Recent end-to-end spoken dialogue models enable natural interaction. However, as user demands become increasingly complex, models that rely solely on conversational abilities often struggle to cope. Incorporating agentic capabilities is therefore essential: by enabling tool use, these models can extend their knowledge boundaries and better solve real-world tasks. Yet, existing research has largely concentrated on core perception and generation, with comparatively limited exploration of such tool-augmented extensions. To bridge this gap, we present VoxMind, an integrated framework designed to equip end-to-end spoken dialogue models with comprehensive agentic abilities. Leveraging our curated 470-hour AgentChat dataset, we incorporate a "Think-before-Speak" mechanism, enabling the model to internalize structured reasoning as a critical prerequisite for planning and response generation. Furthermore, to mitigate latency bottlenecks caused by large-scale tool integration, we propose a Multi-Agent Dynamic Tool Management architecture. By asynchronously delegating retrieval tasks to an auxiliary agent aligned with the main model’s reasoning trajectory, this system effectively decouples inference latency from toolset size. Experimental results confirm that VoxMind achieves significant improvements in agent performance: compared with strong baselines, the task completion rate increases from 34.88% to 74.57%, outperforming Gemini-2.5-Pro on spoken agent tasks while preserving general conversational quality. The source code and associated data are publicly available at https://github.com/MM-Speech/VoxMind.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
As large language models (LLMs) are increasingly applied to legal domain-specific tasks, evaluating their ability to perform legal work in real-world settings has become essential. However, existing legal benchmarks rely on simplified and highly standardized tasks, failing to capture the ambiguity, complexity, and reasoning demands of real legal practice. Moreover, prior evaluations often adopt coarse, single-dimensional metrics and do not explicitly assess fine-grained legal reasoning. To address these limitations, we introduce PLawBench, a Practical Law Benchmark designed to evaluate LLMs in realistic legal practice scenarios. Grounded in real-world legal workflows, PLawBench models the core processes of legal practitioners through three task categories: public legal consultation, practical case analysis, and legal document generation. These tasks assess a model’s ability to identify legal issues and key facts, perform structured legal reasoning, and generate legally coherent documents. PLawBench comprises 850 questions across 13 practical legal scenarios, with each question accompanied by expert-designed evaluation rubrics, resulting in approximately 12,500 rubric items for fine-grained assessment. Using an LLM-based evaluator aligned with human expert judgments, we evaluate 10 state-of-the-art LLMs. Experimental results show that none achieves strong performance on PLawBench, revealing substantial limitations in the fine-grained legal reasoning capabilities of current LLMs and highlighting important directions for future evaluation and development of legal LLMs. Data is available at: https://anonymous.4open.science/r/PLawbench-B524/.
Collaboration and information sharing empower Multi-Agent Systems (MAS) but also introduce a critical security risk known as Agent Cascading Injection (ACI). In such attacks, a compromised agent exploits inter-agent trust to propagate malicious instructions, causing cascading failures across the system. However, existing studies consider only limited attack strategies and simplified MAS settings, limiting their generalizability and comprehensive evaluation. To bridge this gap, we introduce ACIArena, a unified framework for evaluating the robustness of MAS. ACIArena offers systematic evaluation suites spanning multiple attack surfaces (i.e., external inputs, agent profiles, inter-agent messages) and attack objectives (i.e., instruction hijacking, task disruption, information exfiltration). Specifically, ACIArena establishes a unified specification that jointly supports MAS construction and attack–defense modules. It covers six widely used MAS implementations and provides a benchmark of 1,356 test cases for systematically evaluating MAS robustness. Our benchmarking results show that evaluating MAS robustness solely through topology is insufficient; robust MAS require deliberate role design and controlled interaction patterns. Moreover, defenses developed in simplified environments often fail to transfer to real-world settings; narrowly scoped defenses may even introduce new vulnerabilities. ACIArena aims to provide a solid foundation for advancing deeper exploration of MAS design principles.
