news

Oct 7, 2025	Our paper The Mirage of Performance Gains: Why Contrastive Decoding Fails to Mitigate Object Hallucinations in MLLMs? has been accepted to NeurIPS 2025! This work reveals that contrastive decoding fails to genuinely mitigate hallucinations. Any apparent improvements are merely artifacts of confounding factors, not true effectiveness.
Sep 15, 2025	Excited to share that I’ve begun a research internship at Tencent! My project will involve exploring how to boost the common sense reasoning abilities of MLLMs. This will primarily be achieved through post-training approaches, focusing on efficient data construction and innovative training strategy design.
Feb 28, 2025	Our paper Lifting the Veil on Visual Information Flow in MLLMs: Unlocking Pathways to Faster Inference has been accepted to CVPR 2025! This work investigates the internal visual information flow patterns in MLLMs and proposes a novel training-free inference acceleration method based on our findings.
Feb 28, 2025	Our paper ClearSight: Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models has been accepted to CVPR 2025! In this work, we present a method to mitigate object hallucination in MLLMs by strengthening attention to visual input.