Reads MHSA, FFN, and residual-stream outputs at every layer through the model’s own logit lens, then compresses their entropy trajectories into a 3L-dimensional representation. The three module-wise signals provide complementary evidence for hallucination detection without storing hidden states or sampling multiple generations.
EMNLP
DECENT: Distribution-aware Prototypical Calibration with Fine-grained Voting for VLM Misclassification Detection
Yijun Gong, Bohan Yang, Zhi Zhang, Yupeng Qin, Wenpeng Xing, Changting Lin, Meng Han, Ziyue Qiao†, and Xiao Luo
Calibrates VLM predictions against distribution-aware class prototypes and aggregates fine-grained voting evidence across the decision space. The two-stage design targets subtle misclassifications that may remain hidden behind a single confident output.
TMLR
COLOR: Coarse-to-fine Outlier Detection with Subspace Purification for Open-Vocabulary Semantic Segmentation
Yijun Gong, Bohan Yang, Zhi Zhang, Ziyue Qiao†, and Xiao Luo
Transactions on Machine Learning Research (TMLR) , 2026
First purifies open-vocabulary features into a cleaner semantic subspace, then applies coarse-to-fine outlier scoring at complementary granularities. The design aims to preserve known-class segmentation while separating unfamiliar regions that conventional confidence scores can overlook.
TNNLS
Latent Fusion Jailbreak: Blending Harmful and Harmless Representations to Elicit Unsafe LLM Outputs
Wenpeng Xing, Bohan Yang, Mohan Li, Chunqiang Hu, Haitao Xu, Ningyu Zhang, Bo Lin, and Meng Han†
Pairs harmful queries with structurally similar benign counterparts and uses refusal-loss gradients to choose layers and token positions for hidden-state interpolation. Across five open-weight models and four safety benchmarks, it reaches 94.01% average attack success; latent-space adversarial training reduces attack success by over 80%.
Under review
Rethinking Test-Time Scaling: From Answer Discovery to Candidate Selectability
Builds a unified generation-and-selection benchmark spanning three open-weight models, five datasets, seven generation policies, seven label-free selectors, and four candidate budgets. Larger pools improve oracle coverage far more than returned-answer accuracy, revealing candidate selectability as the main bottleneck.
ICIP
QuadBox: Accelerating 3D Gaussian Splatting with Geometry-Aware Boxes
Wraps projected Gaussians with four tight tile-aligned boxes and traverses their intersections in one pass using simple interval tests. By removing irrelevant Gaussian-tile checks, the method accelerates 3DGS rendering by 1.85x on public datasets.