For a long time, reasoning was regarded as a distinctly human ability, far beyond the reach of machine learning. Today, frontier AI systems routinely work through proofs, debug programs, and plan multi-step solutions to novel problems. This is functionally a form of reasoning, even if it is likely achieved very differently from how humans reason.
In contrast to retrieval or memorization, reasoning involves multi-step computation that composes what the system already knows in new ways, producing results that could not be obtained from recall alone. The steps are typically sequential, and finding the right chain often involves search or exploration among alternative strategies. The mark of genuine reasoning is systematicity: a process that applies to every problem sharing the same structure, including longer or harder ones than any seen in training, rather than memorized patterns that happen to cover the training distribution.
Modern LLMs are transformers pre-trained on internet-scale human text, then trained on high-quality reasoning traces and refined with reinforcement learning. Reasoning emerges somewhere in this pipeline, but we have a limited understanding of how, or how far it extends. I’m interested in understanding how these models learn to reason from data and experience, how they represent and organize the reasoning process internally, and how they execute long chains of computation, and in developing methods that improve their out-of-distribution generalization.
Papers in this thread:
- CoT Information: a learning theory showing how supervising intermediate steps, rather than only final answers, makes reasoning easier to learn, and formally quantifying the gain in sample efficiency.
- Putting transformer theories to the test: compares and empirically evaluates competing theoretical accounts of which computations transformers can and cannot perform in a single forward pass.
- Recursive latent space reasoning: explores recurrence in latent space as an alternative to autoregressive reasoning in token space, combining input-adaptive recurrence, supervision of latent steps, and discrete latent states, enabling transformers to extrapolate to problems far larger than those seen in training.
- Uncovering latent reasoning strategies in language models: LLMs implicitly represent a broad repertoire of abstract strategies that generalize across problems; this paper studies how to recover this latent structure as an explicit variable that can be interpreted and controlled.
More broadly, I’m interested in studying and designing better algorithms for learning from experience, including for long-horizon problems and continual learning. Current reinforcement learning methods for LLMs make little use of the rich structure these models learn, exploring only passively by sampling from their existing distribution. Making this structure explicit and controllable could enable a model to actively direct its exploration through abstract strategy space.