Hirokazu Shirado

Atsushi Ueshima, Shayegan Omidshafiei, Hirokazu Shirado

Cooperation is challenging in biological systems, human societies, and multi-agent systems in general. While a group can benefit when everyone cooperates, it is tempting for each agent to act selfishly instead. Prior human studies show that people can overcome such social dilemmas while choosing interaction partners, i.e., strategic network rewiring. However, little is known about how agents, including humans, can learn about cooperation from strategic rewiring and vice versa. Here, we perform multi-agent reinforcement learning simulations in which two agents play the Prisoner's Dilemma game iteratively. Each agent has two policies: one controls whether to cooperate or defect; the other controls whether to rewire connections with another agent. This setting enables us to disentangle complex causal dynamics between cooperation and network rewiring. We find that network rewiring facilitates mutual cooperation even when one agent always offers cooperation, which is vulnerable to free-riding. We then confirm that the network-rewiring effect is exerted through agents' learning of ostracism, that is, connecting to cooperators and disconnecting from defectors. However, we also find that ostracism alone is not sufficient to make cooperation emerge. Instead, ostracism emerges from the learning of cooperation, and existing cooperation is subsequently reinforced due to the presence of ostracism. Our findings provide insights into the conditions and mechanisms necessary for the emergence of cooperation with network rewiring.

Yuxuan Li, Kyzyl Monteiro, Hirokazu Shirado et al.

Policymakers in domains such as emergency management, public health, and urban planning must make decisions under deep uncertainty, where outcomes depend on how large populations interpret information, coordinate, and adopt over time. Existing tools only partially support this process: tabletop exercises enable collaborative discussion but lack dynamic feedback, while computational simulations capture population dynamics but are designed for offline analysis. We present WhatIf, an interactive system that enables policymakers to steer, inspect, and compare LLM-powered social simulations in real time. Informed by a formative study in emergency preparedness planning, we derive four design requirements for interactive policy simulations: fluid steering, real-time scale, collaborative exploration, and multi-level interpretability. We developed WhatIf guided by these requirements and evaluated it with five preparedness professionals across three disaster evacuation scenarios. Our findings show that participants used the system as a space for iterative branching and comparison rather than evaluating fixed plans; reflected on tacit planning assumptions when agent behavior violated expectations; surfaced previously unrecognized planning vulnerabilities; and grounded their reasoning in inspectable agent-level cases rather than aggregate outputs alone. These findings suggest broader design implications for LLM-powered social simulation systems: designing such systems as interactive, shared reasoning environments -- rather than offline predictive tools -- can better support expert decision-making under deep uncertainty.

Aoi Naito, Hirokazu Shirado

Artificial intelligence (AI) is understood to affect the content of people's decisions. Here, using a behavioral implementation of the classic Newcomb's paradox in 1,305 participants, we show that AI can also change how people decide. In this paradigm, belief in predictive authority can lead individuals to constrain decision-making, forgoing a guaranteed reward. Over 40% of participants treated AI as such a predictive authority. This significantly increased the odds of forgoing the guaranteed reward by a factor of 3.39 (95% CI: 2.45-4.70) compared with random framing, and reduced earnings by 10.7-42.9%. The effect appeared across AI presentations and decision contexts and persisted even when predictions failed. When people believe AI can predict their behavior, they may self-constrain it in anticipation of that prediction.

Zhonghao He, Tianyi Qiu, Hirokazu Shirado et al.

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 and confirmation bias, 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 are unpredictable from the current belief. We propose the unsupervised, regression-based Martingale Score to measure violations of this property, which signal 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 find such violations to be widespread across models and setups, where the current belief positively predicts future belief updates, a phenomenon which we term belief entrenchment. We identify the models, reasoning techniques, and domains 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 a reasoning process.

