Jinman Zhao

Local Branch Routing enables language models to leverage contextual evidence for decision-making without the computational burden of full solution searches, leading to substantial improvements in reasoning accuracy.

Pretraining action modules with motion priors can drastically enhance VLA model performance, achieving faster convergence and better success rates in complex robot manipulation tasks.

Code that passes existing LLM code generation benchmarks may still be subtly wrong: CodeHacker finds the flaws.