I am a computational neuroscientist at NIH working at the intersection of neuroscience, neural networks, and AI interpretability. I have developed recurrent neural network and reinforcement learning models to investigate the neural mechanisms underlying learning and decision-making. More recently, I have extended this computational perspective to the mechanistic interpretability of language models, using representational analyses, causal interventions, and circuit-level approaches to uncover the internal algorithms that neural networks develop to perform reasoning.
01 / Latent reasoning
01 / CoT (chain of thoughts) / latent reasoning / interpretability
Five model variants solve the same graph-reachability task, but strong in-distribution performance masks different strategies. Representational and circuit analyses trace a soft forward-search computation in the latent models.
Not All Thinking Is Created Equal: Latent Reasoning Discovers a Recurrent Search Algorithm for Depth Generalization
01 Behavior
All five variants perform well on in-distribution (ID) problems, but their success reflects different computational strategies. Direct and Pause-token rely substantially on answer-candidate in-degree, whereas CoT exploits degree differences among the query root's immediate successors, favoring those with lower in-degree minus out-degree.
02 Representation
Representational similarity analysis (RSA) reveals that Full-latent and Bottleneck-latent develop internal states aligned with successive forward-search frontiers. As recurrent computation progresses, their representations correspond to nodes reachable at increasing depths from the query root.
03 Circuit
In the Bottleneck-latent model, attention head L4H1 retrieves graph relations through an address-content organization: keys select premise sources, while values provide destination information. L4MLP works jointly with the residual stream to filter and integrate retrieved information into the next latent state, guiding subsequent attention queries. In parallel, heads L4H1, L4H2, and L4H4 match the evolving state to candidates' graph context and write answer-relevant evidence into latent representations for final decoding. Together, these interacting pathways support a 'soft' forward-search algorithm that propagates reachability across recurrent steps.
I developed a GRU-based recurrent neural network that integrates past sensory, action, and reward information to predict future events. Through this predictive objective, the model learns task structure and reproduces flexible, adaptive behavior across multiple behavioral tasks. Notably, network units spontaneously develop activity patterns resembling decision-related neural responses, despite never being trained to match neural recordings.
Sequence prediction induces adaptive and flexible behavior and interpretable internal task variables.
Read paperA recurrent neural network framework for flexible and adaptive decision making based on sequence learning
I developed a reservoir-network model of the orbitofrontal cortex (OFC) in which sensory and reward-history inputs generate task-state representations that go beyond sensory cues. Using a biologically plausible learning rule, the model showed that these extended state representations can support model-based-like behavior with a relatively simple training algorithm, without explicitly encoding the full task structure.
Reward history acts as an additional cue that expands the represented state space.
Read paperA neural network model for the orbitofrontal cortex and task space acquisition during reinforcement learning
02 / Publications
Cheng & Zhang
Zhang, Costa, Langdon & Schoenbaum
Costa* & Zhang* et al. · co-first authors
Zhang* & Takahashi* et al. · co-first authors
Zhang, Cheng & Yang
Zhang* & Cheng* et al. · co-first authors