A hybrid LLM-agent system recasts cognitive-algorithm discovery as program refinement, consistently improving model fit on a problem-solving task

It proposes a hybrid system that treats cognitive-algorithm discovery as program refinement, taking human-created cognitive models as probabilistic programs and using a system of LLM agents to identify model-behavior mismatches, propose constrained code-level modifications, and verify structural fidelity. Traditional cognitive modeling is interpretable and benefits from human expertise but lacks flexibility and scalability, while emerging LLM-based de novo generation is scalable and flexible but lacks a role for human expertise and has mostly been applied to simpler tasks than algorithm recovery; this work combines human expertise with LLM agents and explicitly targets algorithm recovery. Evidence comes from the system design and evaluation pipeline described in the abstract, i.e., a methods-level argument; implementation details, agent counts, and constraint forms are not given in the text.

Revised models consistently improve model fit relative to ancestral models on human behavior in a problem-solving paradigm that exposes a variety of cognitive algorithms. This moves the refinement pipeline from a design claim to a fit improvement on behavioral data, with the improvement reported as consistent across the evaluation. Evidence is the evaluation conclusion reported in the abstract that revised models consistently improve fit; specific fit metrics, sample sizes, or effect sizes are not provided.

The revision process reveals a small set of recurring innovations that capture meaningful behavioral variability in this task. This indicates refinement converges on a few reusable structural innovations rather than producing many scattered edits, thereby accounting for behavioral differences in the task. Evidence is a qualitative summary in the abstract; the specific innovations and their number are not listed.

Revisions propagate to a probabilistic inference module that performs inference for latent variables and data likelihood computations. It connects LLM-agent code-level modifications with probabilistic modeling's inference and likelihood evaluation, so structural changes can be quantitatively examined within the probabilistic-program framework. Evidence is the pipeline description in the abstract, i.e., an architectural statement.

The result is meant for settings where cognitive-algorithm discovery is framed as program refinement: researchers already have cognitive models expressible as probabilistic programs and are willing to specify constraints on code-level modifications. In that setting, the system can identify model-behavior mismatches, propose constrained modifications, and verify structural fidelity, after which the probabilistic inference module performs latent-variable inference and likelihood computation. Its direct audience is cognitive scientists modeling algorithmic reasoning, and the evaluation setting is one problem-solving paradigm that exposes a variety of cognitive algorithms.

The text is abstract-level information and does not give specific fit metrics, effect sizes, sample sizes, agent configurations, or the concrete form of researcher constraints, nor does it list the small set of recurring innovations. A careful reader would still watch how far these innovations depend on the specific problem-solving paradigm, how the probabilistic-program and constraint setup shapes the direction of revisions, and what the concrete criteria for structural-fidelity verification are. These are directions for follow-up work rather than shortcomings of the present work.