Large language models (LLMs) have demonstrated better safety performance in high-resource languages than in low-resource languages. We attribute this issue as a mismatch gap between language-agnostic semantic understanding ability and language dominant safety alignment biased toward high-resource languages. Based on above insights, we empirically identify the semantic bottleneck in LLMs: intermediate layers in which the geometry of model representations is governed primarily by shared semantic content rather than language identity. Then, we propose Language-Agnostic Semantic Alignment (LASA), which anchors safety alignment directly in semantic bottlenecks. Experiments show that LASA substantially improves safety across all languages: average attack success rate (ASR) drops from 24.7% to 2.8% on LLaMA-3.1-8B-Instruct and remains within 3–4% across Qwen2.5 and Qwen3 Instruct models (7B–32B). Besides, our analysis and method offer a representation-level perspective on LLM safety, suggesting that safety alignment requires anchoring safety understanding not in surface text, but in the model’s language-agnostic semantic space.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Current paradigms for empowering Large Language Models (LLMs) with multilingual capabilities rely heavily on massive instruction tuning. We challenge this view, proposing that the barrier is topological alignment, not data quantity. We introduce Hybrid Cross-Alignment (HCA), fusing a frozen NLLB encoder with a Qwen decoder via a closed-loop dual-adapter architecture. HCA utilizes a Source-Side Adapter to precondition encoder features and a Query-Residual Adapter to preserve generative stability, bridged by an adaptive gated cross-modal interface. Our core discovery is Universal Alignment Generalization.” We demonstrate that training HCA on a single language pair (German-English) unlocks state-of-the-art zero-shot transfer to dozens of unseen languages. Crucially, our Oracle” experiments reveal that this single-pair training recovers over 96.7% of the performance achievable by training on all available pairs. This proves that a universal, language-agnostic projection protocol exists. With a total inference footprint of 5.25B parameters, our model significantly outperforms larger baselines, surpassing TowerPlus-9B (+9.0 COMET on low-resource languages) and Aya-101 (13B). Furthermore, performance scales linearly with encoder size; upgrading from 600M to 1.3B yields immediate gains (+3.4 points on Gujarati) with minimal retraining cost.
Temporal Precision Matters: Brain-Tuning Speech Language Models with Millisecond-Resolution Neural Signals
PDF ↗Brain-tuning enhances brain alignment and downstream performance by fine-tuning speech language models with neural recordings. However, previous work relies primarily on fMRI, whose temporal resolution integrates neural activity over seconds, blending distinct processing stages into a single supervision signal and precluding temporally targeted training. We introduce ECoG-tuning, which leverages electrocorticography’s millisecond precision to train speech language models. We design temporally targeted windows—a speech window capturing acoustic-phonetic encoding and a language window capturing higher-order linguistic processing—grounded in neuroscientific findings about temporal encoding hierarchies. Evaluating three models on the Podcast ECoG dataset, we find that ECoG-tuning significantly improves brain alignment over pretrained and distillation baselines. Notably, full spatiotemporal dynamics yield 7–17% higher alignment than time-averaged supervision across models, and language-window tuning produces larger gains in higher-order language regions, indicating that temporal precision provides additional training value. Moreover, ECoG-tuned models consistently improve or maintain downstream performance. Overall, our work provides initial evidence that electrophysiology is a viable brain-tuning modality, demonstrating how neuroscientific insights into processing hierarchies can inform principled model training strategies. Code is available at [https://github.com/Mochizuki-BUPT/ECoG-Tuning-main](https://github.com/Mochizuki-BUPT/ECoG-Tuning-main).
Embodied-Reasoner: Synergizing Visual Search, Reasoning, and Action for Embodied Interactive Tasks
PDF ↗Recent advances in reasoning models have demonstrated remarkable capabilities on mathematical and coding tasks. However, their effectiveness in embodied domains, where the agent must continuously interact with environments and process observation-action interleaved trajectories, remains largely unexplored. We present Embodied-Reasoner, a reasoning model for interactive embodied tasks. Unlike mathematical reasoning that relies primarily on logical deduction, embodied scenarios demand spatial understanding, temporal reasoning, and ongoing self-reflection based on interaction history. To address these challenges, we synthesize 9.3k coherent Observation-Thought-Action trajectories containing 64k ego-centric images and 90k diverse reasoning processes (analysis, spatial reasoning, reflection, planning, and verification). We develop a three-stage training recipe that progressively enhances the model’s capabilities through imitation learning, rejection sampling tuning on self-exploration trajectories, and reflection tuning. The evaluation shows that our model significantly outperforms advanced visual reasoning models, e.g., exceeds OpenAI o1, o3-mini, and Claude-3.7 by +9%, 24%, and +13%. Analysis reveals that our model exhibits fewer repeated searches and logical inconsistencies, with particular advantages in complex long-horizon tasks. Real-world testing further validates the effectiveness of our approach.