Yuxuan Li, Hirokazu Shirado, Sauvik Das

While advances in fairness and alignment have helped mitigate overt biases exhibited by large language models (LLMs) when explicitly prompted, we hypothesize that these models may still exhibit implicit biases when simulating human behavior. To test this hypothesis, we propose a technique to systematically uncover such biases across a broad range of sociodemographic categories by assessing decision-making disparities among agents with LLM-generated, sociodemographically-informed personas. Using our technique, we tested six LLMs across three sociodemographic groups and four decision-making scenarios. Our results show that state-of-the-art LLMs exhibit significant sociodemographic disparities in nearly all simulations, with more advanced models exhibiting greater implicit biases despite reducing explicit biases. Furthermore, when comparing our findings to real-world disparities reported in empirical studies, we find that the biases we uncovered are directionally aligned but markedly amplified. This directional alignment highlights the utility of our technique in uncovering systematic biases in LLMs rather than random variations; moreover, the presence and amplification of implicit biases emphasizes the need for novel strategies to address these biases.

Yuxuan Li, Aoi Naito, Hirokazu Shirado

Multi-agent systems built on large language models (LLMs) promise enhanced problem-solving through distributed information integration, but may also replicate collective reasoning failures observed in human groups. Yet the absence of a theory-grounded benchmark makes it difficult to systematically evaluate and improve such reasoning. We introduce HiddenBench, the first benchmark for evaluating collective reasoning in multi-agent LLMs. It builds on the Hidden Profile paradigm from social psychology, where individuals each hold asymmetric pieces of information and must communicate to reach the correct decision. To ground the benchmark, we formalize the paradigm with custom tasks and show that GPT-4.1 groups fail to integrate distributed knowledge, exhibiting human-like collective reasoning failures that persist even with varied prompting strategies. We then construct the full benchmark, spanning 65 tasks drawn from custom designs, prior human studies, and automatic generation. Evaluating 15 LLMs across four model families, HiddenBench exposes persistent limitations while also providing comparative insights: some models (e.g., Gemini-2.5-Flash/Pro) achieve higher performance, yet scale and reasoning are not reliable indicators of stronger collective reasoning. Our work delivers the first reproducible benchmark for collective reasoning in multi-agent LLMs, offering diagnostic insight and a foundation for future research on artificial collective intelligence.

Yuxuan Li, Hirokazu Shirado

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-of-thought 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.

Faria Huq, Elijah L. Claggett, Hirokazu Shirado

Social cohesion is difficult to sustain in societies marked by opinion diversity, particularly in online communication. As large language model (LLM)-driven messaging assistance becomes increasingly embedded in these contexts, it raises critical questions about its societal impact. We present an online experiment with 557 participants who engaged in multi-round discussions on politically controversial topics while freely reconfiguring their discussion groups. In some conditions, participants received real-time message suggestions generated by an LLM, either personalized to the individual or adapted to their group context. We find that subtle shifts in linguistic style during communication, mediated by AI assistance, can scale up to reshape collective structures. While individual-focused assistance leads users to segregate into like-minded groups, relational assistance that incorporates group members' stances enhances cohesion through more receptive exchanges. These findings demonstrate that AI-mediated communication can support social cohesion in diverse groups, but outcomes critically depend on how personalization is designed.

Yuxuan Li, Sauvik Das, Hirokazu Shirado

There is growing interest in using Large Language Models as agents (LLM agents) for social simulations to inform policy, yet real-world adoption remains limited. This paper addresses the question: How can LLM agent simulations be made genuinely useful for policy? We report on a year-long iterative design engagement with a university emergency preparedness team. Across multiple design iterations, we iteratively developed a system of 13,000 LLM agents that simulate crowd movement and communication during a large-scale gathering under various emergency scenarios. These simulations informed actual policy implementation, shaping volunteer training, evacuation protocols, and infrastructure planning. Analyzing this process, we identify three design implications: start with verifiable scenarios and build trust gradually, use preliminary simulations to elicit tacit knowledge, and treat simulation and policy development as evolving together. These implications highlight actionable pathways to making LLM agent simulations that are genuinely useful for policy.