Xue Yang

Bench2Dex is meant as a platform for studying visuo-tactile learning across dexterous hands and offers a shared setting for algorithm development while tactile hardware and simulation models are still evolving.

DMRL transforms the way advertising recommendation systems optimize skill documents, achieving superior performance through structured editing and advanced reward estimation techniques.

FIRM-Video reveals that a checklist-driven approach can significantly enhance the reliability of text-to-video reward modeling, achieving state-of-the-art performance in evaluation metrics.

CoQui achieves superior image generation quality with fewer qubits by decoupling pixel control from quantum resource constraints.

WorldScape Policy 2.0 achieves unprecedented long-horizon autonomous planning by integrating reasoning-augmented memory with multimodal instruction processing.

Current LLMs only achieve 27.3% accuracy in reasoning about scientific lineage, revealing a critical gap in their compositional capabilities.

Role-aware training can boost video diffusion models' physical consistency by up to 39.4% without sacrificing visual fidelity.

ACE-Brain-0.5 unifies spatial reasoning and action generation in embodied AI, achieving remarkable performance improvements across multiple benchmarks.

Large-scale structured academic visual data can transform image generation from mere aesthetic appeal to verifiable knowledge-grounded creation.

ACE revolutionizes context management for LLM agents, enabling them to adaptively retain critical information without loss, leading to superior decision-making performance.

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

Current AI agents struggle to reliably rediscover scientific knowledge, with top performers averaging only 21.5 out of a possible score, revealing critical gaps in their research capabilities.

Current omnimodal LLMs that ace offline benchmarks still fumble basic real-time interactions, highlighting a critical gap in their ability to handle streaming audio-visual data.

Model-generated skills can actually hurt agent performance, and bigger models don't necessarily make for better skill extractors or consumers.

SkillOpt transforms agent skill development into a reproducible optimization process, achieving state-of-the-art results by treating skills as trainable parameters.

LLM-powered agents can now produce surprisingly strong photographs in complex 3D environments, suggesting a path towards embodied AI with aesthetic awareness.

Visual degradations can cripple the spatial reasoning abilities of even state-of-the-art MLLMs, but targeted finetuning can restore—and even surpass—human-level performance.

LLMs can have their personalities surgically altered by tweaking just 0.5% of their neurons, preserving general capabilities while achieving competitive control.

Current image editing models stumble when domain-specific knowledge is required, as revealed by a new benchmark spanning disciplines from natural science to social science.

Hallucinations in RL-based image editing and generation are tamed with FIRM, a new framework that trains robust reward models on curated datasets to provide more accurate guidance.