Research

Learning and symbolic reasoning

Learning model Knowledge-guided learning

Concepts & optimization guidance Logical feedback

Symbolic reasoning Learning-guided reasoning

Learning → Reasoning

Learned predictions can guide optimization. LEAP: learning-augmented approximation for integer programs, including MAX-k-SAT. paper

Reasoning → Learning

Knowledge as supervision: when concepts are learnable paper and how to learn from multiple valid explanations. paper

AI Infra

Kernels

Automatic generation and optimization of compute kernels.
Systems

Millrace: pipelined feature preparation and draft-model training. paper
Algorithms

LongSpark: fixed-cost parallel drafting for long-context speculative decoding. paper