Research Scientist, Meta AI (FAIR)
5 papers at NeurIPS 2025
We study the mechanism of chain of continuous thought on the graph reachability problem, and show it can reason by maintaining a superposition of multiple search traces both theoretically and empirically.
Semi-ring structure exists in 2-layer neural nets for reasoning tasks on Abelian group (e.g., modular addition), trained with L2 loss, which enables constructing global solutions analytically from non-optimal ones instead of gradient descent.
We introduce NaturalReasoning, a 2.8M-question dataset spanning diverse domains, enabling effective knowledge distillation and unsupervised self-training to enhance LLM reasoning capabilities.
A new LLM jailbreak objective that enables more nuanced control over jailbroken responses, exploits undergeneralization of safety alignment, and improves success rates of existing jailbreaks from 14% to 80%.