Correct When Paired, Wrong When Split: Decoupling and Editing Modality-Specific Neurons in MLLMs
PDF ↗Although Knowledge Editing provides an efficient mechanism for updating the knowledge of Multimodal Large Language Models (MLLMs), we find that current paradigms still suffer from an important yet remain underexplored issue : editing decoupling failure, where entity-related knowledge can be updated when the model is triggered by multimodal inputs (text–image query pairs), however, it often reverts to outdated pre-edit facts when the paired inputs are split into unimodal ones. Our in-depth empirical analysis reveals that the entity knowledge in MLLMs is not stored as a unified representation, but is instead distributed across disentangled modality-specific pathways. As a result, updates biased toward multimodal queries fail to propagate effectively to unimodal circuits. To bridge this gap, we propose DECODE, which explicitly disentangles and localizes modality-specific neuron groups for targeted knowledge. Extensive experiments demonstrate that DECODE consistently achieves effective knowledge updates under different modality triggers, thereby mitigating editing decoupling failures.
Fine-tuning-as-a-service, while commercially successful for Large Language Model (LLM) providers, exposes models to harmful fine-tuning attacks. As a widely explored defense paradigm against such attacks, unlearning attempts to remove malicious knowledge from LLMs, thereby essentially preventing them from being used to perform malicious tasks. However, we highlight a critical flaw: the inherent general adaptability of LLMs allows them to easily bypass selective unlearning by rapidly relearning or repurposing their general capabilities for harmful tasks. To address this fundamental limitation, we propose a paradigm shift: instead of selective removal, we advocate for inducing model collapse, effectively forcing the model to ”unlearn everything”, specifically in response to updates characteristic of malicious adaptation. This collapse directly neutralizes the very general capabilities that attackers exploit, tackling the core issue unaddressed by selective unlearning. We introduce the Collapse Trap (CTRAP) as a practical mechanism to implement this concept conditionally. Embedded during alignment, CTRAP pre-configures the model’s reaction to subsequent fine-tuning dynamics. If updates during fine-tuning constitute a persistent attempt to reverse safety alignment, the pre-configured trap triggers a progressive degradation of the model’s core language modeling abilities, ultimately rendering it inert and useless for the attacker. Crucially, this collapse mechanism remains dormant during benign fine-tuning, ensuring the model’s utility and general capabilities are preserved.
Many SLT systems quietly assume that brief chunks of signing map directly to spoken-language words. That assumption breaks down because signers often create meaning on the fly using context, space, and movement. We revisit SLT and argue that it is mainly a cross-modal reasoning task, not just a straightforward video-to-text conversion. We thus introduce a reasoning-driven SLT framework that uses an ordered sequence of latent thoughts as an explicit middle layer between the video and the generated text. These latent thoughts gradually extract and organize meaning over time. On top of this, we use a plan-then-ground decoding method: the model first decides what it wants to say, and then looks back at the video to find the evidence. This separation improves coherence and faithfulness. We also built and released a new large-scale gloss-free SLT dataset with stronger context dependencies and more realistic meanings. Experiments across several benchmarks show consistent gains over existing gloss-free methods. Our code and data are available at https://github.com/fletcherjiang/SignThought.
While critical for alignment, Supervised Fine-Tuning (SFT) incurs the risk of catastrophic forgetting, yet the layer-wise emergence of instruction-following capabilities remains elusive. We investigate this mechanism via a comprehensive analysis utilizing information-theoretic, geometric, and optimization metrics across model scales (1B-32B). Our experiments reveal a distinct depth-dependent pattern: middle layers (20%-80%) are stable, whereas final layers exhibit high sensitivity. Leveraging this insight, we propose Mid-Block Efficient Tuning, which selectively updates these critical intermediate layers. Empirically, our method outperforms standard LoRA up to 10.2% on GSM8K (OLMo2-7B) with reduced parameter overhead, demonstrating that effective alignment is architecturally localized rather than distributed. The code is publicly available at https://github.com/lshowway/base.
