Large vision-language models (LVLMs) remain highly sensitive to text embedded in images: a misleading word printed, painted, or pasted onto an object can override visual evidence and flip the model’s answer. Existing defenses are often tied to specific vision encoders, require training or feature-level intervention, or do not operate on the autoregressive decoding interface of modern LVLMs. We propose TypoCD, a training-free contrastive decoding method for typographic robustness. TypoCD constructs a text-localized counterfactual view by detecting visible text regions and inpainting only the typographic content, then contrasts the attacked-image and typography-removed logits during decoding. This localized contrastive direction targets the attacker’s textual signal while preserving unrelated visual content, allowing TypoCD to subtract the typography-induced logit shift without modifying model parameters or internal features. Across three open-source LVLM backbones and synthetic, real-world, and safety-oriented typographic benchmarks, TypoCD recovers attacked-image accuracy to within a few percentage points of clean-image performance without per-backbone tuning, closing a gap that often exceeds 40 percentage points under attack. When combined with a simple prompt prefix, TypoCD also substantially reduces jailbreak attack success on safety-oriented typographic attacks.

Autonomous web agents increasingly operate over sensitive personal data, yet existing privacy-preserving approaches rely on static filtering and globally enforced safeguards misaligned with compositional agent execution. We introduce PrivAgentFlow, a memory-guided agentic workflow framework that formulates privacy preservation as adaptive node-level governance over compositional web-agent workflows. PrivAgentFlow decomposes complex tasks into compositional workflow nodes, where each node operates under a distinct privacy policy conditioned on its execution role and information access requirements, enabling privacy governance to be localized at the execution level rather than globally enforced. Such localized governance is refined through execution memory, where historical task failures and privacy violations serve as corrective signals for improving node-level privacy control without exposing raw user data during planning. We provide theoretical insights showing that compositional web-agent workflows induce subtask-specific privacy–utility trade-offs, formally characterizing the advantage of heterogeneous node-level governance over uniform privacy control. Across 246 web-agent tasks on three platforms, PrivAgentFlow improves task utility by 23.1 percentage points (0.415→0.646), cuts environment-level privacy leakage by 92.3% (0.366→0.028), and reduces API-level leakage from 100% to 3.7% over the strongest prior baseline. Our code is available at https://github.com/haorentang/PrivAgentFlow.

Large skill libraries are now standard equipment for LLM agents, and which subset reaches the prompt decides what the agent can do. Enumeration and top-K embedding retrieval both treat skills as independent documents, assuming relevance is a property of one skill at a time. It is not: skills depend on, conflict with, specialize, and duplicate one another, so the closest match to a task can name it while omitting a prerequisite it needs. Graph retrieval encodes these relations but spends them in a fixed pipeline, handing the agent an opaque bundle. We present SkillDAG, which exposes a typed skill graph to the LLM as an agent-callable interface rather than consuming it for the agent: search returns vector matches, typed-edge neighbors, and conflict signals, and a propose-then-commit protocol lets it register execution-backed edges as it works. Across two benchmarks and three model families, SkillDAG outperforms full-library prompting, vector retrieval, and Graph-of-Skills, gaining +14.3 points of ALFWorld success and tying only at ceiling; its ranking holds flat across a 10× pool expansion that degrades the fixed pipeline. Inter-skill structure, these results suggest, is worth more to an agent that can query and revise it than to a retriever that consumes it.

