Xinjie Shen

Xinjie Shen, Rongzhe Wei, Peizhi Niu et al.

Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, increasingly capable attackers can distribute their intent across multiple benign-looking turns. Recent studies show that even modern commercial models with advanced guardrails remain vulnerable to such attacks despite advances in safety alignment and external guardrails. In this work, we address this challenge by detecting the earliest turn at which delivering the candidate response would make the accumulated interaction sufficient to enable harmful action. This objective requires precise turn-level intervention that identifies the harm-enabling closure point while avoiding premature refusal of benign exploratory conversations. To further support training and evaluation, we construct the Multi-Turn Intent Dataset (MTID), which contains branching attack rollouts, matched benign hard negatives, and annotations of the earliest harm-enabling turns. We show that MTID helps enable a turn-level monitor TurnGate, which substantially outperforms existing baselines in harmful-intent detection while maintaining low over-refusal rates. TurnGate further generalizes across domains, attacker pipelines, and target models. Our code is available at https://github.com/Graph-COM/TurnGate.

Zehao Jin, Ruixuan Deng, Junran Wang et al.

Activation steering has emerged as a promising alternative for controlling language-model behavior at inference time by modifying intermediate representations while keeping model parameters frozen. However, large-scale evaluations such as AxBench show that existing steering methods are often outperformed by simple in-context prompting and generalize poorly to unseen concepts. We hypothesize that these limitations arise from unvalidated simplifying assumptions shared across prior methods, which typically restrict steering interventions to fixed, single-step, position-invariant transforms. We propose FLAS (Flow-based Activation Steering), which learns a general, concept-conditioned velocity field $v_t(h,t,c)$ that transports unsteered activations to steered ones without relying on these assumptions. On AxBench, FLAS is the first learned method to consistently outperform prompting, reaching held-out harmonic means of $1.015$ on Gemma-2-2B-IT and $1.113$ on Gemma-2-9B-IT without per-concept tuning. Analysis of the learned flow shows curved, multi-step, token-varying trajectories, which suggests that previous hypotheses on activation space geometry might be incomplete.

Xinjie Shen, Danyang Wu, Jitao Lu et al.

How to utilize the pseudo labels has always been a research hotspot in machine learning. However, most methods use pseudo labels as supervised training, and lack of valid assessing for their accuracy. Moreover, applications of pseudo labels in graph neural networks (GNNs) oversee the difference between graph learning and other machine learning tasks such as message passing mechanism. Aiming to address the first issue, we found through a large number of experiments that the pseudo labels are more accurate if they are selected by not overlapping partial labels and defined as negative node pairs relations. Therefore, considering the extraction based on pseudo and partial labels, negative edges are constructed between two nodes by the negative pseudo partial labels extraction (NP$^2$E) module. With that, a signed graph are built containing highly accurate pseudo labels information from the original graph, which effectively assists GNN in learning at the message-passing level, provide one solution to the second issue. Empirical results about link prediction and node classification tasks on several benchmark datasets demonstrate the effectiveness of our method. State-of-the-art performance is achieved on the both tasks.

Xinjie Shen, Zhichao Geng, Yang Yang

With increasing demands for efficiency, information retrieval has developed a branch of sparse retrieval, further advancing towards inference-free retrieval where the documents are encoded during indexing time and there is no model-inference for queries. Existing sparse retrieval models rely on FLOPS regularization for sparsification, while this mechanism was originally designed for Siamese encoders, it is considered to be suboptimal in inference-free scenarios which is asymmetric. Previous attempts to adapt FLOPS for inference-free scenarios have been limited to rule-based methods, leaving the potential of sparsification approaches for inference-free retrieval models largely unexplored. In this paper, we explore $\ell_0$ inspired sparsification manner for inference-free retrievers. Through comprehensive out-of-domain evaluation on the BEIR benchmark, our method achieves state-of-the-art performance among inference-free sparse retrieval models and is comparable to leading Siamese sparse retrieval models. Furthermore, we provide insights into the trade-off between retrieval effectiveness and computational efficiency, demonstrating practical value for real-world applications.

Yujian Liu, Ruoxuan Wu, Xinjie Shen et al.

In the realm of digital pathology, multi-magnification Multiple Instance Learning (multi-mag MIL) has proven effective in leveraging the hierarchical structure of Whole Slide Images (WSIs) to reduce information loss and redundant data. However, current methods fall short in bridging the domain gap between pretrained models and medical imaging, and often fail to account for spatial relationships across different magnifications. Addressing these challenges, we introduce the Concentric Dual Fusion Attention-MIL (CDFA-MIL) framework,which innovatively combines point-to-area feature-colum attention and point-to-point concentric-row attention using concentric patch. This approach is designed to effectively fuse correlated information, enhancing feature representation and providing stronger correlation guidance for WSI analysis. CDFA-MIL distinguishes itself by offering a robust fusion strategy that leads to superior WSI recognition. Its application has demonstrated exceptional performance, significantly surpassing existing MIL methods in accuracy and F1 scores on prominent datasets like Camelyon16 and TCGA-NSCLC. Specifically, CDFA-MIL achieved an average accuracy and F1-score of 93.7\% and 94.1\% respectively on these datasets, marking a notable advancement over traditional MIL approaches.