论文检索

Dongliang Chen, Xinlin Zhuang, Junjie Xu, Luojian Xie, Zehui Wang, Jiaxi Zhuang, Haolin Yang, Liang Dou, Xiao He, Xingjiao Wu 等

Multi-objective alignment for text-to-image generation is commonly implemented via static linear scalarization, but fixed weights often fail under heterogeneous rewards, leading to optimization imbalance where models overfit high-variance, high-responsiveness objectives (e.g., OCR) while under-optimizing perceptual goals. We identify two mechanistic causes: variance hijacking, where reward dispersion induces implicit reweighting that dominates the normalized training signal, and gradient conflicts, where competing objectives produce opposing update directions and trigger seesaw-like oscillations. We propose APEX (Adaptive Priority-based Efficient X-objective Alignment), which stabilizes heterogeneous rewards with Dual-Stage Adaptive Normalization and dynamically schedules objectives via \mathcal{P}^3 Adaptive Priorities that combine learning potential, conflict penalty, and progress need. On Stable Diffusion 3.5, APEX achieves improved Pareto trade-offs across four heterogeneous objectives, with balanced gains of +1.31 PickScore, +0.35 DeQA, and +0.53 Aesthetics while maintaining competitive OCR accuracy, mitigating the instability of multi-objective alignment.

Maike Züfle, Ondrej Klejch, Nicholas Sanders, Jan Niehues, Alexandra Birch, Tsz Kin Lam

Spoken conversational systems require more than accurate speech generation to have human-like conversations: to feel natural and engaging, they must produce conversational behaviour that adapts dynamically to the context. Current spoken conversational systems, however, rarely allow such customization, limiting their naturalness and usability. In this work, we present the first open, instruction-following full-duplex conversational speech model that can be trained efficiently under typical academic resource constraints. By keeping the audio encoder frozen and finetuning only the language model, our model requires just 2,000 hours of data, without relying on large-scale pretraining or multi-stage optimization. The model can follow explicit instructions to control speaker voice, conversation topic, conversational behaviour (e.g., backchanneling and interruptions), and dialogue initiation. We propose a single-stage training protocol and systematically analyze design choices. Both the model and training code is released to enable reproducible research on controllable full-duplex speech systems.

Jinyang Du, Ruihao Gong, Linghan Ai, Zining Wang, Yunke Peng, Yao Wang, Lei Yan, Wxuefei, Yaoyuan Wang, Jinyang Guo 等

Training large language models (LLMs) at 4-bit precision offers substantial efficiency gains but remains challenging due to the limited dynamic range and coarse numerical resolution. Existing 4-bit training pipelines typically rely on max-scaling, which is ill-suited for heavy-tailed LLM tensor distributions and leads to severe under-utilization of the FP4 quantization grid in the low-magnitude region. This effect causes pronounced representation collapse and large rounding errors for the values that dominate LLM computation. In this work, we derive the theoretically optimal scaling for FP4 under heavy-tailed inputs, revealing why max-scaling is intrinsically suboptimal. Guided by this analysis, we propose Half-S, a simple and efficient scaling strategy that uses half-scaling as a hardware-friendly default and falls back to an MSE-based clipping threshold when needed, yielding a close approximation to the theoretical optimum under real LLM statistics. Extensive experiments on large-scale pretraining and downstream fine-tuning show that Half-S consistently narrows the gap to BF16 in both convergence and final model quality, while preserving the efficiency benefits of 4-bit computation. Under native FP4 support, Half-S is estimated to provide up to 1.8\times end-to-end training speedup. These results indicate that Half-S provides a simple and effective correction to max-scaling, substantially improving the stability and accuracy of 4-bit LLM training.

Wenpeng Xing, Zhonghao Qi, Yupeng Qin, Yilin Li, Caini Chang, Jiahui Yu, Changting Lin, Zhenzhen Xie, Meng Han

While Large Language Models (LLMs) have achieved remarkable performance, they remain vulnerable to jailbreak. The integration of Large Language Models (LLMs) with external tools via protocols such as the Model Context Protocol (MCP) introduces critical security vulnerabilities, including prompt injection, data exfiltration, and other threats. To counter these challenges, we propose MCP-Guard, a robust, layered defense architecture designed for LLM–tool interactions. MCP-Guard employs a three-stage detection pipeline that balances efficiency with accuracy: it progresses from lightweight static scanning for overt threats and a deep neural detector for semantic attacks, to our fine-tuned E5-based model achieves 96.01% accuracy in identifying adversarial prompts. Finally, an LLM arbitrator synthesizes these signals to deliver the final decision. To enable rigorous training and evaluation, we introduce MCP-AttackBench, a comprehensive benchmark comprising 70,448 samples augmented by GPT-4. This benchmark simulates diverse real-world attack vectors that circumvent conventional defenses in the MCP paradigm, thereby laying a solid foundation for future research on securing LLM-tool ecosystems.

