Rui Liu

Real-time, open-ended video understanding is now possible: AURA enables VideoLLMs to proactively respond to live video streams, moving beyond simple captioning.

Stop letting sparse rewards bottleneck your VLN agent: SACA disentangles failed trajectories into valid prefixes and divergence points for dense supervision, unlocking SOTA performance.

LLMs still fail to follow complex instructions that entangle content, formatting, control flow, and real-world constraints, despite progress on simpler benchmarks.

Current GUI agents are reactive, but PIRA-Bench offers a challenging new environment for training agents to *proactively* anticipate user intentions from continuous visual inputs, a crucial step towards truly intelligent AI assistants.

LLM reasoning can be compressed without sacrificing accuracy by selectively encouraging exploration only on difficult questions, preventing premature entropy collapse.

A 1.7B parameter model can now rival much larger audio language models, thanks to a novel architecture and data synthesis pipeline.