Recent agentic search frameworks enable deep research via iterative planning and retrieval, reducing hallucinations and enhancing factual grounding. However, they remain text-centric, overlooking the multimodal evidence that characterizes real-world expert reports. We introduce a pressing task: multimodal long-form generation. Accordingly, we propose Deep-Reporter, a unified agentic framework for grounded multimodal long-form generation. It orchestrates: (i) Agentic Multimodal Search and Filtering to retrieve and filter textual passages and information-dense visuals; (ii) Checklist-Guided Incremental Synthesis to ensure coherent image-text integration and optimal citation placement; and (iii) Recurrent Context Management to balance long-range coherence with local fluency. We develop a rigorous curation pipeline producing 8K high-quality agentic traces for model optimization. We further introduce M^2LongBench, a comprehensive testbed comprising 247 research tasks across 9 domains and a stable multimodal sandbox. It enables unified multimodal assessment, fair comparison, and accessible evaluation without commercial APIs. Extensive experiments demonstrate that long-form multimodal generation is a challenging task, especially in multimodal selection and integration, and effective post-training can bridge the gap. Our code is available at https://github.com/fangda-ye/Deep-Report.
From Competition to Synergy: Unlocking Reinforcement Learning for Subject-Driven Image Generation
PDF ↗Subject-driven image generation models face a fundamental trade-off between identity preservation (fidelity) and prompt adherence (editability). While online reinforcement learning (RL), specifically GPRO, offers a promising solution, we find that a naive application of GRPO leads to competitive degradation, as the simple linear aggregation of rewards with static weights causes conflicting gradient signals and a misalignment with the temporal dynamics of the diffusion process. To overcome these limitations, we propose Customized-GRPO, a novel framework featuring two key innovations: (i) Synergy-Aware Reward Shaping (SARS), a non-linear mechanism that explicitly penalizes conflicted reward signals and amplifies synergistic ones, providing a sharper and more decisive gradient. (ii) Time-Aware Dynamic Weighting (TDW), which aligns the optimization pressure with the model’s temporal dynamics by prioritizing prompt-following in the early, identity preservation in the later. Extensive experiments demonstrate that our method significantly outperforms naive GRPO baselines, successfully mitigating competitive degradation. Our model achieves a superior balance, generating images that both preserve key identity features and accurately adhere to complex textual prompts.
The Visual Iconicity Challenge: Evaluating Vision-Language Models on Sign Language Form–Meaning Mapping
PDF ↗Iconicity, the resemblance between linguistic form and meaning, is pervasive in sign languages, offering a natural testbed for visual grounding in vision–language models (VLMs). We introduce the Visual Iconicity Challenge, a video-based benchmark that adapts psycholinguistic measures to evaluate VLMs on three tasks: (i) phonological sign-form prediction, (ii) transparency (inferring meaning from visual form), and (iii) graded iconicity ratings. We assess 17 state-of-the-art VLMs in zero- and few-shot settings on Sign Language of the Netherlands and compare them to human baselines. VLMs mirror human phonological difficulty patterns (e.g., handshape harder than location) and achieve moderate to strong alignment with human iconicity ratings. However, they still fail to infer lexical meaning from visual form alone and show a systematic object-based bias that inverts the human preference for action-based signs. Crucially, models with stronger phonological form prediction correlate better with human iconicity judgments, indicating shared sensitivity to visually grounded structure. Our findings validate these diagnostic tasks, show that explicit reasoning narrows the open-to-closed-model calibration gap, and motivate human-centric signals for modelling iconicity in multimodal models.
Proactive dialogue has emerged as a critical and challenging research problem in advancing large language models (LLMs). Existing works predominantly focus on domain-specific or task-oriented scenarios, which leads to fragmented evaluations and limits the comprehensive exploration of models’ proactive dialogue abilities. In this work, we propose ProactiveEval, a unified framework for evaluating proactive dialogue capabilities of LLMs. This framework decomposes proactive dialogue into target planning and dialogue guidance, establishing evaluation metrics across various domains. Moreover, it also enables the automatic generation of diverse and challenging evaluation data. Based on the proposed framework, we develop 328 evaluation environments spanning 6 distinct domains. Through experiments with 22 different types of LLMs, we show that DeepSeek-R1 and Claude-3.7-Sonnet exhibit exceptional performance on target planning and dialogue guidance tasks, respectively. Finally, we investigate how reasoning capabilities influence proactive behaviors and discuss their implications for future model development. Our code and data are available at the https://github.com/liutj9/ProactiveEval.