Enabling Large Language Models (LLMs) to effectively utilize tools in multi-turn interactions is essential for building capable autonomous agents. However, acquiring diverse and realistic multi-turn tool-use data remains a significant challenge. In this work, we propose a novel text-based paradigm. We observe that textual corpora naturally contain rich, multi-step problem-solving experiences, which can serve as an untapped, scalable, and authentic data source for multi-turn tool-use tasks. Based on this insight, we introduce GEM, a data synthesis pipeline that enables the generation and extraction of multi-turn tool-use trajectories from text corpora through a four-stage process: relevance filtering, workflow tool extraction, trajectory grounding, and complexity refinement. To reduce the computational cost, we further train a specialized Trajectory Synthesizer via supervised fine-tuning. This model distills the complex generation pipeline into an efficient, end-to-end trajectory generator. Experiments demonstrate that our GEM-32B achieve a 14.9% improvement on the BFCL V3 Multi-turn benchmark. Our models partially surpass the performance of models trained on -bench (Airline and Retail) in-domain data, highlighting the superior generalization capability derived from our text-based synthesis paradigm. Notably, our Trajectory Synthesizer matches the quality of the full pipeline while significantly reducing inference latency and costs.
Tokenizers play a critical role in large language model studies. Despite recent advances, existing tokenizers fail to explicitly leverage historical tokenization results when making subsequent token decisions, nor do they selectively utilize such history based on contextual relevance. We propose SPEAK, a tokenizer that integrates spiking neurons to explicitly leverage historical tokenization results. Furthermore, we introduce an entropy-aware reset mechanism that selectively leverages history based on contextual relevance, which is determined by token-level entropy. High-entropy tokens are treated as contextual boundaries, whereas low-entropy tokens between consecutive such boundaries exhibit strong contextual relevance. Accordingly, we induce hard reset at high-entropy tokens to discard irrelevant historical tokenization results, and soft reset at low-entropy tokens to preserve and leverage relevant history. Experiments on 2 language models and 5 datasets spanning 16 languages demonstrate superior cross-lingual adaptability, with competitive performance and efficiency. Our code is publicly available at https://github.com/zju-bmi-lab/SPEAK.
Sparrow: Text-Anchored Window Attention with Visual-Semantic Glimpsing for Speculative Decoding in Video LLMs
PDF ↗Although speculative decoding is widely used to accelerate Vision-Language Models (VLMs) inference, it faces severe performance collapse when applied to Video Large Language Models (Vid-LLMs). The draft model typically falls into the trap of attention dilution and negative visual gain due to key-value cache explosion and context window mismatches. We observe a visual semantic internalization phenomenon in Vid-LLMs, indicating that critical visual semantics are implicitly encoded into text hidden states during deep-layer interactions, which renders raw visual inputs structurally redundant during deep inference. To address this, we propose the Sparrow framework, which first utilizes visually-aware text-anchored window attention via hidden state reuse to fully offload visual computation to the target model, and leverages intermediate-layer visual state bridging to train the draft model with semantic-rich intermediate states, thereby filtering out low-level visual noise. Additionally, a multi-token prediction strategy is introduced to bridge the training-inference distribution shift. Experiments show that Sparrow achieves an average speedup of 2.82x even with 25k visual tokens, effectively resolving the performance degradation in long sequences and offering a practical solution for real-time long video tasks.
Large Language Models are increasingly utilized as Role-Playing Agents (RPAs) to simulate personas in interactive settings. However, current RPAs often produce flattened and stereotypical personas with limited depth and fidelity. This limitation arises from two core challenges: insufficient modeling of complex personal histories and internal logic, and ungrounded reasoning that fails to preserve persona coherence as dialogue context evolves. To address these challenges, we propose ThinkPersona, a role-playing agent trained to explicitly ground responses in individual identity. We introduce Persona Graphs as structured representations that encode life trajectories, values, relationships, and events as interconnected knowledge. We construct 1,201 Persona Graphs from real-world interviews and derive a Question–Reasoning–Answer (QRA) dataset of 23,401 samples that supervises reasoning over persona evidence. Fine-tuning on QRA enables ThinkPersona to internalize persona logic and generate persona-consistent responses in long-context dialogues. Experiments on three benchmarks show that ThinkPersona improves role-playing fidelity, behavioral consistency, and grounded reasoning over existing methods, while preserving general instruction-following capabilities. Our code and dataset are available at https://github.com/Hualeez/ThinkPersona.
