henhaobio
Sign in

bioRxivAnimal Behavior & CognitionPreprint

Discovering Symbolic Cognitive Models from Human and Animal Behavior

P. CastroiD, Nenad TomaševiD, A. Anand, N. Sharma, Rishika MohantaiD, A. DeviD, K. PerliniD, S. Jain, Kyle Levin, Noémi ÉltetőiD, Will DabneyiD, A. NovikoviD, Glenn C TurneriD, M. EcksteiniD, Nathaniel D. DawiD, Kevin J. MilleriD, Kimberly StachenfeldiD

PDFDOI

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (6 Feb 2025) · CC BY 4.0

LLM-driven evolutionary program synthesis (CogFunSearch) discovers interpretable Python cognitive models that predict bandit-task choices of humans, rats and fruit flies better than leading hand-designed models.

Researchers usually write symbolic cognitive models by hand, so much of the huge space of possible models goes unexamined, and widely used models underfit compared with flexible recurrent networks. CogFunSearch automates this search: an LLM writes and evolves programs whose parameters are then fit to behavioral data, producing readable hypotheses whose fit comes close to RNN-level performance.

Key results

  1. ShownFor each dataset, the single best evolved program fit held-out test subjects better than the strongest published model for that species (PFQ for humans, RHG for rats, DFQ for fruit flies).

    Wilcoxon signed-rank W = 1772 (human), 0 (rat), 1479 (fly), all p < 0.002; subjects better fit: 95.6% human, 100% rat, 83.8% fly; mean normalized-likelihood gain 5.17 ± 0.15 (human), 0.54 ± 0.06 (rat), 0.48 ± 0.05 (fly) percentage points. · 3.1 Best Discovered Programs Fit Data Well; Figure legends

  2. ShownThe search found well-fitting models reliably across replicate runs and seed programs. The three informative seeds performed similarly, while the minimal LowInfo seed tended to produce weaker programs.

    The top program beat its baseline on held-out data in 35 of 36 runs (three datasets, four seed programs, three replicates, each sampling at least 150,000 programs). LowInfo vs informative seeds: Wilcoxon rank-sum p from 0.05 to 0.65 across nine comparisons. · 3. Results; 3.1 Best Discovered Programs Fit Data Well; D.2 Differences across Seed Program

  3. ShownThe best programs come close to the fit of GRU recurrent networks. They beat a cross-subject GRU on human data and trail the GRUs only narrowly on rat and fly data. For rats and flies, synthetic behavior generated by the programs reproduces real trial-history patterns.

    GRUs were tuned by a hidden-unit sweep with early stopping. On rat data the program bridged over 50% of the gap to the RNN. For the rat and fly datasets, lagged logistic regression of reward-seeking and choice perseveration matched between the real data and ten synthetic datasets per program. · 3.1 Best Discovered Programs Fit Data Well; C. Extended Dataset descriptions; Figure legends

  4. ShownWithin each run's program library, fit trades off against code complexity. Programs on the efficient frontier include some that are both simpler and better-fitting than the hand-designed baselines.

    Halstead difficulty and an LLM pairwise complexity ranking (sorted with MergeSort) both rise with normalized likelihood. Frontier programs beat the baselines on both axes for most seeds and datasets. One rat frontier program uses 5 parameters versus seven in RHG. · 3.2 Quality-of-Fit Trades Off With Complexity; D.3 Complexity Statistics; D.3.1 LLM-based pair-wise complexity scores; Example Program 1

  5. SuggestedTop programs contain interpretable mechanisms absent from the baselines. The human model mostly tracks reward-independent choice history and decays values toward their mean; the rat model pulls chosen values toward the unchosen arm; fly forgetting applies only after positive prediction errors.

    The decay-toward-mean motif arose independently in three Structured2 human runs, each of which also added a lapse ceiling on choice probabilities. The best rat program kept the RHG state variables. The best fly program mixed biased softmax with ϵ-greedy exploration. · Top Scoring Human Program; Top Scoring Rat Program; Top Scoring Fruit Fly Program

Methods, data and reproducibility

CogFunSearch extends FunSearch. In the outer loop, an LLM mutates pairs of high-scoring JAX Python programs within an evolutionary search. In the inner loop, up to ten parameters are fit by gradient descent and each program is scored by normalized choice likelihood under even/odd-session cross-validation; final validation uses held-out subjects.