We investigate belief-like representations in decoder-only autoregressive LLMs using linear controlled probes on residual stream activations and single attention heads. Following Herrmann and Levinstein’s (2025) criteria (Accuracy, Use, Coherence, and Uniformity) we find that large models exhibit strong truth sensitivity (Accuracy), and steering activations along probe directions reliably changes downstream behavior (Use). Coherence, measured via calibrated probes and cross-dataset probing, is moderate across models, while training on diverse data yields domain-consistent truth directions (Uniformity). The results are particularly encouraging at the head level and align with some standard philosophical accounts of belief, e.g., minimal functionalism, supporting the view that LLMs can maintain propositional attitudes under such theoretical frameworks.
Beyond Static Alignment: Adaptive Arbitration for Semantic Incongruence in Semi-Supervised Multimodal Sentiment Analysis
PDF ↗Multimodal sentiment analysis is fundamentally challenged by semantic incongruence, where ambiguous visual signals often conflict with explicit textual cues. In semi-supervised scenarios, naively fusing such noisy features contaminates the joint representation, while conventional static alignment strategies fail to effectively arbitrate conflicting modalities in this task, leading to error reinforcement during self-training. To this end, we propose a novel Adaptive Arbitration for Semantic Incongruence (A2SI) framework for semi-supervised multimodal sentiment analysis, which emphasizes stable cross-modal representations and reliable supervision. Specifically, we first constrain unreliable visual representations by leveraging the reliable textual modality as an anchor to align divergent embeddings and reduce representation noise. Based on this, we further consider the reliability of supervision signals and calibrate pseudo-labels by adaptively weighting evidentiary confidence from heterogeneous views. Finally, to prevent error accumulation caused by unreliable samples, we introduce a progressive arbitration mechanism that verifies pseudo-labeled data from dual perspectives, enabling the model to dynamically balance sample diversity and label purity throughout self-training. Extensive experiments on the MVSA-Single and MVSA-Multiple datasets demonstrate that A2SI consistently outperforms state-of-the-art methods under label-limited settings.
Learning to Conceal Risk: Controllable Multi-turn Red Teaming for LLMs in the Financial Domain
PDF ↗Large Language Models (LLMs) are increasingly deployed in finance, where unsafe behavior can lead to serious regulatory risks. However, most red-teaming research focuses on overtly harmful content and overlooks attacks that appear legitimate on the surface yet induce regulatory-violating responses. We address this gap by introducing a controllable black-box multi-turn risk-concealed redteaming framework (CoRT) that progressively conceals surface-level risk while exploiting regulatory-violating behaviors. CoRT contains two key components: (i) a Risk Concealment Attacker (RCA) that generates multiturn prompts via iterative refinement, and (ii) a Risk Concealment Controller (RCC) that predicts a turn-level Risk Concealment Score (RCS) to steer RCA’s follow-up style. We also build a domain-specific benchmark, FinRisk-Bench, with 522 instructions spanning six financial risk categories. Experiments on nine widely used LLMs show that CoRT (RCA) achieves 93.19% average attack success rate (ASR), and CoRT (RCA+RCC) further improves the average ASR to 95.00%. Our code and FinRisk-Bench are available at https://github.com/gcheng128/CoRT.
