Can LLMs be used to discover interpretable models of human and animal behavior?🤔 Turns out: yes! Thrilled to share our latest preprint where we used FunSearch to automatically discover symbolic cognitive models of behavior. 1/12

The fields of neuroscience & psychology use cognitive models to build understanding of the cognitive processes underlying human/animal behavior. These models have historically been built "by hand", often taking multiple years to refine. We took a data-first approach... 2/12

We used 3 existing datasets of humans, rats, and fruit flies performing bandit tasks in a sequence of trials. these datasets allow us to verify the accuracy of our models on both aggregate, and per-subject levels. each dataset has an existing state-of-the-art baseline model. 3/12

FunSearch is an optimization process that uses LLMs to evolve Python programs. we adapted it to our use-case to evolve parameterized symbolic cognitive models using a bi-level optimization: each program's params are fit on the training data for evaluation. 4/12

After running this process, we evaluated the best programs discovered and found they are more accurate than all existing baselines (left plot). They are also on par with a fully-trained RNN, which serve as a soft upper-bound, sacrificing interpretability for accuracy. 5/12

A really cool outcome of this is that we can use our discovered programs to generate synthetic data which reproduces the scientifically important aspects of the real underlying data. 7/12

We also analyzed the interpretability-accuracy tradeoff of our generated programs, and found a form of "efficient frontier". Although we selected our final programs mostly on accuracy, researchers can examine a diversity of programs along this frontier. 8/12

Although optimizing for accuracy, our discovered models are interpretable! We devote a full section to this analysis, but I include some samples here which demonstrate some of our findings. 9/12

Lots more details, analyses, and results in the paper, which I encourage you to check out! www.biorxiv.org/content/10.1... 10/12

Last August I decided to jump on this new project, totally different from what I normally do. It's been a blast working with such fantastic collaborators, and I've felt like a PhD student again, having stepped into a new field! Really looking forward to the next few months! 11/12

Shout out to some of my (taggeable) collaborators @nenadtomasev.bsky.social @nathanieldaw.bsky.social @noemielteto.bsky.social , and in particular kevin and @neurokim.bsky.social from whom i've learnt so much! 💙 12/12