Knowledge-Based Visual Question Answering (KB-VQA) requires retrieving external textual knowledge to answer visual queries. However, current methods often suffer from two critical misalignments during retrieval: distributional misalignment, driven by the severe stylistic shift from descriptive pre-training image captions to formal knowledge texts, and granularity misalignment, where lengthy knowledge documents contain excessive noise that drowns out key information, making it difficult to accurately match the visual query. To resolve these challenges, we propose RGCC, a Retriever-Guided Context Compression framework. By mining alignment-friendly data and further training a context compression model, RGCC condenses lengthy documents into compact, retriever-preferred representations that significantly improve cross-modal retrieval. Extensive experiments on the InfoSeek and E-VQA benchmarks demonstrate that RGCC achieves state-of-the-art retrieval performance and boosts downstream question-answering accuracy without additional online inference overhead.

LLM-based agents face challenges in personalized interactive tasks due to heterogeneous, multi-typed, and implicitly constrained long-term traces. Existing memory mechanisms struggle with accurate routing and retrieval, especially on-device where personalization is critical. Most methods use single-vector representations, blurring type distinctions and relational structure. We propose HGP, a hybrid graph memory framework. HGP employs a lightweight self-enhancement classifier for personalized memory routing and constructs episodic, semantic, and procedural memories as graphs. It also extracts working memory as a state trajectory to capture current state and implicit constraints, ensuring reliable decision- making. The classifier reduces large-model calls, enabling on-device deployment, while graph storage enables accurate retrieval and incremental user profile refinement. Experiments on two benchmarks show that on PAL-Set solution selection, HGP achieves an S-score of 35.58, nearly 7 points above the strongest baseline.

Large language models remain fragile against malicious fine-tuning, motivating training-time defenses against harmful persona drift. Preventative Steering injects undesirable-trait persona vectors during fine-tuning and removes them at evaluation time, yet the mechanism behind its lasting protection remains unclear. Analyzing its temporal optimization dynamics, we find that the defense emerges from an early compensatory adaptation phase followed by a steady-state phase where the corrective signal decays; in parameter space, attention output projections emerge as the dominant residual-write route for defensive updates. Through Intervention Delta Preservation (IDP) and IDP Continuation experiments, we further show that preserving or reinjecting the weight offset fails to maintain protection, indicating that preventative steering relies on active adaptation rather than a static defense. Motivated by this finding, we propose Progressive Intensity Scheduling (PIS), which starts with a moderate injection strength and increases it after static-strength alignment begins to decay. Across the evaluated Qwen2.5 and Gemma-3 models, PIS improves safety robustness over static-strength steering while reducing harmful trait expression.

The enhanced capabilities of LLM-based agents bring an urgent need to address safety risks in planning and tool use. Due to the helpfulness–harmlessness trade-off in LLM alignment, agents typically also inherit the flaw of Over-Refusal, which is a passive failure mode. However, the proactive planning and action capabilities of agents introduce another emerging danger on the opposite side of this trade-off. We term this phenomenon **Toxic Proactivity**: an active failure mode in which an agent, driven by the optimization for Machiavellian helpfulness, prioritizes utility over ethics. Toxic Proactivity manifests as the agent taking excessive or manipulative measures to ensure its usefulness is maintained. Existing research pays little attention to identifying this behavior, as it often lacks the subtle context required for such strategies to unfold. To reveal this risk, we introduce a novel evaluation framework based on dilemma-driven interactions between dual models, enabling the simulation and analysis of agent behavior over multi-step behavioral trajectories. Through extensive experiments with multiple mainstream LLMs, we demonstrate that Toxic Proactivity is a widespread behavioral phenomenon and reveal two major tendencies: direct violation and strategic misalignment. We further present a systematic benchmark for evaluating Toxic Proactive behavior across contextual settings.

Self-Refinement has emerged as a promising paradigm for improving Large Language Model (LLM) performance by iteratively improving its response at inference time. Existing training methods typically treat refinement as a decoupled task and rely on pre-generated responses, thereby detaching training from the model’s actual reasoning trajectories and hindering performance. In this paper, we propose FlipTS (Failure-to-Success Learning for self-refinement with Potential-guided Trajectory-Aware Sampling), an end-to-end reinforcement learning (RL) framework that optimizes the entire self-refinement loop online. We identify a fundamental bottleneck in this end-to-end setting: the sparsity of informative refinement behaviors Failure-to-Success (F2S). FlipTS addresses this by utilizing a non-stationary Bayesian model to estimate data potential and applying trajectory-aware sampling to enrich the training process with valuable refinement signals. Experiments on Qwen3-4B-Instruct, Qwen3-1.7B and LLaMA-3.1-8B-Instruct across challenging mathematical benchmarks demonstrate that FlipTS consistently outperforms both offline and online RL baselines. Notably, it exhibits robust generalization to factuality, scientific reasoning and legal reasoning tasks without domain-specific tuning. Our code will be open-sourced.

