based merely on aggregate metrics: we uncover LLMs implicit tendencies and show to what extent these align with human intuitions.

Poster #192

Title: Novel Insights Into the Wisdom of Crowds by Process-Consistent Modeling

Authors: Tobias R. Rebholz*, Marco Biella, Mandy Hütter

Abstract: Advice taking and related research is dominated by deterministic weighting indices, specifically ratio-of-differences-based formulas for investigating informational influence. Their arithmetic is intuitively simple, but they pose several measurement problems and restrict research to a particular paradigmatic approach. As a solution, we propose process-consistent mixed-effects regression modeling to specify how strongly peoples’ judgments are influenced by externally provided evidence. The formal derivation of the proposed weighting measures is accompanied by a detailed elaboration on their most important technical and statistical subtleties. Essentially, by implementing appropriate multilevel models, our approach explicitly distinguishes between components of endogenous (i.e., final judgments) and exogenous nature, such as independent initial judgments and advice. Corresponding mixed-effects regression coefficients of various exogenous sources of information thus also reflect individual weighting, but are based on a conceptually consistent representation of the endogenous judgment process. We use this modeling approach to revisit empirical findings from several paradigms investigating algorithm aversion, sequential collaboration, and advice taking. In summary, we replicate and extend the original finding of algorithm appreciation, and initially demonstrate a lack of evidence for systematic order effects in sequential collaboration or for differential weighting of multiple individual pieces of advice. In addition to opening new avenues for innovative research, process-consistent modeling of information sampling and utilization has the potential to increase the reproducibility and replicability of our science. Furthermore, the proposed method is relevant beyond advice taking, as mixed-effects regression weights can also inform research on related cognitive phenomena such as multidimensional belief updating, anchoring effects, hindsight bias, or attitude change.

Poster #193

Title: Off The Rails: Procedural Dilemma Generation for Moral Reasoning

Authors: Jan-Philipp Fränken*, Ayesha Khawaja, Kanishk Gandhi, Noah Goodman, Tobias Gerstenberg

Abstract: As AI systems like language models are increasingly integrated into making decisions that affect people, it's critical to ensure that these systems have sound moral reasoning. To test whether they do, we need to develop systematic evaluations. Recent work has introduced a method for procedurally generating LLM evaluations from abstract causal templates, and tested this method in the context of social reasoning (i.e., theory-of-mind). In this paper, we extend this method to the domain of moral dilemmas. We develop a framework that translates causal graphs into a prompt template which can then be used to procedurally generate a large and diverse set of moral dilemmas using a language model. Using this framework, we created the OffTheRails dataset which consists of 50 scenarios and 500 unique test items. We evaluated the quality of our model-written test items using two independent human experts and found that 90% of the test-items met the desired structure. We collect moral permissibility and intention judgments from 100 human crowdworkers and compared these judgments with those from GPT-4 and Claude-2 across eight control conditions. Both humans and GPT-4 assigned higher intentionality to agents when a harmful outcome was evitable and a necessary means. However, our findings did not match previous findings on permissibility judgments. This difference may be a result of not controlling the severity of harmful outcomes during scenario generation. We conclude by discussing future extensions of our benchmark to address this limitation.

Poster #194

Title: Prosocial Interaction Between Human and Autonomous Agent

Authors: Kumar Akash, Teruhisa Misu,  Shashank Mehrotra, Xinyue Hu*, Mark Steyvers

Abstract: In this research project, we aim to investigate prosocial behavior within a hybrid system of human-machine interactions.We explore to what extent humans engage in prosocial behaviors toward autonomous agents. We intend to construct a computational model that explains these complex social behaviors. Our online behavioral experiments involve gathering data on how humans engage in prosocial actions within simple grid-game settings, particularly when interacting with AI agents. These insights will serve as an empirical foundation for developing more effective, prosocially-inclined AI systems, thereby reinforcing prosocial tendencies within hybrid human-machine collectives.