Feihu Jin, Ying Tan

Fine-tuning large language models (LLMs) using standard first-order (FO) optimization oftendrives training toward sharp, poorly generalizing minima. Conversely, zeroth-order (ZO) methods offer stronger exploratory behaviorwithout relying on explicit gradients, yet suffer from slow convergence. More critically, our analysis reveals that in generative tasks, the vast output and search space significantly amplify estimation variance, rendering ZO methods both noisy and inefficient. To address these challenges, we propose Hi-ZFO (Hierarchical Zeroth- and First-Order optimization), a hybrid framework designed to synergize the precision of FO gradients with the exploratory capability of ZO estimation. Hi-ZFO adaptively partitions the model through layer-wise importance profiling, applying precise FO updates to critical layers while leveraging ZO optimization for less sensitive ones. Notably, ZO in Hi-ZFO is not merely a memory-saving surrogate; it is intentionally introduced as a source of "beneficial stochasticity" to help the model escape the local minima where pure FO optimization tends to stagnate. Validated across diverse generative, mathematical, and code reasoning tasks, Hi-ZFO consistently achieves superior performance while significantly reducing the training time. These results demonstrate the effectiveness of hierarchical hybrid optimization for LLM fine-tuning.

Kun Huang, Rui Qiu, Xiaoming Li, Salah Uddin

Recent advances in Large Vision–language Models (VLMs) suggest their potential for multimodal misinformation detection. However, existing multimodal misinformation detectors often fail to effectively integrate them, relying instead on passive aggregation of multimodal features and social signals. Such correlation-driven paradigms are vulnerable to spurious associations and multimodal noise, and lack explicit verification mechanisms. In this paper, we propose Logic-Guided Adaptive Reasoning (LoGAR), a verification-oriented framework that integrates VLMs into multimodal misinformation detection through explicit rationale-guided reasoning. LoGAR leverages a VLM to generate an explicit verification rationale, which serves as a global semantic anchor to condition the entire reasoning process. Concretely, the rationale functions as an active query to guide multimodal feature fusion and as a conditioning signal to modulate message passing over heterogeneous social graphs, enabling hypothesis-aware evidence aggregation. Furthermore, LoGAR introduces an instance-aware adaptive depth mechanism that dynamically determines the required reasoning depth. Experimental results on multiple multimodal misinformation benchmarks demonstrate that LoGAR consistently outperforms state-of-the-art methods while significantly reducing computational cost.

Meng Zhang, Ruochun Jin, Yuanxi Peng, Wenjing Yang, Haotian Wang, Liting Sun, Kun Hu, Silin Yang, Zhang Ke-di

Text-to-SQL aims to bridge the gap between human intent and relational databases. While LLMs have shown proficiency in generating simple SQL queries, they struggle with complex analytical tasks. Moreover, models fine-tuned on SQL generation often suffer from catastrophic forgetting, which lose the versatility of procedural reasoning and pertaining to generation constraints. Inspired by the usage of high-resource programming languages as LLM reasoning intermediaries, we propose CORES model, which leverages Python as a procedural reasoning pivot to enhance both complex SQL generation and tabular reasoning. It decomposes complex queries into Python reasoning traces before generating the final SQL, which bridges the gap between procedural reasoning and declarative expression. In order to internalize this reasoning capability, we fine-tune LLMs via GRPO with tailored process reward functions that mitigate the sparse feedback problem. We experimentally verify the effectiveness of CORES on six text-to-SQL benchmarks, where ours outperforms baselines by 6.44% on average, while maintains good capability on three tableQA benchmarks.

Tristan Williams, Franziska Weeber, Sebastian Padó, Alan Akbik

Large language models are increasingly used to represent human opinions, values, or beliefs, and their steerability towards these ideals is an active area of research. Existing work focuses predominantly on aligning marginal response distributions, treating each alignment evaluation example independently. While essential, this may overlook deeper latent structures that characterise real populations and underpin cultural values theories. We propose a framework for evaluating the representativeness of aligned models through multivariate correlation patterns in addition to marginal distributions. We show the value of our evaluation scheme by comparing two model steering techniques (persona prompting and demographic fine-tuning) and evaluating them against human responses from the World Values Survey. While the demographic fine-tuned model better approximates marginal response distributions, persona prompting performs marginally better at reproducing the empirical correlation structure between survey items. Despite this reversal, neither technique aligns with human correlation patterns. We conclude that representativeness is a distinct aspect of value alignment and an evaluation focused on marginals can mask structural failures, leading to overly optimistic conclusions about model representativeness.

