
Large language models (LLMs) increasingly stand in for human participants, yet a plausible response is not evidence of a human decision process. The gap matters most when agents interact, where locally reasonable choices compound through delays and feedback into collective outcomes. Behavioral models fit decision rules to human data; we ask whether their qualitative content alone can drive LLM agents, an approach we call cognitive grounding. Here we show, across 231 sessions of the Beer Distribution Game, a supply-chain experiment in which players’ orders amplify demand swings, that grounding moves seven LLMs toward human-calibrated behavior. Ungrounded agents reproduce the game’s instability but miss its magnitude, timing and propagation; grounding improves trajectory fit for every model and generally shifts inferred decision parameters toward the human-calibrated reference, although unevenly across roles and models. The effect holds against independent empirical data and under alternative prompting and memory designs. Cognitively grounded LLMs are therefore candidate behavioral surrogates, contingent on task-specific validation. The study shows that guiding AI agents with descriptions of how people make decisions in a supply-chain game helps agents reproduce collective human behavior.
By training a reservoir computing system without access to its intermediate states and by eliminating the coupling between its nodes, a physical learning framework achieves efficient learning both in simulation and on optoelectronic hardware.
Escalating energy demands for high-performance machine learning have sparked growing interest in unconventional computing paradigms rooted in physical systems. Central to this direction is leveraging physical properties to design computing frameworks integrating strong learning capabilities, efficient training strategies and practical implementability. Here we present a simple, efficient and scalable framework aimed at achieving this, validated on an optoelectronic platform. This framework features our proposed training mechanism, state-skipping direct feedback alignment, together with our single-map architecture based on simple one-dimensional dynamics. The former eliminates the need to access the system’s intermediate states during training, substantially simplifying error backpropagation, while the latter removes complex internal interconnections. Together, they enhance training efficiency and practical feasibility. Compared with conventional deep neural networks, our approach substantially reduces computational costs and trainable parameter count while achieving comparable performance. Integrated into modern architectures, our scheme attains improved performance with an approximately 40% reduction in computational resources. Notably, our approach challenges the scaling laws of conventional counterparts, underscoring its strong scalability and practical promise. This study presents a scalable, efficient framework for unconventional computing with physical neural networks. It replaces dense connections with uncoupled one-dimensional dynamics, simplifies training, improves performance and reduces resource use.
Evolution unfolds in populations of reproducing individuals, with the dynamics shaped by numerous factors, including population structure, update rules, mutation rates, traits ('strategies') and the reproductive success derived from traits ('payoffs'). Developing a unifying principle to predict which traits will evolve and when, remains a major challenge. While rigorous mathematical results exist for static population structures, the general case of dynamic structures, where connections between individuals change over time, is more difficult and less understood. Here we present a structure-coefficient theorem for evolutionary games on dynamic networks with exogenous stochastic transitions. The framework enables arbitrary Markovian network dynamics independent of strategic interactions, together with arbitrary update mechanisms, mutation rates, strategy spaces and payoff matrices. We demonstrate that the combined effects of structure, update rules and mutation on evolutionary outcomes can be fully captured by three universal structure coefficients. Further analysis reveals that, although most sampled dynamic networks do not promote prosocial behavior, a class of dynamic core-periphery networks consistently promotes multiple forms of prosocial behavior, including cooperation, altruism, fairness and trust.