Xinyin Ma

Invisible Safety Threat: Malicious Finetuning for LLM via Steganography

dParallel: Learnable Parallel Decoding for dLLMs

SparseD: Sparse Attention for Diffusion Language Models

ConciseHint: Boosting Efficient Reasoning via Continuous Concise Hints during Generation

Dimple: Discrete Diffusion Multimodal Large Language Model with Parallel Decoding

Revisit Model Adaptation from Parameters to Features

Gradient Inversion Attacks Beyond SGD

Privacy-Aware Hybrid Image Synthesis with Local-Cloud Collaboration

SCoT: Self-Correction at Test-time for Image Generation

MixReasoning: Switching Modes to Think

Disentangling Token Dependencies for Efficient Decoding in Diffusion Language Models

Thinkless: LLM Learns When to Think

VeriThinker: Learning to Verify Makes Reasoning Model Efficient

dKV-Cache: The Cache for Diffusion Language Models