LLMs are increasingly deployed as interactive agents capable of multi-turn dialogues and tool usage, but they remain vulnerable to adaptive attacks. Existing safety mechanisms are mostly reactive, only attempting to uncover pre-existing hidden malicious intent, which makes them largely ineffective against attacks when such attacks are inherently designed to conceal their intent, and causes them to often fail in multi-turn, interactive, or out-of-distribution scenarios. In addition, existing multi-turn defences are largely data-driven, which limits their generalisation ability. To address this challenge, we propose Proactive Safety Alignment (PSA), a framework that anticipates user intent proactively to align LLM agents’ outputs with safety objectives. PSA combines Safe User Request Prediction (SURP), which anticipates likely safe future requests, and Potentially Harmful Generation (PHG), which evaluates whether the current interaction state can support misuse. Evaluations across multiple benchmarks demonstrate that PSA substantially reduces attack success rates, while preserving overall utility on MT-Bench. Our results show that PSA provides an effective and scalable paradigm for multi-turn LLM safety alignment. The code is available at https://github.com/qinmengyuan/PSA.

Current multimodal LLMs encode images as static visual prefixes and rely on text-based reasoning, lacking goal-driven and adaptive visual access. Inspired by human visual perception—where attention is selectively and sequentially shifted from the most informative regions to secondary cues—we propose Structural Sequential Visual CoT **(SSV-CoT)**. First, a question-relevant saliency map identifies and organizes key visual regions, explicitly modeling the spatial distribution of visual importance. Second, reasoning is performed following this discriminative order, inducing a curriculum-like semantic progression from primary to secondary cues. This method is trained end-to-end, using text cot and answer supervision, without relying on region-level annotations or specialized external tools. Experiments on diverse visual reasoning benchmarks show gains, validating structured and sequential visual cognition.

Enhancing the reasoning capabilities of large language model (LLM) agents requires effective orchestration of diverse expert models and tools. However, existing frameworks typically call APIs, based on coarse-grained matching between tasks and the functions of expert models or tools, while overlooking critical factors such as performance variability and cost efficiency among functionally similar alternatives. To address this, we propose Agora, a framework that uses a confidence-calibrated auction to dynamically allocate tasks to expert models and tools. By treating reasoning steps as tradeable items, Agora bases allocation on calibrated competence rather than raw confidence. Across five main benchmarks, Agora improves or remains competitive with single-model, routing, and cascade baselines under matched candidate pools.

Large language models (LLMs) often suffer from capability stagnation in self-improvement training due to fixed difficulty levels. To address this issue, we propose STRETCH (Self-Taught Reasoning Evolution via Targeted CHallenge) inspired by cognitive scaffolding theory. STRETCH introduces a dynamic Stretch Zone mechanism that continuously aligns question difficulty with the model’s solving capability. Within a single parameter space, the model alternates between a Scaffolder that generates adaptive challenges and a Learner that solves them through reinforcement learning. This dual-loop co-evolution process stabilize training and promote progressive reasoning growth. Experiments on both negotiation and operation research benchmarks demonstrate that STRETCH consistently outperforms strong prompting and domain-specific baselines. Further analysis shows that dynamic difficulty alignment is critical for sustained capability improvement and synchronized reasoning evolution.

Legal consultation question answering (Legal CQA) presents unique challenges compared to traditional legal QA tasks, including the scarcity of high-quality training data, complex task composition, and strong contextual dependencies. To address these, we construct JurisCQAD, a large-scale dataset of over 43,000 real-world Chinese legal queries annotated with expert-validated positive and negative responses, and design a structured task decomposition that converts each query into a legal element graph integrating entities, events, intents, and legal issues. We further propose JurisMA, a modular multi-agent framework supporting dynamic routing, statutory grounding, and stylistic optimization. Combined with the element graph, the framework enables strong context-aware reasoning, effectively capturing dependencies across legal facts, norms, and procedural logic. Trained on JurisCQAD and evaluated on a refined LawBench, our system significantly outperforms both general-purpose and legal-domain LLMs across multiple lexical and semantic metrics, demonstrating the benefits of interpretable decomposition and modular collaboration in Legal CQA.

Emerging research on bias attribution and interpretability have revealed how tokens contribute to biased behavior in language models processing English texts. We build on this line of inquiry by adapting the information-theoretic bias attribution score metric for implementation on models handling agglutinative languages—particularly Filipino. We then demonstrate the effectiveness of our adapted method by using it on a purely Filipino model and on three multilingual models—one trained on languages worldwide and two on Southeast Asian data. Our results show that Filipino models are driven towards bias by words pertaining to people, objects, and relationships—entity-based themes that stand in contrast to the action-heavy nature of bias-contributing themes in English (i.e., criminal, sexual, and prosocial behaviors). These findings point to differences in how English and non-English models process inputs linked to sociodemographic groups and bias.

