Wenbo Zhou
Famous AuthorPublications
State-Dependent Safety Failures in Multi-Turn Language Model Interaction
Safety alignment in large language models is typically evaluated under isolated queries, yet real-world use is inherently multi-turn. Although multi-turn jailbreaks are empirically effective, the structure of conversational safety failure remains insufficiently understood. In this work, we study safety failures from a state-space perspective and show that many multi-turn failures arise from structured contextual state evolution rather than isolated prompt vulnerabilities. We introduce STAR, a state-oriented diagnostic framework that treats dialogue history as a state transition operator and enables controlled analysis of safety behavior along interaction trajectories. Rather than optimizing attack strength, STAR provides a principled probe of how aligned models traverse the safety boundary under autoregressive conditioning. Across multiple frontier language models, we find that systems that appear robust under static evaluation can undergo rapid and reproducible safety collapse under structured multi-turn interaction. Mechanistic analysis reveals monotonic drift away from refusal-related representations and abrupt phase transitions induced by role-conditioned context. Together, these findings motivate viewing language model safety as a dynamic, state-dependent process defined over conversational trajectories.
Rethinking Multi-Condition DiTs: Eliminating Redundant Attention via Position-Alignment and Keyword-Scoping
While modern text-to-image models excel at prompt-based generation, they often lack the fine-grained control necessary for specific user requirements like spatial layouts or subject appearances. Multi-condition control addresses this, yet its integration into Diffusion Transformers (DiTs) is bottlenecked by the conventional ``concatenate-and-attend'' strategy, which suffers from quadratic computational and memory overhead as the number of conditions scales. Our analysis reveals that much of this cross-modal interaction is spatially or semantically redundant. To this end, we propose Position-aligned and Keyword-scoped Attention (PKA), a highly efficient framework designed to eliminate these redundancies. Specifically, Position-Aligned Attention (PAA) linearizes spatial control by enforcing localized patch alignment, while Keyword-Scoped Attention (KSA) prunes irrelevant subject-driven interactions via semantic-aware masking. To facilitate efficient learning, we further introduce a Conditional Sensitivity-Aware Sampling (CSAS) strategy that reweights the training objective towards critical denoising phases, drastically accelerating convergence and enhancing conditional fidelity. Empirically, PKA delivers a 10.0$\times$ inference speedup and a 5.1$\times$ VRAM saving, providing a scalable and resource-friendly solution for high-fidelity multi-conditioned generation.
Character as a Latent Variable in Large Language Models: A Mechanistic Account of Emergent Misalignment and Conditional Safety Failures
Emergent Misalignment refers to a failure mode in which fine-tuning large language models (LLMs) on narrowly scoped data induces broadly misaligned behavior. Prior explanations mainly attribute this phenomenon to the generalization of erroneous or unsafe content. In this work, we show that this view is incomplete. Across multiple domains and model families, we find that fine-tuning models on data exhibiting specific character-level dispositions induces substantially stronger and more transferable misalignment than incorrect-advice fine-tuning, while largely preserving general capabilities. This indicates that emergent misalignment arises from stable shifts in model behavior rather than from capability degradation or corrupted knowledge. We further show that such behavioral dispositions can be conditionally activated by both training-time triggers and inference-time persona-aligned prompts, revealing shared structure across emergent misalignment, backdoor activation, and jailbreak susceptibility. Overall, our results identify character formation as a central and underexplored alignment risk, suggesting that robust alignment must address behavioral dispositions rather than isolated errors or prompt-level defenses.
GUIGuard: Toward a General Framework for Privacy-Preserving GUI Agents
GUI agents enable end-to-end automation through direct perception of and interaction with on-screen interfaces. However, these agents frequently access interfaces containing sensitive personal information, and screenshots are often transmitted to remote models, creating substantial privacy risks. These risks are particularly severe in GUI workflows: GUIs expose richer, more accessible private information, and privacy risks depend on interaction trajectories across sequential scenes. We propose GUIGuard, a three-stage framework for privacy-preserving GUI agents: (1) privacy recognition, (2) privacy protection, and (3) task execution under protection. We further construct GUIGuard-Bench, a cross-platform benchmark with 630 trajectories and 13,830 screenshots, annotated with region-level privacy grounding and fine-grained labels of risk level, privacy category, and task necessity. Evaluations reveal that existing agents exhibit limited privacy recognition, with state-of-the-art models achieving only 13.3% accuracy on Android and 1.4% on PC. Under privacy protection, task-planning semantics can still be maintained, with closed-source models showing stronger semantic consistency than open-source ones. Case studies on MobileWorld show that carefully designed protection strategies achieve higher task accuracy while preserving privacy. Our results highlight privacy recognition as a critical bottleneck for practical GUI agents. Project: https://futuresis.github.io/GUIGuard-page/