Zhi Wang

Value Flattening is identified as an important yet overlooked failure mode of critic learning in standard PPO and a simple sparse supervision strategy can mitigate it; SParse Proximal Policy Optimization is introduced, which applies the value loss to only a few well-separated states in each response to mitigate both effects.

TRACE transforms high-performing LLMs into consistently reliable agents, achieving a remarkable 34.6-point boost in task consistency.

MemFuse achieves superior performance in multi-source memory tasks, revealing that effective memory fusion can significantly enhance an agent's reasoning capabilities.

Effective visual tool use hinges on causally grounded supervision, not just imitating tool calls, reshaping how we train multimodal agents.

GeniWorld achieves robust zero-shot generalization in robotic manipulation, outperforming traditional models even with minimal training data.

A unified decision process for multi-modal reasoning reveals that joint optimization of text and image generation can dramatically enhance performance in complex reasoning tasks.

A lightweight Q-value model can boost a 9B VLA's performance beyond that of a 27B model while reducing inference latency by 27%.

By exposing task dependencies through a graph structure, ATG enables LLMs to execute complex tasks more efficiently and reliably than ever before.

A tool-augmented agent achieves high compilation rates but reveals a shocking 29-point gap in semantic faithfulness, challenging the reliability of existing evaluation metrics.

Organizing visual attention before camera motion can dramatically enhance narrative coherence and viewer engagement in dynamic 3D environments.

ForceBand transforms human muscle signals into precise force data, enabling robots to learn manipulation tasks with unprecedented accuracy.

LLMs can effectively decompose interactions into phases and roles, but struggle to generate dynamic, realistic motion without a structured approach.

HDSL achieves a remarkable reduction in editing token usage by over 5 times while maintaining scene integrity and enhancing generation speed.

Turning past programming failures into reusable knowledge boosts automated repair performance by 3.7% on a multimodal benchmark.

By structurally disentangling temporal joint planning from frame-level manipulation, StructBiHOI achieves superior long-horizon stability and motion realism in bimanual hand-object interaction generation.

Shifting the powertrain of a powered prosthetic knee *above* the joint could boost walking speed and cadence, challenging the conventional focus on simply minimizing total mass.

LLMs can be sped up by 21% at inference time without retraining, thanks to a new sparsity method that smartly prunes activations based on the importance of the weights they interact with.

LLMs can be taught to "think longer" and explore more diverse reasoning paths in-context via a simple length-incentivized reward, leading to improved generalization.