Xin Li

RynnBrain 1.1 not only outperforms all competitors in embodied cognition tasks but also redefines how robots can be trained for complex manipulation through innovative 3D grounding techniques.

VLA-Corrector allows VLA models to adaptively replan actions in real-time, drastically reducing compounding errors in dynamic environments.

Fine-tuning large language models with ZO-Act yields consistent performance gains while dramatically reducing variance in optimization.

Goku redefines the landscape of video editing datasets by enabling complex, multi-task editing capabilities that surpass traditional single-task limitations.

Achieving trillion-parameter performance with just 35 billion parameters by scaling agent horizons reveals a new frontier in model efficiency.

DnA achieves a notable 0.8% improvement on ImageNet-1K by effectively filtering out irrelevant features in attention mechanisms.

A novel QUBO/Ising approach enables efficient design of cyclic peptides by reducing the complexity of residue representation without sacrificing interaction fidelity.

Compiling external knowledge into structured, reusable skills boosts agent performance, achieving over 98% success rates in complex tasks.

MLEvolve not only breaks the barriers of information isolation in MLE agents but also achieves state-of-the-art performance in algorithm discovery within half the standard runtime.

Current video MLLMs struggle to grasp fleeting visual events, with top models barely surpassing 39% accuracy on critical momentary tasks.

RynnBrain leapfrogs existing embodied foundation models, offering a unified, open-source spatiotemporal model that excels at physically grounded reasoning and planning across a wide range of benchmarks.