Jian Luan

Fine-grained credit assignment in LLM agents can be unreliable due to downstream randomness, but AdaStep's optimal shrinkage coefficient precisely balances local and global rewards.

AdaptRubric's innovative two-stage framework boosts GUI reward modeling performance by over 3.6 F1 points, showcasing the power of task-adaptive criteria.

TRACE transforms high-performing LLMs into consistently reliable agents, achieving a remarkable 34.6-point boost in task consistency.

Achieving superior translation quality without relying on reference data, this approach sets a new benchmark in multilingual machine translation.

HarnessX reveals that evolving agent harnesses through execution feedback can outperform traditional model scaling, achieving up to 44% performance gains in specific tasks.

Finally, a single model generates realistic and coherent audio scenes from text, rivaling specialized models and even approaching real-world recordings.

ControlFoley lets you generate audio from video with unprecedented control over text descriptions and reference audio, even when those inputs conflict.

Open LLMs can now rival proprietary systems in multilingual translation, as demonstrated by MiLMMT-46's competitive performance against Google Translate and Gemini 3 Pro.

Mobile GUI agents, despite benchmark successes, stumble badly when faced with the messy reality of third-party content, failing almost half the time.