Zheyuan Zhang

GaussianFusion achieves superior 3D perception by leveraging a continuous 3D Gaussian space, outperforming traditional methods while using fewer resources.

Verbal uncertainty in large language models can be effectively aligned with their sampled behavior, leading to improved calibration and reduced overconfidence.

Agents struggle to match human performance in long-horizon tasks, revealing critical gaps in their learning capabilities during deployment.

Stop hand-tuning your multi-agent collaboration: EvolveRouter co-evolves agent prompts and routing strategies, dynamically adapting collaboration size to each query for SOTA question answering.

LLMs are introducing a citation crisis in scientific literature, and CiteAudit offers a scalable solution to detect fabricated references with significantly improved accuracy and interpretability.

Current AI alignment strategies that compress human values into a single reward are doomed to flatten values, erase minority viewpoints, and ignore uncertainty, demanding a shift towards "Edge Alignment" that respects value diversity.

Duality between nodes and hyperedges unlocks unsupervised learning on heterophilic hypergraphs, outperforming supervised methods without needing negative samples.

A new benchmark, OPBench, offers researchers a standardized way to evaluate and improve graph learning models tackling the opioid crisis across diverse real-world scenarios.