Research

Can large language models (LLMs), when acting as agents, match human cognitive capabilities in sequential reasoning? To answer this question, we designed a novel active probabilistic reasoning task that can be played by humans and LLMs. Our minimal task design allows us to disentangle two essential components of decision-making, sampling (gathering evidence) and inference (evaluating evidence). We evaluated a large set of LLMs and find a wide spectrum of performance. Several frontier models reach human-level performance, but do not exceed skilled human players. Strong model performance consistently relies on extensive reasoning. While some LLMs outperform humans in inference, all models consistently lag in sampling capabilities. To probe the source of these differences, we develop a novel Bayesian modeling framework that tracks sampling-policy updates and maps humans and LLMs to different classical observer models. We show that humans tend toward maximum-a-posteriori (MAP) sampling, whereas the best LLMs tend to minimize posterior entropy across options. We further tested whether LLMs can improve via in-context learning, and found that only a subset of top-performing models could learn to solve the task based only on the outcome of their choices.

Recurrent neural networks (RNNs) provide a powerful approach in neuroscience to infer latent dynamics in neural populations and to generate hypotheses about the neural computations underlying behavior. However, past work has focused on relatively simple, input-driven, and largely deterministic behaviors - little is known about the mechanisms that would allow RNNs to generate the richer, spontaneous, and potentially stochastic behaviors observed in natural settings. Modeling with Hidden Markov Models (HMMs) has revealed a segmentation of natural behaviors into discrete latent states with stochastic transitions between them, a type of dynamics that may appear at odds with the continuous state spaces implemented by RNNs. Here we first show that RNNs can replicate HMM emission statistics and then reverse-engineer the trained networks to uncover the mechanisms they implement. In the absence of inputs, the activity of trained RNNs collapses towards a single fixed point. When driven by stochastic input, trajectories instead exhibit noise-sustained dynamics along closed orbits. Rotation along these orbits modulates the emission probabilities and is governed by transitions between regions of slow, noise-driven dynamics connected by fast, deterministic transitions. The trained RNNs develop highly structured connectivity, with a small set of “kick neurons” initiating transitions between these regions. This mechanism emerges during training as the network shifts into a regime of stochastic resonance, enabling it to perform probabilistic computations. Analyses across multiple HMM architectures — fully connected, cyclic, and linear-chain — reveal that this solution generalizes through the modular reuse of the same dynamical motif, suggesting a compositional principle by which RNNs can emulate complex discrete latent dynamics.