Research map / Large language models / Agents, code and math
Discovering Symbolic Cognitive Models from Human and Animal Behavior
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adapts FunSearch with an inner parameter-fitting loop (CogFunSearch) using Gemini 1.5 Flash to evolve Python cognitive models that outperform published baselines on human, rat, and fruit fly bandit-task choice data.
Problem
Symbolic cognitive models of human and animal reward-guided decision-making are hard to discover by hand, and the space of candidate models is vast; prior human-designed models underfit behavior.
Approach
Use LLM-driven evolutionary program search (FunSearch) where the outer loop evolves Python programs for choice prediction and the inner loop fits per-subject parameters, scored by cross-validated normalized likelihood on held-out sessions/subjects.
Key results
- Best discovered programs beat the human-designed baseline on all three datasets (p < 0.002, Wilcoxon signed-rank); 95.6% of human, 100% of rat, and 83.8% of fly test subjects are better fit.
- Average normalized-likelihood improvements of 5.17 (Human), 0.54 (Rat), and 0.48 (Fruit Fly) percentage points over baselines.
- Best programs roughly match a GRU RNN fit, with the RNN narrowly outperforming them, closing most of the gap.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the best CogFunSearch-discovered program instead of PFQ | Human dataset · Human dataset held-out test subjects | normalized likelihood: 5.17 ± 0.15 | Helped |
| Use the best CogFunSearch-discovered program instead of RHG | Rat dataset · Rat dataset held-out test subjects | normalized likelihood: 0.54 ± 0.06 | Helped |
| Use the best CogFunSearch-discovered program instead of DFQ | Fruit Fly dataset · Fruit Fly dataset held-out test subjects | normalized likelihood: 0.48 ± 0.05 | Helped |
| Use an RNN instead of the best CogFunSearch program | Human, rat, and fruit fly datasets · Human, rat, and fruit fly datasets | normalized likelihood | Helped |
| Use a LowInfo seed prompt rather than an informative seed prompt (ablation) | CogFunSearch · Human, rat, and fruit fly datasets | normalized likelihood | Hurt |
| Provide more informative information in the prompt and seed program (ablation) | CogFunSearch · Human, rat, and fruit fly datasets | normalized likelihood | Helped |
Surprising findings
- Three independent human-dataset runs discovered a similar novel motif: decaying Q-values toward their mean each step.
- Discovered programs sometimes include unused state variables and terms multiplied by parameters fit to zero, indicating imperfect code quality.
Limitations
- Requires hundreds of thousands of LLM calls per run, making it computationally costly.
- Highest-scoring programs are more complex than baselines; readability and code quality remain imperfect.
- Discovered programs still do not fully match RNN fit.
Experimental setup
Datasets: Human four-alternative bandit dataset (Eckstein et al. 2024) Rat two-armed bandit dataset (Miller et al. 2021) Fruit fly Y-maze bandit dataset (Mohanta 2022; Rajagopalan et al. 2023)
Benchmarks: Human bandit dataset Rat bandit dataset Fruit fly bandit dataset
Metrics: average normalized likelihood on held-out subjects Halstead difficulty LLM-rated complexity trial-history logistic regression fit
Baselines: PFQ (Eckstein et al. 2024) RHG (Miller et al. 2021) DFQ (Ito & Doya 2009) GRU RNN
Builds on: FunSearch Perseverative Forgetting Q-Learning Reward-Seeking/Habit/Gambler-Fallacy model Differential Forgetting Q-Learning
Closest papers by meaning
- Generating Computational Cognitive models using Large Language Models (NeurIPS 2025 · similarity 0.90)
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.89)
- Language Models Trained to do Arithmetic Predict Human Risky and Intertemporal Choice (ICLR 2025 · similarity 0.88)
- Position: Principles of Animal Cognition to Improve LLM Evaluations (ICML 2025 · similarity 0.87)
- Using Artificial Populations to Study Psychological Phenomena in Neural Models (AAAI 2024 · similarity 0.87)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.