Jun Zhou

CausalMix reveals that dynamic data mixture optimization can significantly enhance LLM performance, adapting seamlessly to changing data distributions without the need for costly retraining.

PhysEditWorld reveals that explicit control over physical parameters can transform how game world models interact with their environments, leading to more realistic and manipulable simulations.

MiniOpt achieves the highest solving accuracy for compact models while requiring significantly fewer training resources than traditional methods.

An optimal knowledge distribution can significantly enhance LLM knowledge boundaries, outperforming traditional synthesis methods across multiple benchmarks.

Stellar slashes memory overhead and query latency by orders of magnitude, transforming multimodal document retrieval efficiency.

Shrinkage Bias in E2M1 formats could be the hidden culprit behind training instability in LLMs, but uniform grids like E1M2/INT4 offer a robust solution.

GraphPO slashes redundancy in reasoning model training, enabling more efficient exploration and improved performance on complex tasks.

QK normalization can be effectively integrated into MLA without the overhead of full key caching, leading to improved performance and efficiency.

Structural inconsistency signals can effectively thwart black-box jailbreaks, achieving a dramatic drop in attack success rates without compromising model utility.

Ling-2.6 and Ring-2.6 achieve unprecedented efficiency in agentic intelligence, enabling instant responses and deep reasoning at trillion-parameter scale.

Multi-format training can drastically enhance language model consistency across different answer formats, with just 30% of training data needing augmentation to achieve significant gains.

SearchSwarm reveals that effective delegation in LLMs can significantly boost performance on long-horizon tasks, achieving state-of-the-art results in complex research scenarios.

Code agents are only eliminating 50% of code smells, revealing critical gaps in their understanding of maintainability and cross-file dependencies.

Bridging the perception-reasoning gap in visual planning, MGSD boosts model performance by over 19% while relying solely on visual inference during deployment.

Current vision-language models falter in ultra-resolution reasoning, with errors primarily stemming from evidence grounding and local perception.

Current AI agents falter in autonomous development, revealing critical gaps in robustness and alignment as they struggle against human-engineered solutions.

Forget memorizing manipulation artifacts: comparing against a reference library of authentic examples lets you spot forgeries and adapt to new domains without retraining.

Ditch SwiGLU's quadratic instability: PowLU offers a rational power function that stabilizes LLM pre-training without sacrificing performance.

GPT-4o now has open-source competition: Ming-Omni matches its modality support in a single, unified model capable of perception and generation across image, text, audio, and video.