Still Between Us? Evaluating and Improving Voice Assistant Robustness to Third-Party Interruptions
PDF ↗While recent Spoken Language Models (SLMs) have been actively deployed in real-world scenarios, they lack the capability to discern Third-Party Interruptions (TPI) from the primary user’s ongoing flow, leaving them vulnerable to contextual failures. To bridge this gap, we introduce TPI-Train, a dataset of 88K instances designed with speaker-aware hard negatives to enforce acoustic cue prioritization for interruption handling, and TPI-Bench, a comprehensive evaluation framework designed to rigorously measure the interruption-handling strategy and precise speaker discrimination in deceptive contexts. Experiments demonstrate that our dataset design mitigates semantic shortcut learning—a critical pitfall where models exploit semantic context while neglecting acoustic signals essential for discerning speaker changes. We believe our work establishes a foundational resource for overcoming text-dominated unimodal reliance in SLMs, paving the way for more robust multi-party spoken interaction. The code for the framework is publicly available at https://tpi-va.github.io
Self-report questionnaires remain the default tool for probing the psychological characteristics of Large Language Model (LLM) agents, yet classical instruments (BFI, BDI, MBTI, BSS) inherit three well-known threats under LLMs: contamination from training corpora, directional bias under social-desirability framing, and limited responsiveness to context beyond the item text. We ask whether a *projective* paradigm can be adapted into a usable psychometric tool for LLM agents. We introduce **GenPT** (Generative Projective Testing), which reformulates TAT, Rorschach, and SCT with newly generated stimuli and organises assessment as a three-stage pipeline (Behavior Collection \to Interpretation \to Diagnosis) grounded in SCORS-G and a Simplified Rorschach Analysis System. On personality traits (Big Five, MBTI) and mental-health risks (depression, suicide ideation), questionnaires exhibit systematic directional shifts under social-desirability framing, most strongly on suicide ideation, whereas GenPT’s collected behavioral patterns stay near the symmetric baseline; under a longitudinal counselling context, GenPT-based depression assessment shifts by roughly an order of magnitude more than its questionnaire counterpart. Questionnaires remain competitive on clean-persona trait tasks where items align lexically with the persona description. Overall, GenPT complements rather than replaces self-report when contamination resistance, bias asymmetry, and context sensitivity matter. Code and stimuli: https://github.com/sci-m-wang/GenPT.
While large language models have achieved remarkable success in various natural language processing tasks, their potential in grammatical error correction remains underexplored. Recent work has applied reinforcement learning with rule-based rewards to CGEC, but these approaches rely on coarse-grained binary signals (exact match or not) that fail to capture fine-grained quality distinctions among correction candidates. In this paper, we propose Edit-Aware Reward Model (EARM), a novel reward modeling framework that explicitly incorporates edit-awareness into preference learning for CGEC. EARM introduces a dual-granularity training objective that jointly optimizes sentence-level and token-level weighted Bradley-Terry ranking losses, where edit tokens receive higher importance weights. When integrated with GRPO, our approach achieves 61.29/63.08 on FCGEC/NaCGEC (single output), and 65.04/64.59 with best-of-16 reranking, surpassing previous best by 5.41 and 1.80 points. Extensive experiments demonstrate that learned edit-aware rewards significantly outperform rule-based alternatives for CGEC preference optimization.
Social media text data are often used to train Machine Learning (ML) models to identify users exhibiting high-risk mental health behaviors. However, sharing this sensitive data poses privacy risks and limits the growth of benchmark datasets. We comprehensively evaluate whether privacy-preserving ML techniques can enable safer data sharing while preserving performance. Specifically, we apply federatedlearning (FL) and Differentially Private FL for two widely-studied mental health prediction tasks: depression detection on X (Twitter) and suicide crisis detection on Reddit. We simulate realistic data-sharing scenarios by treating each user as a client in a non-IID setting, evaluating across different client fractions, aggregation strategies, and privacy budgets. While FL achieves comparable performance to centralized training (centralized 𝐹 1 = 85.63; best FL model 𝐹 1 = 83.16) on depression identification, we find that Differentially Private FL has a large performance-privacy trade-off (up to 𝐹 1 = 27.01 drop) even with low levels of noise (𝜖 = 50). This is due to the distortion of highly informative yet sparse mental health linguistic markers related to mental health, like health topics and emotion words. This research empirically demonstrates the potential and limitations of current privacy preservation techniques for mental health inference tasks.
Existing state-of-the-art symbolic music generation models represent symbolic music as a sequence of attribute tokens with fixed unidirectional dependencies. However, from the perspective of music theory, the attributes of a musical note are inherently a set rather than a sequence. Building on this insight, we propose Amadeus, a novel symbolic music generation framework that adopts a two-level architecture: an autoregressive model for note sequences and a bidirectional discrete diffusion model for note attributes. This design enables flexible attribute control and adjustable decoding speed during inference. To further enhance sequential modeling, we introduce the Conditional Information Enhancement Module (CIEM). We also constructed AMD (Amadeus MIDI Dataset)—the largest open-source symbolic music dataset to date—supporting both pre-training and fine-tuning. We trained two models of different scales, Amadeus and Amadeus-M, and conducted extensive experiments, demonstrating substantial improvements over state-of-the-art methods across both objective and subjective metrics.
