Institution profile

Nankai University

Academic institutionasia · cn
Official website
Research library954linked papers
Opportunities0open roles
Selected work

Representative Papers

Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models

Jun 19, 2025International Conference on Learning Representations

In black-box LLM-as-a-Service (LLMaaS) settings, stealthy backdoor alignment attacks—where models violate safety alignment upon inputs containing hidden triggers—are notoriously difficult to detect. Method: We propose BEAT, the first sample-agnostic, black-box detectable defense leveraging distortions in refusal signals. Its core insight is the “probe concatenation effect”: a stable, significant drop in refusal rate upon backdoor activation. Instead of analyzing output semantics, BEAT monitors the stability of safety signals via multi-sample output distribution estimation, probe concatenation perturbations, and KL-divergence–based distortion quantification—requiring neither gradients nor internal model access. Results: Evaluated on closed- and open-source models including GPT-3.5-turbo, BEAT achieves AUC > 0.96 in detecting diverse backdoor attacks and generalizes effectively against mainstream jailbreak techniques.

9 citationsRead paper

MathScape: Evaluating MLLMs in multimodal Math Scenarios through a Hierarchical Benchmark

Aug 14, 2024arXiv.org

Existing mathematical reasoning benchmarks heavily rely on synthetic images, failing to capture the complexity of real-world multimodal reasoning involving photographs and textual mathematics. Method: We introduce MathScape—the first hierarchical multimodal benchmark for photorealistic mathematical problems—featuring a novel “scene–semantics–task” three-level taxonomy that systematically integrates authentic images with formal mathematical semantics, thereby addressing the longstanding gap in joint vision-language mathematical reasoning evaluation. Contribution/Results: Leveraging 11 state-of-the-art multimodal large language models (MLLMs), we conduct dual-track evaluation assessing both theoretical understanding and practical application. Empirical results reveal that even top-performing models achieve sub-50% average accuracy, exposing critical weaknesses in cross-modal alignment, symbolic parsing, and multi-step reasoning. MathScape establishes a new, high-challenge, fine-grained, and interpretable evaluation paradigm for multimodal mathematical reasoning.

7 citationsRead paper

GS-ROR$^2$: Bidirectional-guided 3DGS and SDF for Reflective Object Relighting and Reconstruction

May 22, 2024

To address weak geometric constraints in 3D Gaussian Splatting (3DGS) and the high computational cost of ray marching plus poor sharp-detail modeling in Signed Distance Function (SDF)-based methods for reconstructing reflective objects, this paper proposes GS-SDF, a bidirectional collaborative framework. Our method jointly optimizes explicit 3DGS and implicit SDF representations through mutual supervision: (1) SDF-guided Gaussian optimization enforces depth and normal consistency to enhance geometric fidelity; (2) Gaussian-predicted normals refine the SDF implicit field to improve mesh quality; and (3) SDF-aware pruning eliminates floating-point artifacts. The framework integrates implicit SDF modeling, bidirectional supervision, deferred shading, and alpha compositing rendering. Experiments show that GS-SDF achieves high-fidelity geometry reconstruction, real-time relightable rendering, and produces high-quality meshes with physically plausible normals—outperforming unimodal approaches—while incurring only a 17% increase in training time.

6 citationsRead paper

A Comprehensive Graph Pooling Benchmark: Effectiveness, Robustness and Generalizability

Jun 13, 2024arXiv.org

Existing graph pooling methods lack standardized, multi-dimensional fairness benchmarks for systematic evaluation. Method: We introduce the first unified evaluation framework covering 17 pooling methods across 28 graph datasets, assessing effectiveness (classification, regression, node-level tasks), robustness (noise injection, adversarial perturbations), and generalization (distribution shift, out-of-distribution generalization), while incorporating efficiency, backbone compatibility (GCN, GAT, GIN), and parameter sensitivity analysis. We conduct extensive experiments with canonical poolers (e.g., TopK, SAGPool, ASAP, DiffPool) and provide visualization and ablation studies. Contribution/Results: Our framework reveals performance boundaries across task types, graph scales, and perturbation regimes. The open-sourced code and benchmark fill a critical gap in graph learning, enabling reproducible, comparable, and fair research—establishing a new standard for evaluating graph pooling methods.

2 citationsRead paper

Infinite-World: Scaling Interactive World Models to 1000-Frame Horizons via Pose-Free Hierarchical Memory

Feb 02, 2026

Existing world models struggle to maintain long-term visual consistency in real-world videos, primarily due to noisy pose estimates and sparse viewpoint revisits. To address this, this work proposes a Hierarchical Pose-free Memory Compressor (HPMC) coupled with an uncertainty-aware three-state logic action discretization scheme, along with a revisit-dense fine-tuning strategy. This framework enables efficient long-horizon modeling without relying on geometric priors or accurate pose information. The approach substantially improves the visual fidelity, action controllability, and spatial coherence of generated videos, achieving— for the first time in real-world settings—coherent interactive generation spanning over 1,000 frames.

1 citationsRead paper
Recent publications

Latest Papers