Autonomous LLM agents are increasingly deployed in complex environments as tool-using systems. However, their safety remains fragile, as minor reasoning or retrieval errors can be amplified into hazardous actions within the agentic workflow. Existing defenses, often limited to static prompts or post-hoc guardrails, fail to provide runtime intervention or cross-architecture portability. In this paper, we propose Safety Sidecar, a model-agnostic, plug-and-play module designed to provide standardized runtime safety control and auditability for arbitrary agent workflows. Safety Sidecar operationalizes reflection as a closed-loop controller: it dynamically monitors decision traces, retrieves evidence-based repair exemplars from a reflective memory, and enforces risk-mitigating revisions before execution. Crucially, it employs external verifiers to gate both action release and memory updates, producing a transparent, auditable trail of retrieved evidence and applied constraints.We instantiate and systematically evaluate Safety Sidecar in secure code generation—a high-stakes domain with objective vulnerability signals. Experimental results across eight CWE scenarios and four representative LLMs demonstrate that Safety Sidecar consistently improves the secure-solution rate by 2.9–11.2 percentage points while maintaining competitive functional correctness. Efficiency analysis shows the framework is practical for deployment, with reflection adding only 3.2s to end-to-end latency and a negligible average cost of 5.37 × 10^-4 per scenario. Our findings position Safety Sidecar as a portable and efficient control layer for enhancing the safety, compliance, and auditability of LLM-based agents.

Gradient conflict between forgetting and retaining objectives is a well-known obstacle in machine unlearning, where opposing gradients on shared parameters cause destructive interference. In Multimodal Large Language Models (MLLMs), cross-modal coupling entangles visual and textual representations within the same parameters, making this conflict considerably more severe. We propose EASE (Efficient Alignment-aware Selective Erasing), which explicitly balances the conflicting gradient directions. EASE first dynamically locates the conflicting parameters at each step by detecting opposite-sign gradients, then performs a manifold-constrained optimization within the two-dimensional subspace spanned by the task gradients to find the linear combination that maximally aligns them, yielding a valid descent direction for both objectives without any manually tuned trade-off coefficient. Nonconflicting parameters keep their standard updates, keeping the method lightweight. Experiments across multiple MLLM architectures and unlearning benchmarks show that EASE achieves superior unlearning efficiency and faster convergence while matching or surpassing existing methods in multimodal quality.

Advancing from usable to collaborative autonomy requires driving systems to execute passenger instructions safely and reliably. This work formulates instruction realization as scheduling across multiple motion planners and presents a dual-loop framework that provides a transparent decision chain from natural language to vehicle control. The outer loop uses a small language model (SLM) for high-level, low-frequency semantic reasoning and schedule generation, while the inner loop performs low-level, high-frequency schedule execution and vehicle control. To compensate for the SLM’s limited capacity, the framework integrates receding-horizon scheduling to segment long-horizon instruction tasks, a domain-specific language (DSL) that restricts SLM outputs to a scheduling-oriented subspace, and reinforcement learning in high-fidelity urban traffic to refine the SLM’s DSL proficiency and scheduling performance. Experiments show that the framework improves instruction-completion rates while maintaining high safety and compliance relative to multiple baselines.

Large language models (LLMs) face significant challenges in sustaining long-term memory for agentic applications due to limited context windows. To address this limitation, many work has proposed diverse memory mechanisms to support long-term, multi-turn interactions, leveraging different approaches tailored to distinct memory storage objects, such as KV caches. In this survey, we present a unified taxonomy that organizes memory systems for long-context scenarios by decoupling memory abstractions from model-specific inference and training methods. We categorize LLM memory into three primary paradigms: natural language tokens, intermediate representations and parameters. For each paradigm, we organize existing methods by three management stages, including memory construction, update, and query, so that long-context memory mechanisms can be described in a consistent way across system designs, with their implementation choices and constraints made explicit. Finally, we outline key research directions for long-context memory system design.

In the realm of multi-intent spoken language understanding, recent advancements have leveraged the potential of prompt learning frameworks. However, critical gaps exist in these frameworks: the lack of explicit modeling of dual-task dependencies and the oversight of task-specific semantic differences among utterances. To address these shortcomings, we propose DC-Instruct, a novel generative framework based on Dual-task Inter-dependent Instructions (DII) and Supervised Contrastive Instructions (SCI). Specifically, DII guides large language models (LLMs) to generate labels for one task based on the other task’s labels, thereby explicitly capturing dual-task inter-dependencies. Moreover, SCI leverages utterance semantics differences by guiding LLMs to determine whether a pair of utterances share the same or similar labels. This can improve LLMs on extracting and discriminating task-specific semantics, thus enhancing their SLU reasoning abilities. Extensive experiments on public benchmark datasets show that DC-Instruct markedly outperforms current generative models and state-of-the-art methods, demonstrating its effectiveness in enhancing dialogue language understanding and reasoning.

