CHI ’25, Yokohama, Japan

Large language models demonstrate strong problem-solving abilities through reasoning techniques such as chain-of-thought prompting and reflection.However, it remains unclear whether these reasoning capabilities extend to a form of social intelligence: making effective decisions in cooperative contexts.We examine this question using economic games that simulate social dilemmas.First, we apply chain-ofthought and reflection prompting to GPT-4o in a Public Goods Game.We then evaluate multiple off-the-shelf models across six cooperation and punishment games, comparing those with and without explicit reasoning mechanisms.We find that reasoning models consistently reduce cooperation and norm enforcement, favoring individual rationality.In repeated interactions, groups with more reasoning agents exhibit lower collective gains.These behaviors mirror human patterns of "spontaneous giving and calculated greed."Our findings underscore the need for LLM architectures that incorporate social intelligence alongside reasoning, to help address-rather than reinforce-the challenges of collective action.

Recent advances in reasoning techniques have substantially improved the performance of large language models (LLMs), raising expectations for their ability to provide accurate, truthful, and reliable information. However, emerging evidence suggests that iterative reasoning may foster belief entrenchment, rather than enhancing truth-seeking behavior. In this study, we propose a systematic evaluation framework for *belief entrenchment* in LLM reasoning by leveraging the Martingale property from Bayesian statistics. This property implies that, under rational belief updating, the expected value of future beliefs should remain equal to the current belief, i.e., belief updates cannot be predicted from solely the current belief. We propose the unsupervised, regression-based *Martingale Score* to measure violations of this property, signaling a deviation from the Bayesian ability of updating on new evidence. In open-ended problem domains, including event forecasting, value-laden questions, and academic paper review, we found such violations to be widespread across models, reasoning paradigms, problem domains, and system prompts, where the future beliefs are consistently predictable from the model's current belief, a phenomenon which we term *belief entrenchment*. Through comprehensive experiments, we identify the models (e.g., GPT-4o), reasoning techniques (e.g., chain of thought), and domains (e.g., forecasting) more prone to belief entrenchment. Finally, we validate the Martingale Score by showing that it predicts ground-truth accuracy on problem domains where ground truth labels are available. This indicates that, while designed as an unsupervised metric that operates even in domains without access to ground truth, the Martingale Score is a useful proxy of the truth-seeking ability of the LLM reasoning process.

Human biases toward interacting with similar individuals contribute to the formation of ideological echo chambers and erode trust in others. However, facilitating constructive discussions can potentially counteract these biases and build trust, even when a consensus is not reached. This work explores such potential by developing an algorithmic assessment of linguistic features that promote trust and examining the impact of strategically pairing individuals based on this assessment. Guided by the conversational grounding theory, we first analyze the linguistic features of 123 interpersonal dyadic discussions (2,809 messages) on our online chat system and develop a classifier to identify individuals who use the communication style that promotes trust development. We then conduct a randomized controlled experiment with 530 human subjects in 265 pairs to measure the effect of assigning discussion partners based on the classifier's assessment of participants' prior interactions with a chatbot. Our results show that algorithmically assigned pairs exhibit higher trust in their conversation partners than random pairs, irrespective of opinion similarity. We discuss the implications of our strategic pairing approach for enhancing collaboration and trust in various social settings.

Safe control methods are often designed to behave safely even in worst-case human uncertainties. Such design can cause more aggressive human behaviors that exploit its conservatism and result in greater risk for everyone. However, this issue has not been systematically investigated previously. This paper uses an interaction-based payoff structure from evolutionary game theory to model humans’ short-sighted, self-seeking behaviors. The model captures how prior human-machine interaction experience causes behavioral and strategic changes in humans in the long term. We then show that deterministic worst-case safe control techniques and equilibrium-based stochastic methods can have worse safety and performance trade-offs than a basic method that mediates human strategic changes. This finding suggests an urgent need to fundamentally rethink the safe control framework used in human-technology interaction in pursuit of greater safety for all.

This paper proposes a multi-objective, highly generalized and efficient control framework for a virtual character performing various physical interactions in a dynamically simulated world. The framework comprises 1) a generalized inverse dynamics that determines joint forces satisfying multiple objectives considering priorities, unactuated joints and inequality constraints about contacts, 2) a generalized stabilizer available in various contact situations based on the long-term momentum stabilization and 3) a motion primitive network for realizing composite motions by modularizing and interconnecting motion functions. By applying the proposed framework, various interactions with a virtual humanoid, such as basic reflections and carrying objects, are realized in real time with a tactile sensation through the two-armed multi-fingered haptic device we developed.

This paper discusses a novel approach in developing a texture sensor emulating the major features of a human finger. The aim of this study is to realize precise and quantitative texture sensing. Three physical properties, roughness, softness, and friction are known to constitute texture perception of humans. The sensor is designed to measure the three specific types of information by adopting the mechanism of human texture perception. First, four features of the human finger that were focused on in designing the novel sensor are introduced. Each feature is considered to play an important role in texture perception; the existence of nails and bone, the multiple layered structure of soft tissue, the distribution of mechanoreceptors, and the deployment of epidermal ridges. Next, detailed design of the texture sensor based on the design concept is explained, followed by evaluating experiments and analysis of the results. Finally, we conducted texture perceptive experiments of actual material using the developed sensor, thus achieving the information expected. Results show the potential of our approach.