In this paper, we present the system proposed by our team OldJoe, for the 8th edition of the AVeriTeC shared task, as part of the FEVER workshop. The objective of this task is to verify the factuality of real-world claims. Our approach integrates open source large language models, SQL, and in-context learning. We begin with embedding the knowledge store using a pretrained embedding language model then storing the outputs in a SQL database. Subsequently, we prompt an LLM to craft relevant questions based on the input claim, which are then used to guide the retrieval process. We further prompt the LLM to generate answers to the questions and predict the veracity of the original claim. Our system scored 0.49 on the HU-METEOR AVeriTeC score on the dev set and 0.15 on the Ev2R recall on the test set. Due to the time constraint we were unable to conduct additional experiments or further hyperparameter tuning. As a result, we adopted this pipeline configuration centered on the Qwen3-14B-AWQ model as our final submission strategy. The full pipeline is available on GitHub: https://github.com/farahft/OldJoe

Vision-language models (VLMs) achieve promising results in medical reasoning but struggle with hallucinations, vague descriptions, Inconsistent logic and poor localization. To address this, we propose a agent framework named Medical Visual Reasoning Agent (Med-VRAgent). The approach is based on Visual Guidance and Self-Reward paradigms and Monte Carlo Tree Search (MCTS). By combining the Visual Guidance with tree search, Med-VRAgent improves the medical visual reasoning capabilities of VLMs. We use the trajectories collected by Med-RAgent as feedback to further improve the performance by fine-tuning the VLMs with the proximal policy optimization (PPO) objective. Experiments on multiple medical VQA benchmarks demonstrate that our method outperforms existing approaches.

Narrative-driven recommendation queries are common in question-answering platforms, AI search engines, social forums, and some domain-specific vertical applications. Users typically submit free-form text requests for recommendations, e.g., “Any mind-bending thrillers like Shutter Island you’d recommend?” Such special queries have traditionally been addressed as generic QA task under the RAG paradigm. This work formally introduces narrative recommendation as a distinct task and contends that the RAG paradigm is inherently ill-suited for it, owing to information loss in LLMs when retrieving information from from multiple long and fragmented contexts, and limitations in ranking effectiveness. To overcome these limitations, we propose a novel retrieve-rank paradigm by theoretically demonstrating its superiority over RAG paradigm. Central to this new paradigm, we specially focus on the information retrieval stage and introduce Open-domain Candidate Generation (OCG)-Agent that generatively retrieves structurally adaptive and semantically aligned candidates, ensuring both extensive candidate coverage and high-quality information. We validate effectiveness of new paradigm and OCG-Agent’s retrieve mechanism under real-world datasets from Reddit and corporate education-consulting scenarios. Further extensive ablation studies confirming the rationality of each OCG-Agent component.

Pretrained multilingual models exhibit the same social bias as models processing English texts. This systematic review analyzes emerging research that extends bias evaluation and mitigation approaches into multilingual and non-English contexts. We examine these studies with respect to linguistic diversity, cultural awareness, and their choice of evaluation metrics and mitigation techniques. Our survey illuminates gaps in the field’s dominant methodological design choices (e.g., preference for certain languages, scarcity of multilingual mitigation experiments) while cataloging common issues encountered and solutions implemented in adapting bias benchmarks across languages and cultures. Drawing from the implications of our findings, we chart directions for future research that can reinforce the multilingual bias literature’s inclusivity, cross-cultural appropriateness, and alignment with state-of-the-art NLP advancements.

Recommender systems have become increasingly vital in our daily lives, helping to alleviate the problem of information overload across various user-oriented online services. The emergence of Large Language Models (LLMs) has yielded remarkable achievements, demonstrating their potential for the development of next-generation recommender systems. Despite these advancements, LLM-based recommender systems face inherent limitations stemming from their LLM backbones, particularly issues of hallucinations and the lack of up-to-date and domain-specific knowledge.Recently, Retrieval-Augmented Generation (RAG) has garnered significant attention for addressing these limitations by leveraging external knowledge sources to enhance the understanding and generation of LLMs. However, vanilla RAG methods often introduce noise and neglect structural relationships in knowledge, limiting their effectiveness in LLM-based recommendations. To address these limitations, we propose to retrieve high-quality and up-to-date structure information from the knowledge graph (KG) to augment recommendations. Specifically, our approach develops a retrieval-augmented framework, termed K-RagRec, that facilitates the recommendation generation process by incorporating structure information from the external KG. Extensive experiments have been conducted to demonstrate the effectiveness of our proposed method.

Tool learning empowers large language models (LLMs) as agents to use external tools and extend their utility. Existing methods employ one single LLM-based agent to iteratively select and execute tools, thereafter incorporating execution results into the next action prediction. Despite their progress, these methods suffer from performance degradation when addressing practical tasks due to: (1) the pre-defined pipeline with restricted flexibility to calibrate incorrect actions, and (2) the struggle to adapt a general LLM-based agent to perform a variety of specialized actions. To mitigate these problems, we propose ConAgents, a Cooperative and interactive Agents framework, which coordinates three specialized agents for tool selection, tool execution, and action calibration separately. ConAgents introduces two communication protocols to enable the flexible cooperation of agents. To effectively generalize the ConAgents into open-source models, we also propose specialized action distillation, enhancing their ability to perform specialized actions in our framework. Our extensive experiments on three datasets show that the LLMs, when equipped with the ConAgents, outperform baselines with substantial improvement (i.e., up to 14% higher success rate).

