Public implementations associated with my work in reinforcement learning, efficient AI systems, neural architecture design, and computer vision.
A synthetic-data and curriculum-reinforcement-learning framework for generating correct, high-performance Triton kernels from PyTorch programs.
Decoupled Reward Policy Optimization improves the performance–efficiency trade-off of reasoning models by separating correctness and length signals.
Discriminative Constrained Optimization reinforces large reasoning models while controlling their output distributions.
GitHub repository · NeurIPS paper · arXiv
A framework for jointly searching efficient LLM architectures under accuracy and deployment constraints.
GitHub repository · OpenReview · arXiv
A collection of training-free architecture-search methods and efficient vision applications, including DeepMAD, MAE-DET, and TinyNAS.
GitHub repository · DeepMAD paper · MAE-DET paper · Demo
A zero-shot neural architecture search method for high-performance image recognition without training every candidate network.
GitHub repository · Paper · arXiv
An object-detection architecture built around a heavy-neck design that shifts computation toward multi-scale feature fusion.
A k-nearest-neighbor attention mechanism for improving vision transformers.
A transformer-based entropy model for learned image compression.
Adaptive neighbor discovery for face clustering in a learned structure space.
Neural architecture design methods that explicitly target efficient GPU execution.
A structured model-pruning method that combines knapsack optimization with inner knowledge distillation.
Code for a non-convex, one-pass framework for generalized factorization machines and rank-one matrix sensing.
GitHub repository · Paper · arXiv
An implementation accompanying the second-order linear model for efficient learning with feature interactions.