Deep Research Agents generate analyst-grade reports, yet evaluating them remains challenging due to the absence of a single ground truth and the multidimensional nature of research quality. Recent benchmarks propose distinct methodologies, yet they suffer from the Mirage of Synthesis, where strong surface-level fluency and citation alignment can obscure underlying factual and reasoning defects. We characterize this gap by introducing a taxonomy across four verticals that exposes a critical capability mismatch: static evaluators inherently lack the tool-use capabilities required to assess temporal validity and factual correctness. To address this, we propose **DREAM** (Deep Research Evaluation with Agentic Metrics), a framework that instantiates the principle of capability parity by making evaluation itself agentic. DREAM structures assessment through an evaluation protocol combining query-agnostic metrics with adaptive metrics generated by a tool-calling agent, enabling temporally aware coverage, grounded verification, and systematic reasoning probes. Controlled evaluations demonstrate DREAM is significantly more sensitive to factual and temporal decay than existing benchmarks, offering a scalable, reference-free evaluation paradigm.
Large language models (LLMs), have shown strong potential in scientific discovery, yet existing methods still face substantial challenges in the design of research workflows and multi-role collaboration mechanisms. To mitigate these issues, we propose EvoSci, a multi-agent scientific collaboration framework, which integrates bio-inspired evolution with knowledge graph modeling. To iteratively generate, evaluate, and refine research ideas, EvoSci incorporates multiple role-based agents, including mentor, researcher, and reviewer. By combining collaborative reasoning, shared memory, and evolutionary feedback, EvoSci significantly enhances the coherence and creativity of scientific exploration. Experiments on real-world research topics demonstrate that EvoSci significantly outperforms strong baselines in LLM-based structured peer-review and comparative ranking evaluations, achieving the highest overall peer-review score (ICLR 4.90) and top ranking (Top-10 = 54). These results suggest its superiority in both scientific idea generation and continuous discovery.
Individual investors are significantly outnumbered and disadvantaged in financial markets, overwhelmed by abundant information and lacking professional analysis. Equity research reports stand out as crucial resources, offering valuable insights. By leveraging these reports, large language models (LLMs) can enhance investors’ decision-making capabilities and strengthen financial analysis. However, two key challenges limit their effectiveness: (1) the rapid evolution of market events often outpaces the slow update cycles of existing knowledge bases, (2) the long-form and unstructured nature of financial reports further hinders timely and context-aware integration by LLMs. To address these challenges, we tackle both data and methodological aspects. First, we introduce the Event-Enhanced Automated Construction of Financial Knowledge Graph (FinKario), a dataset comprising over 305,360 entities, 210,328 relational triples, and 19 distinct relation types. FinKario automatically integrates real-time company fundamentals and market events through prompt-driven extraction guided by professional institutional templates, providing structured and accessible financial insights for LLMs. Additionally, we propose a Two-Stage, Graph-Based retrieval strategy (FinKario-RAG), optimizing the retrieval of evolving, large-scale financial knowledge to ensure efficient and precise data access. Extensive experiments show that FinKario with FinKario-RAG achieves superior stock trend prediction accuracy, outperforming financial LLMs by 18.81% and institutional strategies by 17.85% on average in backtesting. [Our code is available at <https://github.com/Jackson906E/FinKario>.]
Early Long-context Document Visual Question Answering (DocVQA) methods struggle with preserving visual semantics or handling finite context windows. Conversely, recent RAG-based approaches suffer from "semantic gaps" and "structural disconnections" due to passive retrieval mechanisms that ignore logical dependencies. To address these challenges, we introduce TRACE (Traversal Retrieval-Augmented Chain of Evidence). By navigating a Bi-Layered Graph that encodes both physical adjacency and semantic relevance, TRACE transforms retrieval from static matching into adaptive evidence chain construction. Furthermore, we propose M5BookVQA, a benchmark designed to assess deep, multi-hop reasoning in books, addressing the limitations of existing datasets. Extensive experiments show that TRACE achieves an average accuracy improvement of 14.07% on M5BookVQA and exhibits robust generalization with a 13.38% gain across four established benchmarks. Our source code is available at https://github.com/shimurenhlq/TRACE.