Recent advancements in large language models (LLMs) have been remarkable. Users face a choice between using cloud-based LLMs for generation quality and deploying local-based LLMs for lower computational cost. The former option is typically costly and inefficient, while the latter usually fails to deliver satisfactory performance for reasoning steps requiring deliberate thought processes. In this work, we propose a novel LLM utilization paradigm that facilitates the collaborative operation of large cloud-based LLMs and smaller local-deployed LLMs. Our framework comprises two primary modules: the local agent instantiated with a relatively smaller LLM, handling less complex reasoning steps, and the cloud agent equipped with a larger LLM, managing more intricate reasoning steps. This collaborative processing is enabled through an adaptive mechanism where the local agent introspectively identifies errors and proactively seeks assistance from the cloud agent, thereby effectively integrating the strengths of both locally-deployed and cloud-based LLMs, resulting in significant enhancements in task completion performance and efficiency. We evaluate AdaSwitch across 7 benchmarks, ranging from mathematical reasoning and complex question answering, using various types of LLMs to instantiate the local and cloud agents. The empirical results show that AdaSwitch effectively improves the performance of the local agent, and sometimes achieves competitive results compared to the cloud agent while utilizing much less computational overhead.

Collaborative tasks are ubiquitous activities where a form of communication is required in order to reach a joint goal. Collaborative building is one of such tasks. We wish to develop an intelligent builder agent in a simulated building environment (Minecraft) that can build whatever users wish to build by just talking to the agent. In order to achieve this goal, such agents need to be able to take the initiative by asking clarification questions when further information is needed. Existing works on Minecraft Corpus Dataset only learn to execute instructions neglecting the importance of asking for clarifications. In this paper, we extend the Minecraft Corpus Dataset by annotating all builder utterances into eight types, including clarification questions, and propose a new builder agent model capable of determining when to ask or execute instructions. Experimental results show that our model achieves state-of-the-art performance on the collaborative building task with a substantial improvement. We also define two new tasks, the learning to ask task and the joint learning task. The latter consists of solving both collaborating building and learning to ask tasks jointly.

Open-domain question answering systems need to answer question of our interests with structured and unstructured information. However, existing approaches only select one source to generate answer or only conduct reasoning on structured information. In this paper, we pro- pose a Document-Entity Heterogeneous Graph Network, referred to as DEHG, to effectively integrate different sources of information, and conduct reasoning on heterogeneous information. DEHG employs a graph constructor to integrate structured and unstructured information, a context encoder to represent nodes and question, a heterogeneous information reasoning layer to conduct multi-hop reasoning on both information sources, and an answer decoder to generate answers for the question. Experimental results on HybirdQA dataset show that DEHG outperforms the state-of-the-art methods.

The MultiWOZ 2.0 dataset has greatly boosted the research on dialogue state tracking (DST). However, substantial noise has been discovered in its state annotations. Such noise brings about huge challenges for training DST models robustly. Although several refined versions, including MultiWOZ 2.1-2.4, have been published recently, there are still lots of noisy labels, especially in the training set. Besides, it is costly to rectify all the problematic annotations. In this paper, instead of improving the annotation quality further, we propose a general framework, named ASSIST (lAbel noiSe-robuSt dIalogue State Tracking), to train DST models robustly from noisy labels. ASSIST first generates pseudo labels for each sample in the training set by using an auxiliary model trained on a small clean dataset, then puts the generated pseudo labels and vanilla noisy labels together to train the primary model. We show the validity of ASSIST theoretically. Experimental results also demonstrate that ASSIST improves the joint goal accuracy of DST by up to 28.16% on MultiWOZ 2.0 and 8.41% on MultiWOZ 2.4, compared to using only the vanilla noisy labels.

Dialogue State Tracking (DST) aims to keep track of users’ intentions during the course of a conversation. In DST, modelling the relations among domains and slots is still an under-studied problem. Existing approaches that have considered such relations generally fall short in: (1) fusing prior slot-domain membership relations and dialogue-aware dynamic slot relations explicitly, and (2) generalizing to unseen domains. To address these issues, we propose a novel Dynamic Schema Graph Fusion Network (DSGFNet), which generates a dynamic schema graph to explicitly fuse the prior slot-domain membership relations and dialogue-aware dynamic slot relations. It also uses the schemata to facilitate knowledge transfer to new domains. DSGFNet consists of a dialogue utterance encoder, a schema graph encoder, a dialogue-aware schema graph evolving network, and a schema graph enhanced dialogue state decoder. Empirical results on benchmark datasets (i.e., SGD, MultiWOZ2.1, and MultiWOZ2.2), show that DSGFNet outperforms existing methods.