Reasoning and agentic AI
Since the early days of LLMs, we have been working to improve the reasoning capabilities of the models. Our focus includes developing analytical tools to measure and evaluate these capabilities, as well as new methodologies to augment reasoning. In particular, we are actively working on dynamic reasoning, tool-calling and agentic systems.
Selected publications:
- AbstRaL: Augmenting LLMs’ Reasoning by Reinforcing Abstract Thinking
- Algorithm discovery with llms: Evolutionary search meets reinforcement learning
- How far can transformers reason? the globality barrier and inductive scratchpad
- Generalization on the unseen, logic reasoning and degree curriculum
- Polynomial‐time universality and limitations of deep learning
- Sgd learning on neural networks: leap complexity and saddle-to-saddle dynamics
- The merged-staircase property: a necessary and nearly sufficient condition for sgd learning of sparse functions on two-layer neural networks
- Learning to reason with neural networks: Generalization, unseen data and boolean measures
- Polynomial Freiman-Ruzsa, Reed-Muller codes and Shannon capacity
- Reed-Muller codes have vanishing bit-error probability below capacity: a simple tighter proof via camellia boosting
- A proof that Reed-Muller codes achieve Shannon capacity on symmetric channels
- Proof of the achievability conjectures for the general stochastic block model
- Community detection and stochastic block models: recent developments
- Community detection in general stochastic block models: Fundamental limits and efficient algorithms for recovery
- Exact recovery in the stochastic block model