Abstractive related work generation has attracted increasing attention in generating coherent related work that helps readers grasp the current research. However, most existing models ignore the inherent causality during related work generation, leading to spurious correlations which downgrade the models’ generation quality and generalizability. In this study, we argue that causal intervention can address such limitations and improve the quality and coherence of generated related work. To this end, we propose a novel Causal Intervention Module for Related Work Generation (CaM) to effectively capture causalities in the generation process. Specifically, we first model the relations among the sentence order, document (reference) correlations, and transitional content in related work generation using a causal graph. Then, to implement causal interventions and mitigate the negative impact of spurious correlations, we use do-calculus to derive ordinary conditional probabilities and identify causal effects through CaM. Finally, we subtly fuse CaM with Transformer to obtain an end-to-end related work generation framework. Extensive experiments on two real-world datasets show that CaM can effectively promote the model to learn causal relations and thus produce related work of higher quality and coherence.

The task of joint dialog sentiment classification (DSC) and act recognition (DAR) aims to simultaneously predict the sentiment label and act label for each utterance in a dialog. In this paper, we put forward a new framework which models the explicit dependencies via integrating prediction-level interactions other than semantics-level interactions, more consistent with human intuition.Besides, we propose a speaker-aware temporal graph (SATG) and a dual-task relational temporal graph (DRTG) to introduce temporal relations into dialog understanding and dual-task reasoning. To implement our framework, we propose a novel model dubbed DARER, which first generates the context-, speaker- and temporal-sensitive utterance representations via modeling SATG, then conducts recurrent dual-task relational reasoning on DRTG, in which process the estimated label distributions act as key clues in prediction-level interactions.Experiment results show that DARER outperforms existing models by large margins while requiring much less computation resource and costing less training time.Remarkably, on DSC task in Mastodon, DARER gains a relative improvement of about 25% over previous best model in terms of F1, with less than 50% parameters and about only 60% required GPU memory.

Recent joint multiple intent detection and slot filling models employ label embeddings to achieve the semantics-label interactions.However, they treat all labels and label embeddings as uncorrelated individuals, ignoring the dependencies among them. Besides, they conduct the decoding for the two tasks independently, without leveraging the correlations between them.Therefore, in this paper, we first construct a Heterogeneous Label Graph (HLG) containing two kinds of topologies: (1) statistical dependencies based on labels’ co-occurrence patterns and hierarchies in slot labels; (2) rich relations among the label nodes.Then we propose a novel model termed ReLa-Net.It can capture beneficial correlations among the labels from HLG.The label correlations are leveraged to enhance semantic-label interactions. Moreover, we also propose the label-aware inter-dependent decoding mechanism to further exploit the label correlations for decoding. Experiment results show that our ReLa-Net significantly outperforms previous models.Remarkably, ReLa-Net surpasses the previous best model by over 20% in terms of overall accuracy on MixATIS dataset.

Recent graph-based models for joint multiple intent detection and slot filling have obtained promising results through modeling the guidance from the prediction of intents to the decoding of slot filling.However, existing methods (1) only model the unidirectional guidance from intent to slot; (2) adopt homogeneous graphs to model the interactions between the slot semantics nodes and intent label nodes, which limit the performance.In this paper, we propose a novel model termed Co-guiding Net, which implements a two-stage framework achieving the mutual guidances between the two tasks.In the first stage, the initial estimated labels of both tasks are produced, and then they are leveraged in the second stage to model the mutual guidances.Specifically, we propose two heterogeneous graph attention networks working on the proposed two heterogeneous semantics-label graphs, which effectively represent the relations among the semantics nodes and label nodes.Experiment results show that our model outperforms existing models by a large margin, obtaining a relative improvement of 19.3% over the previous best model on MixATIS dataset in overall accuracy.