This direction isn’t fully settled yet, but the starting point is applying LLMs to ARC-style abstract reasoning puzzles. Rather than chasing benchmark-specific tricks, the goal is to study what makes a reasoning process transfer across tasks the model hasn’t seen - i.e., generalizable reasoning rather than memorized pattern-matching on a single benchmark.
Generalizable Reasoning with LLMs
Using large language models to solve abstract reasoning tasks like the ARC (Abstraction and Reasoning Corpus) benchmark - and, more broadly, understanding what makes a reasoning strategy generalize rather than overfit to one benchmark.