Puns are a common form of rhetorical wordplay that exploits polysemy and phonetic similarity to create humor. In multimodal puns, visual and textual elements synergize to ground the literal sense and evoke the figurative meaning simultaneously. Although Vision-Language Models (VLMs) are widely used in multimodal understanding and generation, their ability to understand puns has not been systematically studied due to a scarcity of rigorous benchmarks. To address this, we first propose a multimodal pun generation pipeline. We then introduce MultiPun, a dataset comprising diverse types of puns alongside adversarial non-pun distractors. Our evaluation reveals that most models struggle to distinguish genuine puns from these distractors. Moreover, we propose both prompt-level and model-level strategies to enhance pun comprehension, with an average improvement of 16.5% in F1 scores. Our findings provide valuable insights for developing future VLMs that master the subtleties of human-like humor via cross-modal reasoning.
LLM agents operating in open environments face escalating risks from indirect prompt injection, particularly within the tool stream where manipulated metadata and runtime feedback hijack execution flow. Existing defenses encounter a critical dilemma as advanced models prioritize injected rules due to strict alignment while static protection mechanisms sever the feedback loop required for adaptive reasoning. To reconcile this conflict, we propose VIGIL, a framework that shifts the paradigm from restrictive isolation to a verify-before-commit protocol. By facilitating speculative hypothesis generation and enforcing safety through intent-grounded verification, VIGIL preserves reasoning flexibility while ensuring robust control. We further introduce SIREN, a benchmark comprising 959 tool stream injection cases designed to simulate pervasive threats characterized by dynamic dependencies. Extensive experiments demonstrate that VIGIL outperforms state-of-the-art dynamic defenses by reducing the attack success rate by over 22% while more than doubling the utility under attack compared to static baselines, thereby achieving an optimal balance between security and utility. Our code is available at: https://github.com/Touring-686/vigil.
SGPVT: Self-Generated Proximal Visual Tokens for Mitigating Proximal Collateral Damage in MLLM Unlearning
PDF ↗Machine unlearning in multimodal large language models (MLLMs) aims to remove specific concepts while preserving overall utility. However, existing approaches focus primarily on general utility metrics, overlooking the preservation of semantically related concepts. We present the first systematic analysis of this proximal collateral damage, revealing that forgetting vulnerability correlates strongly with visual embedding similarity in a smooth gradient across the semantic space. Based on this insight, we propose a novel unlearning framework that introduces Self-Generated Proximal Visual Tokens (SGPVTs), which are synthetically perturbed visual representations around the target concept. Our method employs an adaptive cosine-band curriculum with a dual-stream objective: forgetting the target via gradient ascent while distilling knowledge from a frozen teacher model into proximal tokens to prevent degradation. Extensive experiments demonstrate that our approach significantly outperforms existing methods in preserving semantically related concepts while achieving effective target unlearning, eliminating the need for manual retention set curation. Our source code will be released in the near future.
Current conversational agents often follow static learning paradigms and miss the implicit, evolving feedback embedded in users’ follow-up behaviors. We propose IEvoAgent, an evolving conversational agent framework that leverages the structured dependency between agent responses and user reactions. We construct an annotated dataset from LMSYS-Chat-1M and WildChat and find consistent response-conditioned feedback patterns. Based on this finding, IEvoAgent uses a conditional feedback distribution matrix to estimate expected feedback rewards, combining offline KTO alignment with an inference-time prompt-evolution mechanism driven by a dynamic matrix. Experiments on MT-Bench-101, WildBench, and FB-Bench show improvements over open-source baselines, indicating that mining implicit feedback supports better multi-turn alignment under evolving user preferences. Our code and dataset are available at https://github.com/Hualeez/IEvoAgent.
Reward modeling is essential for aligning large language models with human preferences, yet predominant architectures rely on a static pooling strategy to condense sequences into scalar scores. This paradigm, however, suffers from two key limitations: a static inductive bias that misaligns with the task-dependent preference signals, and a representational mismatch, as the backbone’s optimization for generation leaves its representations ill-suited to fine-grained discrimination. To address this, we propose AdaJudge, a unified framework that jointly adapts representation and aggregation. AdaJudge first improves backbone representations into a discrimination-oriented space via gated refinement blocks. It then replaces the static readout with an adaptive multi-view pooling module, which dynamically routes and combines evidence. Extensive experiments on RM-Bench and JudgeBench show that AdaJudge outperforms strong off-the-shelf reward models and traditional pooling baselines.