Jiaming Zhou, Xuxin Cheng, Shiwan Zhao, Yuhang Jia, Cao Liu, Ke Zeng, Xunliang Cai, Yong Qin

Autoregressive (AR) large audio language models (LALMs) such as Qwen-2.5-Omni have achieved strong performance on audio understanding and interaction, but scaling them remains costly in data and computation, and strictly sequential decoding limits inference efficiency. Diffusion large language models (dLLMs) have recently been shown to make effective use of limited training data, and prior work on DIFFA indicates that replacing an AR backbone with a diffusion counterpart can substantially improve audio understanding under matched settings, albeit at a proof-of-concept scale without large-scale instruction tuning, preference alignment, or practical decoding schemes. We introduce DIFFA-2, a practical diffusion-based LALM for general audio understanding. DIFFA-2 upgrades the speech encoder, employs dual semantic and acoustic adapters, and is trained with a four-stage curriculum that combines semantic and acoustic alignment, large-scale supervised fine-tuning, and variance-reduced preference optimization, using only fully open-source corpora. Experiments on MMSU, MMAU, and MMAR show that DIFFA-2 consistently improves over DIFFA and is competitive to strong AR LALMs under practical training budgets, supporting diffusion-based modeling is a viable backbone for large-scale audio understanding.

Basel Mousi, Fahim Dalvi, Shammur Absar Chowdhury, Firoj Alam, Nadir Durrani

Vision–language models (VLMs) can achieve high accuracy while still accepting **culturally plausible but visually incorrect** interpretations. Existing hallucination benchmarks rarely test this failure mode, particularly outside Western contexts and English. We introduce **M^2CQA**, a culturally grounded multimodal benchmark built from images spanning 17 MENA countries, paired with contrastive true and counterfactual statements in English, Arabic, and its dialects. To isolate hallucination beyond raw accuracy, we propose the **CounterFactual Hallucination Rate (CFHR)**, which measures counterfactual acceptance conditioned on correctly answering the true statement. Evaluating state-of-the-art VLMs under multiple prompting strategies, we find that CFHR rises sharply in Arabic, especially in dialects, even when true-statement accuracy remains high.Moreover, reasoning-first prompting consistently increases counterfactual hallucination, while answering before justifying improves robustness. We make the dataset publicly available for the community (https://huggingface.co/datasets/QCRI/M2CQA)).

Yehua Lin, Liping Zheng, Yin Chen

Large language models (LLMs) face challenges in logical reasoning where correctness requires strict deductive procedures. Purely model-based approaches often suffer from hallucinations, while neuro-symbolic methods typically delegate deduction to external solvers, reducing the LLM to a mere translator. To address this, we propose MAC-Reasoner, a multi-agent framework that constructs a Logic-Augmented Context to enhance LLMs’ reasoning. In this framework, a translator agent converts problems into executable symbolic programs. Symbolic information from solver execution is transformed into the Logic-Augmented Context, serving as a verification reference where logical conflicts trigger heightened attention to violated constraints. We evaluate MAC-Reasoner with three backbone LLMs on four challenging benchmarks. Results show consistent and robust improvements over baselines. Furthermore, reasoning traces from MAC-Reasoner can be used for supervised fine-tuning of LLMs to achieve more accurate and efficient logical reasoning.

Yanting Li, Zhuoyang Jiang, Enyan Dai, Lei Wang, Wen-Cai Ye, Li Liu

Goal-directed molecular generation requires satisfying heterogeneous constraints such as protein–ligand compatibility and multi-objective drug-like properties, yet existing methods often optimize these constraints in isolation, failing to reconcile conflicting objectives (e.g., affinity vs. safety), and struggle to navigate the non-differentiable chemical space without compromising structural validity. To address these challenges, we propose CAGenMol, a condition-aware discrete diffusion framework over molecular sequences that formulates molecular design as conditional denoising guided by heterogeneous structural and property signals. By coupling discrete diffusion with reinforcement learning, the model aligns the generation trajectory with non-differentiable objectives while preserving chemical validity and diversity. The non-autoregressive nature of diffusion language model further enables iterative refinement of molecular fragments at inference time. Experiments on structure-conditioned, property-conditioned, and dual-conditioned benchmarks demonstrate consistent improvements over state-of-the-art methods in binding affinity, drug-likeness, and success rate, highlighting the effectiveness of our framework. The code is available at https://github.com/Lee612-1/CAGenMol.