  • Online four-armed bandit with keypress choices and drifting graded rewards (0 to 100 points); up to five sessions per participant (Eckstein et al. 2024) · human · n = 862 participants; 4,134 sessions; 617,871 trials
  • Two-armed bandit with binary liquid rewards at nose ports, drifting reward probabilities and 20% instructed trials (Miller et al. 2021) · rat · n = 20 rats; 1,946 sessions; 1,087,140 trials
  • Y-maze odor-choice two-armed bandit rewarded by red light activating sugar receptors; blockwise reward probabilities; one session per fly · fruit fly · n = 347 flies; 68,000 trials
Methods & tools
FunSearch (LLM-driven evolutionary program search), CogFunSearch (bilevel extension of FunSearch), JAX, AdaBelief optimizer, GRU recurrent neural network comparator, PFQ (Perseverative Forgetting Q-Learning) baseline, RHG (Reward-Seeking/Habit/Gambler-Fallacy) baseline, DFQ (Differential Forgetting Q-Learning) baseline, Wilcoxon signed-rank and rank-sum tests, Halstead difficulty measure, LLM pairwise complexity ranking with MergeSort, Trial-history logistic regression
Reagents & environment
Programs are written in Python/JAX (differentiable and jittable). Parameters are fit with AdaBelief (learning rate 5 × 10^−2, up to 10,000 steps, convergence threshold 10^−2, up to 10 restarts). No package or LLM versions are stated.
Experimental conditions
Each CogFunSearch experiment ran three seeds for seven days each. The best programs needed 1.2M (human), 300K (rat) and 200K (fly) LLM samples. Hardware is not reported.
Reproducibility gaps
The text gives no code repository, data download links or dataset accessions, and does not name the LLM used for program mutation and complexity ranking.
Limitations (6)
  • The best programs still fall slightly short of GRU recurrent networks on rat and fly data. The authors leave closing this gap, and further improving interpretability, to future work. · stated by the authors
  • The search is costly: the best programs required 1.2M (human), 300K (rat) and 200K (fly) LLM samples. Rejection sampling lowers the cost but gives up exploration and may converge to local optima. · stated by the authors
  • The code is not fully readable. Some names are vague or misleading; indexing past the end of the parameter array silently returns the last element in JAX; the best rat program updates only 7 of its 18 hidden-state dimensions; and some terms carry parameters that fit to zero. · stated by the authors
  • Many programs reach similar predictive scores, so choice likelihood alone cannot pick out a single best model. Whether these programs are reparameterizations of the same input-output mapping is unresolved. · stated by the authors
  • The human 'sessions' are actually consecutive blocks, with probable dependencies that are not modeled. The human RNN comparison pits a pooled cross-subject GRU against per-subject program fits, which the authors acknowledge is not like-for-like. · stated by the authors
  • The paper never names the LLM used for program mutation and complexity ranking, and only reward-guided bandit tasks are tested. Both the reproducibility of the search and its generality to other behaviors remain unverified. · noted in this summary

Read this if you build computational models of reward learning or choice behavior and want to see how LLM-guided program search can propose interpretable models that outperform hand-crafted baselines.

Citations
14

Abstract

Symbolic models play a key role in cognitive science, expressing computationally precise hypotheses about how the brain implements a cognitive process. Identifying an appropriate model typically requires a great deal of effort and ingenuity on the part of a human scientist. Here, we adapt FunSearch Romera-Paredes et al. (2024), a recently developed tool that uses Large Language Models (LLMs) in an evolutionary algorithm, to automatically discover symbolic cognitive models that accurately capture human and animal behavior. We consider datasets from three species performing a classic reward-learning task that has been the focus of substantial modeling effort, and find that the discovered programs outperform state-of-the-art cognitive models for each. The discovered programs can readily be interpreted as hypotheses about human and animal cognition, instantiating interpretable symbolic learning and decision-making algorithms. Broadly, these results demonstrate the viability of using LLM-powered program synthesis to propose novel scientific hypotheses regarding mechanisms of human and animal cognition.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISYe '26Kasenberg '26Miller '23Miller '21Luppi '25Riveland '24
72% semantic matchNeuroscience
Yingying Ye, Youle Fang, Xiaoxue Gao+5 more·bioRxiv·Oct 6, 2026