Hao Wang

Achieve near-lossless 2-bit LLMs with a novel quantization-aware training scheme that progressively reduces precision and intelligently handles outlier channels.

Quantizing rollouts in LLM RL pipelines introduces a training-inference gap that QaRL closes, leading to +5.5 performance on math problems.

LLMs are far more alike than you think: shared biases and failure modes mean that ensembling them is less effective than you'd hope.

Achieve robust multimodal fusion even with missing modalities by ensuring the fusion head always receives a complete, fixed-size input via learned proxy tokens.

Stop guessing how long LLM outputs will be – modeling the *distribution* of possible lengths slashes latency by 2x and boosts throughput by 40%.

Polarization cues, often overlooked, can significantly boost camouflaged object detection by explicitly guiding RGB feature learning, leading to state-of-the-art performance.

Current facial expression editing models can't simultaneously preserve identity and accurately manipulate expressions, revealing a critical need for better fine-grained instruction following.

Current robot manipulation benchmarks fail to capture the messy reality of real-world deployment, so this work introduces a new benchmark, ManipArena, to close the sim2real gap.

Agentic coding models can achieve near-SOTA performance by specializing in distinct coding domains before unifying them via on-policy distillation.

Robot swarms can now synchronize and execute diverse tasks in uncertain environments without deadlocks, thanks to a new distributed planning method that dynamically adapts to risk and task dependencies.