Synthesizing symbolic models of human and machine reasoning at scale

Solim LeGris, Brenden M. Lake, Todd M. Gureckis and colleagues at NYU and collaborating institutions introduce Behavior-Conditioned Symbolic Synthesis (BCSS), which uses an LLM to search for interpretable programs that exactly reproduce individual human or model responses, and use it to compare how people and reasoning models make errors.
Ask this paper
Method. An LLM proposes program edits, and the search keeps deterministic symbolic programs that verifiably reproduce each observed response across hundreds of reasoning problems.
Human errors. Programs that explain errors are shorter than the target rule, often fit only some training examples, and most often compute a parameter from the wrong feature.
Shared signatures. Large reasoning models show the same broad error patterns as people.
Reasoning effort. In GPT-OSS-120B, lower reasoning effort produces errors explained by simpler programs and fewer partial fits to the evidence, paralleling the human results.
Interpretation. The authors read both as satisficing: settling on a rule that fits enough of the evidence rather than all of it.
Abstract
Symbolic models offer a powerful account of cognition, but their scope has historically been limited by the need for hand-engineering and computational tractability. We show how advances in artificial intelligence (AI) can scale symbolic modeling and help overcome these engineering limitations. We introduce Behavior-Conditioned Symbolic Synthesis (BCSS)---an approach for automated search of interpretable, deterministic symbolic programs that verifiably reproduce behavior---and apply it to model both human and machine behavior across hundreds of challenging reasoning problems. Analyzing these programs reveals that both people and machines tend to oversimplify, adopt rules that are partial explanations for their observations, and misinfer dependencies between problem features. These shared signatures are modulated by available resources, with reasoning effort predicting the extent to which machines simplify and explain their observations. Together, these results point to underlying mechanisms reminiscent of satisficing in both people and machines, and illustrate how BCSS offers a general approach for interpreting and comparing rule-based reasoning of humans and black-box AI systems.