Poster #195

Title: Selection History Modulates the Effect of Automation Accuracy and Task Difficulty on Aided Decision-Making

Authors: Cheng-Ta Yang*, Cheng-You Cheng, Shang Shu Huang, Ming-Hui Cheng,  Peng-Fei Zhu, Hao-Lun Fu

Abstract: We examined the impact of automation accuracy and task difficulty on human decision-making. We hypothesized that highly accurate aids would improve performance only under difficult conditions, and this effect would be influenced by individual selection history. Using a categorization task, we manipulated automation accuracy (high/low) and task difficulty (easy/difficult) with three types of aids presented in separate blocks or randomly intermixed to 36 participants. We used a capacity measure based on the single-target self-terminating (STST) rule within the framework of Systems Factorial Technology (SFT) to assess decision efficiency. Results showed that high-accuracy aids reduced accuracy and increased RTs compared to unaided decisions, regardless of automation accuracy and task difficulty. Notably, high-accuracy aids provided incorrect answers under difficult conditions, leading to a significant decline in performance. However, the STST capacity results showed that high-accuracy aids had supercapacity processing under difficult conditions in the block design, but not in the mixed design. These findings suggest that effective top-down control is essential to utilize high-accuracy aids to improve decision efficiency when the task is relatively difficult. Our study suggests that individuals may rely more on high-accuracy aids as task demands increase. Furthermore, these capacity differences imply that participants may utilize different decision strategies in terms of mental architecture to integrate current percept and aided information for decision-making. Our research provides novel insights into the potential benefits and limitations of automated aids for information processing efficiency. 

Poster #196

Title: Social Contract AI: A Simulator for Learning Social Contracts

Authors: Samuel Kwok*, Peixuan Ye*, Jan-Philipp Fränken, Kanishk Gandhi, Tobias Gerstenberg, Noah Goodman

Abstract: We explore the idea of aligning an AI assistant by inverting a model of users’ (unknown) preferences from observed interactions. To validate our proposal, we run proof-of-concept simulations in the economic ultimatum game, formalizing user preferences as policies that guide the actions of simulated players. We find that the AI assistant accurately aligns its behavior to match standard policies from the economic literature (e.g., selfish, altruistic). However, the assistant’s learned policies lack robustness and exhibit limited generalization in an out-of-distribution setting when confronted with a currency (e.g., grams of medicine) that was not included in the assistant’s training distribution. Additionally, we find that when there is inconsistency in the relationship between language use and an unknown policy (e.g., an altruistic policy combined with rude language), the assistant’s learning of the policy is slowed. Overall, our preliminary results suggest that developing simulation frameworks in which AI assistants need to infer preferences from diverse users can provide a valuable approach for studying practical alignment questions.

Poster #197

Title: The relationships between reward sensitivity, cognitive effort cost, and flexible system-switching in perceptual categorization

Authors: Li Xin Lim* & Sébastien Hélie

Abstract: Much research suggests that switching between perceptual categorization systems is extremely difficult, with about half of the participants failing to switch on a trial-by-trial basis. While Hélie (2017) showed that trial-by-trial switching can be improved by additional training and the inclusion of a switch cue allowing for preparation time, little work has focused on identifying the individual characteristics that (dis)allowed some participants to execute frequent system switching in category learning. As a first attempt, Hélie, Lim, and colleagues (2023) proposed a computational model of the switching circuit which suggested important roles for executive functions and reward processing. They tested the role of executive functions using a battery of working memory (WM) tasks and found evidence that a composite score of WM updating can partially predict switching ability. The present experiment used a within-subject design to test the importance of reward processing as a predictor of system-switching ability. Specifically, participants learned both a rule-based and an information-integration category structure (separately) before being asked to flexibly switch between the two categorization tasks. Critically, some categorization trials included a switch cue while others did not. Participants also performed the shell game task (SGT; to estimate cognitive effort cost) and the Iowa gambling task (IGT; to estimate reward sensitivity). The results show that individuals who learned to use the optimal categorization system in the training phase had a lower cognitive effort cost in the SGT. Additionally, individuals who could switch between optimal systems from trial-to-trial without a preparation switch cue exhibited higher sensitivity to punishment than reward in the IGT. These results provide insights into the individual characteristics that enable system-switching in category learning and can aid in designing effective training and learning programs.

Poster #198

Title: The Role of Causal Stability in Children’s Causal Reasoning

Authors: Ny Vasil*, Anais Jimenez, Kate Marcotullio, Shihan Gao, Samhita Katteri, Mei Murphy, Tania Lombrozo, Alison Gopnik

Abstract: Adults have been shown to favor stable causal relationships – those that hold robustly across background contexts – in their actions and causal/explanatory generalizations (Vasilyeva et al, 2018). Here we explore how this preference develops. We present results from a developmental study with 110 4-7-year-olds investigating whether children pay attention to causal stability when they explain observations and design interventions in novel contexts. We report developmental shifts in reliance on causal stability in a range of inferential tasks, highlight the important role of perceived average causal strength in determining children’s causal preferences, and discuss the implications of our findings for theories of early causal learning.