Tianyu Zhang

Quality-aware pretraining strategies, particularly approach-aware episode conditioning, can significantly outperform plain behavioral cloning in VLA tasks, with optimal choices depending on label granularity and model architecture.

A single MLLM can now perform temporal grounding and video QA with significant inference speedups by treating them as structured decision tasks.

Interactive lesion segmentation can now adaptively refine predictions in real-time, significantly improving accuracy in heterogeneous PET/CT datasets.

PRAXIS reveals that systematically extracting and surfacing tacit knowledge can dramatically enhance LLM performance in domain-specific code generation, outperforming traditional methods.

ClinRAG-GRAPH achieves impressive pCR prediction accuracy while ensuring interpretability and robustness against imaging biases across multiple centers.

Targeted information exposure in LLaMA-3-8B drops from 144 instances to just 1, all while preserving contextual fidelity.

Synthesized breast MRIs can now better mimic real-world lesion complexity thanks to a diffusion model that explicitly handles multi-scale features and heterogeneous enhancement.

LLMs struggle with structured 2D tasks when inputs are serialized into 1D, revealing a surprising performance gap compared to vision-augmented models that directly process the 2D layout.