Xianglong Liu

Distilling video diffusion models just got a whole lot faster and better at capturing motion, thanks to a new method that directly optimizes score gradients.

Achieve the same performance with half the data: MIRA distills source-specific rubrics into scalable data scorers, enabling efficient and effective data selection for LLM mid-training.

AgentDoG 1.5 proves you can achieve GPT-5.4-level agent safety with open-source models trained on just 1k samples, slashing deployment overhead by two orders of magnitude.

LLMs that can generate HTML are finally useful: HTMLCure's closed-loop repair engine turns superficially correct but broken pages into high-quality training data, rivaling the performance of much larger models.

DisagFusion unlocks up to 20x higher throughput for diffusion model serving by intelligently splitting the workload across heterogeneous GPUs and dynamically adapting to workload shifts.

Pushing super-resolution models to the extreme of 2-bit quantization doesn't have to mean sacrificing accuracy, thanks to QuantSR+'s clever combination of quantization-aware operators, architecture, and training.

LLM agents can achieve near-impregnable defense against prompt injection with minimal utility loss by borrowing classic operating system virtualization techniques.

Text-to-image models can be tricked into generating images containing malicious text with over 90% success, even when standard jailbreak methods fail.

Binarizing weights and ternarizing activations in Transformers can deliver 16-24x kernel speedup and comparable accuracy to full-precision models, finally making ultra-low-bit quantization practical.

Industrial code generation gets a reasoning boost: InCoder-32B-Thinking leverages error-driven feedback and a code world model to achieve top-tier performance on complex hardware-aware tasks.

A new 32B code LLM trained specifically for industrial tasks crushes existing models on specialized domains like chip design and GPU kernel optimization, while remaining competitive on general coding benchmarks.

Code LLMs can achieve SOTA performance in agentic tasks by explicitly modeling the dynamic evolution of software logic across different training stages.