Deep Research Agents (DRAs) aim to solve complex, long-horizon research tasks involving planning, retrieval, multimodal understanding, and report generation, yet their evaluation remains challenging due to dynamic web environments and ambiguous task definitions. We propose **DR³-Eval**, a realistic and reproducible benchmark for evaluating deep research agents on multimodal, multi-file report generation. DR³-Eval is constructed from authentic user-provided materials and paired with a per-task static research sandbox corpus that simulates open-web complexity while remaining fully verifiable, containing supportive documents, distractors, and noise. Moreover, we introduce a multi-dimensional evaluation framework measuring Information Recall, Factual Accuracy, Citation Coverage, Instruction Following, and Depth Quality, and validate its alignment with human judgments. Experiments with our developed multi-agent system DR³-Agent based on multiple state-of-the-art language models demonstrate that DR³-Eval is highly challenging and reveals critical failure modes in retrieval robustness and hallucination control.

Large reasoning models (LRMs) have attracted much attention due to their exceptional performance. However, their performance mainly stems from thinking, a long Chain of Thought (CoT), which significantly increase computational overhead. To address this overthinking problem, existing work focuses on using reinforcement learning (RL) to train hybrid reasoning models that automatically decide whether to engage in thinking or not based on the complexity of the query. Unfortunately, using RL will suffer the the reward hacking problem, e.g., the model engages in thinking but is judged as not doing so, resulting in incorrect rewards.To mitigate this problem, existing works either employ supervised fine-tuning (SFT), which incurs high computational costs, or enforce uniform token limits on non-thinking responses, which yields limited mitigation of the problem.In this paper, we propose Thinking-Based Non-Thinking (TNT). It does not employ SFT, and sets different maximum token usage for responses not using thinking across various queries by leveraging information from the solution component of the responses using thinking. Experiments on five mathematical benchmarks demonstrate that TNT reduces token usage by around 50\\%$ compared to DeepSeek-R1-Distill-Qwen-1.5B/7B and DeepScaleR-1.5B, while significantly improving accuracy. In fact, TNT achieves the optimal trade-off between accuracy and efficiency among all tested methods. Additionally, the probability of reward hacking problem in TNT’s responses, which are classified as not using thinking, remains below $10\\%$ across all tested datasets.

Large Language Models (LLMs), due to substantial computational requirements, are vulnerable to resource consumption attacks, which can severely degrade server performance or even cause crashes, as demonstrated by denial-of-service (DoS) attacks designed for LLMs. However, existing works lack mitigation strategies against such threats, resulting in unresolved security risks for real-world LLM deployments. To this end, we propose the Pluggable and Dynamic DoS-Defense Framework (PD³F), which employs a two-stage approach to defend against resource consumption attacks from both the input and output sides. On the input side, we propose the Resource Index to guide Dynamic Request Polling Scheduling, thereby reducing computing resource usage induced by malicious prompts under high-concurrency scenarios. On the output side, we introduce the Adaptive End-Based Suppression mechanism, which reduces excessive malicious generation. Experiments across six models demonstrate that PD³F significantly mitigates resource consumption attacks, improving users’ access capacity by up to 500% during adversarial load. PD³F represents a step toward the resilient and resource-aware deployment of LLMs against resource consumption attacks.

Syntactic information is essential for both sentiment analysis(SA) and aspect-based sentiment analysis(ABSA). Previous work has already achieved great progress utilizing Graph Convolutional Network(GCN) over dependency tree of a sentence. However, these models do not fully exploit the syntactic information obtained from dependency parsing such as the diversified types of dependency relations. The message passing process of GCN should be distinguished based on these syntactic information. To tackle this problem, we design a novel weighted graph convolutional network(WGCN) which can exploit rich syntactic information based on the feature combination. Furthermore, we utilize BERT instead of Bi-LSTM to generate contextualized representations as inputs for GCN and present an alignment method to keep word-level dependencies consistent with wordpiece unit of BERT. With our proposal, we are able to improve the state-of-the-art on four ABSA tasks out of six and two SA tasks out of three.

Open-vocabulary slots, such as file name, album name, or schedule title, significantly degrade the performance of neural-based slot filling models since these slots can take on values from a virtually unlimited set and have no semantic restriction nor a length limit. In this paper, we propose a robust adversarial model-agnostic slot filling method that explicitly decouples local semantics inherent in open-vocabulary slot words from the global context. We aim to depart entangled contextual semantics and focus more on the holistic context at the level of the whole sentence. Experiments on two public datasets show that our method consistently outperforms other methods with a statistically significant margin on all the open-vocabulary slots without deteriorating the performance of normal slots.