Zheng Liu

Coding agents routinely fail to build agents that truly meet user requirements, with the gap between demonstration and delivery being the norm.

Watching physical scenes change over time turns out to be an ideal supervisory signal for static visual features, allowing an independently encoded frame supervised across signed temporal intervals to excel at both single-image and multi-frame downstream tasks.

Generative world action models can run up to 3× faster in closed-loop control without fine-tuning or task degradation by dynamically repairing and reusing intermediate computational states instead of recomputing them from scratch.

Atomically thin van der Waals materials can now generate polarization-entangled Bell states with 92% fidelity simply by metasurface-sculpting their intrinsic nonlinear susceptibility tensor.

The DeepSeek-V4.1-Flash model, a multimodal Mixture-of-Experts model with 552B backbone parameters and support for contexts of up to one million tokens, is introduced, substantially improving cost efficiency for agentic workloads and pushing the limits of KV cache compression.

A novel PEFT method termed position anchor tuning (PAT) is proposed, which performs comparably to state-of-the-art methods while incurring significantly lower computational overhead and fewer trainable parameters.

Even top-tier VLMs score just 0.632 when reverse-engineering AI video prompts, frequently hallucinating plausible descriptions that completely fail to reproduce the source video's camera and stylistic controls.

Distilling operational knowledge from GitHub repositories boosts AI research agents' performance by over 134% on key benchmarks.

Achieving 18.5x faster city-scale 3D rendering on mobile VR devices while slashing energy consumption by 92.4% could redefine immersive experiences in urban environments.

AREX achieves superior performance in deep research tasks by recursively refining answers through a novel self-improvement mechanism that outpaces traditional search methods.