LLMs often mistake what sounds true for what is formally valid. This limitation is especially evident in syllogistic reasoning, where plausible arguments can lead models to endorse conclusions that are logically invalid, a phenomenon known as Content Effect (CE).We present Boethius, a schema-guided framework for syllogistic reasoning that disentangles semantic plausibility from logical validity. Boethius adopts an auditable, quasi-formal reasoning process with two complementary stages: a Schema Module, which deduces the underlying logical form by analysing the formal structure of the premises, and an Instantiation Module, which instantiates this form over the concrete argument and evaluates validity independently of content-level semantics.Our results show that Boethius consistently outperforms existing approaches, improving syllogistic reasoning accuracy while substantially reducing CE. These gains hold for both large models in a pure in-context learning setting and smaller models trained via schema-guided trajectories using supervised fine-tuning and optimisation-based refinement.
CNSL-bench: Benchmarking the Sign Language Understanding Capabilities of MLLMs on Chinese National Sign Language
PDF ↗Sign language research has achieved significant progress due to the advances in large language models (LLMs). However, the intrinsic ability of LLMs to understand sign language, especially in multimodal contexts, remains underexplored. To address this limitation, we introduce \textbf{CNSL-bench}, the first comprehensive \textbf{C}hinese \textbf{N}ational \textbf{S}ign \textbf{L}anguage \textbf{bench}mark designed for evaluating multimodal large language models (MLLMs) in sign language understanding. The proposed CNSL-bench is characterized by: 1) Authoritative grounding, as it is anchored to the officially standardized \textit{National Common Sign Language Dictionary}, mitigating ambiguity from regional or non-canonical variants and ensuring consistent semantic definitions; 2) Multimodal coverage, providing aligned textual descriptions, illustrative images, and sign language videos; and 3) Articulatory diversity, supporting fine-grained analysis across key manual articulatory forms, including air-writing, finger-spelling, and the Chinese manual-alphabet. Using CNSL-bench, we extensively evaluate 21 open-source and proprietary up-to-date MLLMs. Our results reveal that, despite recent advances in multimodal modeling, current MLLMs remain substantially inferior to human performance, exhibiting systematic disparities across input modalities and manual articulatory forms. Additional diagnostic analyses suggest that several performance limitations persist beyond improvements in reasoning and that instruction-following robustness varies substantially across models.
Retrieval-Augmented Defense: Adaptive and Controllable Jailbreak Prevention for Large Language Models
PDF ↗Large Language Models (LLMs) remain vulnerable to jailbreak attacks, which attempt to elicit harmful responses from LLMs. The evolving nature and diversity of these attacks pose many challenges for defense systems, including (1) adaptation to counter emerging attack strategies without costly retraining, and (2) control of the trade-off between safety and utility. To address these challenges, we propose Retrieval-Augmented Defense (RAD), a novel framework for jailbreak detection that incorporates a database of known attack examples into Retrieval-Augmented Generation, which is used to infer the underlying, malicious user query and jailbreak strategy used to attack the system. RAD enables training-free updates for newly discovered jailbreak strategies and provides a mechanism to balance safety and utility. Experiments on StrongREJECT show that RAD substantially reduces the effectiveness of strong jailbreak attacks such as PAP and PAIR while maintaining low rejection rates for benign queries. We propose a novel evaluation scheme and show that RAD achieves a robust safety-utility trade-off across a range of operating points in a controllable manner.
Generative Search Engines (GSEs) have reshaped information retrieval, and Generative Engine Optimization (GEO) emerges to improve the content visibility in GSEs’ responses. Previous methods mainly rely on empirical strategies or query-dependent preferences of GSEs for content optimization. However, they remain limited in effectiveness as they overlook the latent user search demands in queries that drive content retrieval and response generation of GSEs. To address this, we propose Mind Reader, a novel GEO method to effectively improve the content visibility within the generated responses of GSEs through content optimization guided by the extracted latent demands of user search. Specifically, we propose a decomposition-recombination query augmentation module, which enriches the query with latent semantic information by decomposing it into diverse perspectives, capturing underlying semantic information, and recombining them into variants to support subsequent optimization. Then, we propose a reasoning coverage content optimization module. By optimizing content to cover critical reasoning information of GSEs, we align the content with the user search demands, effectively improving the content visibility. Extensive experiments on widely used GEO-Bench and our proposed PC-GEO show that our method significantly outperforms baselines and effectively improves content visibility (with up to 2.44x objective metrics and 1.23x subjective metrics on average).