Previous work has shown that programmatic policies—computer programs written in a domain-specific language—generalize to out-of-distribution problems more easily than neural policies.

Is this really the case? 🧵

In other cases, where clustering seems unable to find relevant structure, the subgoal-based policies do not seem to harm the search, as in Sokoban problems.

The empirical results are strong when clustering effectively detects the problem's underlying structure.

🧵1/ New paper! 📄 Subgoal-Guided Policy Heuristic Search with Learned Subgoals, led by my PhD student @tuero.ca.

arxiv.org/pdf/2506.07255

This paper follows the Levin tree search (LTS) research line and focuses on learning subgoal-based policies.

🧵1/ New paper! 📄 InnateCoder: Learning Programmatic Options with Foundation Models

This is Rubens Moraes' final chapter of his PhD thesis from Universidade Federal de Viçosa, Brazil, in collaboration with Quazi Sadmine and Hendrik Baier.

arXiv: arxiv.org/abs/2505.12508

Are you interested in programmatic representations in machine learning? Please consider submitting your work to these workshops: one at ICML and another at RLC. 🧵

I am re-reading "Learning Differentiable Programs with Admissible Neural Heuristics." This is still one of my favorite papers from the past 5–6 years.

arxiv.org/pdf/2007.12101

I still believe that visual perception could play a key role in solving the ARC-AGI challenge. The fact that VLMs currently perform worse than LLMs doesn’t necessarily prove the challenge isn’t visual. It could be that VLMs lack the bias needed to solve the visual ARC riddles.

x.com/fchollet/sta...

Just arrived in Vancouver for our #NeurIPS2024 tutorial on programmatic RL and code generation. Come join us tomorrow.