Zheng Ruan

Non-thinking inference in hybrid-thinking MLLMs suffers from widespread response-pattern failures, revealing a critical misalignment that can be mitigated with targeted reinforcement learning strategies.

Achieving a 6.37x speedup in inference while expanding OCR capabilities across long-tail tasks sets a new benchmark for lightweight models.

CuRe transforms video captioning reward design by shifting from holistic evaluations to precise claim-level verification, significantly boosting factual accuracy and diversity in generated captions.

Task success rates for agentic phone use soar from 36.67% to 45.33% through a novel combination of real and mock environments in training.

Reliable phone automation hinges on mixed-action capabilities, with agents achieving a 75% success rate in real-world workflows.