Shuicheng Yan

Making the teacher's privileged context learnable end-to-end enables agents to evolve more efficiently, outperforming traditional methods with less than 30% of their rollout budget.

StreamFlow achieves a remarkable 67.73% accuracy in streaming video understanding while cutting latency and memory usage by over 50%.

ToolArtist achieves unprecedented synergy in image generation by unifying reasoning and tool use under a single agent policy, outperforming conventional methods.

Shifting the focus from static state transitions to dynamic, agent-centric feedback could revolutionize how we train and evolve intelligent agents.

DOPD reveals that intelligently routing supervision based on advantage gaps can significantly enhance capability transfer in distillation, outperforming conventional methods.

MambaADv2 achieves superior anomaly detection by combining linear computational efficiency with advanced global and local representation modeling, setting a new standard in unsupervised learning.

SPOT-E transforms frozen VLMs into more reliable evidence readers by dynamically spotlighting critical visual information during inference.

Current memory agents fail to provide reliable governance in shared settings, with no method achieving a balance between utility, access control, and forgetting.

OPD-Evolver outperforms traditional memory systems by up to 11.5%, showcasing a new paradigm in agent evolution that transcends mere memory storage.

TacForeSight enables robots to anticipate contact changes in real-time, outperforming traditional methods in dynamic manipulation tasks.

Real-time audio interaction is now possible with a unified model that not only performs traditional tasks but also proactively responds to audio stimuli.

Scaling up robot data and closing the loop with state decoding and automated reward scoring allows a 2B parameter video world simulator to outperform larger, dedicated robotic world models in real-world policy transfer.

The landscape of deep learning optimizers is vast, but this paper cuts through the noise to reveal the fundamental trade-offs and promising future directions for efficient, robust, and trustworthy training.

A dedicated guard agent, trained via reasoning-intensive methods, can effectively neutralize prompt injection attacks in web-navigating agents without sacrificing performance.

Text-centric agentic search is out: Deep-Reporter shows how to build multimodal agents that leverage both text and visuals for grounded long-form generation.

Real-world proactive agents can now infer latent user needs and act on them in real-time, rivaling state-of-the-art models in intent detection while maintaining low latency.

Language models are increasingly doing their real work in the "invisible" latent space, not the tokens we see.

Current research agent benchmarks miss critical flaws, as MiroEval reveals that process quality is a reliable predictor of research outcome, and multimodal tasks expose weaknesses invisible to output-level metrics.

Autoregressive inference gets a potential 14x speed boost without retraining, thanks to a clever trick of reusing attention weights within semantically coherent chunks.