Taolin Zhang, Haidong Kang, Dongyang Li, Qizhou Chen, Xiaofeng He, Chengyu Wang, Richang Hong

Recently, large language models (LLMs) have demonstrated impressive performance but still suffer from hallucinations. Model editing has been proposed as a means to correct factual inaccuracies. A challenging scenario is sequential model editing (SME), which aims to rectify errors continuously, rather than a one-time task. During SME, the general capabilities of LLMs can be negatively affected due to the introduction of new parameters. In this paper, we propose a queue-based self-correction framework, QueueEDIT, that not only enhances SME performance by addressing long-sequence dependencies but also mitigates the impact of parameter bias on the general capabilities of LLMs. Specifically, we first introduce a structural mapping editing loss to map editing triplets to knowledge-sensitive neurons within the Transformer layers. We then store the located parameters for each piece of edited knowledge in a queue and dynamically align previously edited parameters. At each edit, we select parameters in the queue that are most relevant to currently located parameters to determine whether knowledge associated with previous edits requires realignment. Irrelevant parameters in the queue are frozen, and we update the parameters at the queue head into the LLM to ensure they do not harm general capabilities. Experiments show that QueueEDIT significantly outperforms strong baselines across various SME settings, while maintaining competitive performance in single-turn editing. Resulting LLMs also preserve high performance on general NLP tasks throughout the SME process.

Yubo Hou, Zhisheng Chen, Tao Wan, Zengchang Qin

The stateless architecture of Large Language Models inherently lacks the mechanism to preserve dynamic context, compelling agents to redundantly reprocess history to maintain long-horizon autonomy. While latent memory offers a solution, current approaches are hindered by architectural segregation, relying on auxiliary encoders that decouple memory from the reasoning backbone. We propose FlashMem, a framework that distills intrinsic memory directly from transient reasoning states via computation reuse. Leveraging the property that internal representations uniquely encode input trajectories, FlashMem identifies the last hidden state as a sufficient statistic for the interaction history. This enables a Shared-KV Consolidator to synthesize memory by attending directly to the backbone’s frozen cache, eliminating redundant re-parameterization. Furthermore, a parameter-free Cognitive Monitor leverages attention entropy to adaptively trigger consolidation only when high epistemic uncertainty is detected. Experiments demonstrate that FlashMem matches the performance of heavy baselines while reducing inference latency by 5 times, effectively bridging the gap between efficiency and persistent cognition.

Yu Li, Sizhe Tang, Tian Lan

Reinforcement learning for Large Language Model agents is often hindered by sparse rewards in multi-step reasoning tasks. Existing approaches like Group Relative Policy Optimization treat sampled trajectories as independent chains, assigning uniform credit to all steps in each chain and ignoring the existence of critical steps that may disproportionally impact reasoning outcome. In this paper, we propose T-STAR(Tree-structured Self-Taught Agent Rectification), a framework that recovers the latent correlated reward structure across seemingly independent trajectories. Specifically, we consolidate trajectories into a unified Cognitive Tree by identifying and merging functionally similar steps/nodes. It enables an Introspective Valuation mechanism that back-propagates trajectory-level rewards through the tree to obtain a new notion of variance-reduced relative advantage at step-level. Using the Cognitive Tree, we also develop In-Context Thought Grafting to synthesize corrective reasoning by contrasting successful and failed branches at critical divergence points/steps. Our proposed Surgical Policy Optimization then capitalizes on the rich policy gradient information concentrated at these critical points/steps through a Bradley-Terry type of surgical loss. Extensive experiments across embodied, interactive, reasoning, and planning benchmarks demonstrate that T-STAR achieves consistent improvements over strong baselines, with gains most pronounced on tasks requiring extended reasoning chains.

Gengyang Li, Wang Cai, Yifeng Gao, Yunfang Wu

Chain-of-Thought (CoT) prompting improves reasoning but often produces long and redundant traces that substantially increase inference cost. We present SyncThink, a training-free and plug-and-play decoding method that reduces CoT overhead without modifying model weights. We find that answer tokens attend weakly to early reasoning and focus on ‘</think>‘, indicating an information bottleneck.Building on this observation, SyncThink monitors the model’s own reasoning-transition signal and terminates reasoning. Experiments on GSM8K, MMLU, GPQA, and BBH across three DeepSeek-R1 distilled models show that SyncThink achieves 62.00% average Top@1 accuracy using 656 generated tokens and 28.68s latency, compared to 61.22%, 2141 tokens, and 92.01s for full CoT decoding. On long-horizon tasks such as GPQA, SyncThink can further yield up to +8.1 absolute accuracy by preventing over-thinking.

Dabin Fu, Fanghong Zhang

Named Entity Recognition (NER) plays a fundamental role in information extraction and domain knowledge construction. However, in specialized domains such as wind power fault diagnosis, the scarcity of labeled data makes supervised approaches impractical. Zero-shot NER provides a promising alternative but still struggles with incomplete entity detection and unstable generation boundaries. To address these challenges, we propose ReCoT-NER, a reasoning-enhanced generative framework that integrates Chain-of-Thought (CoT) prompting and recall-oriented loss optimization. The proposed CoT instruction design explicitly decomposes NER into two reasoning stages: entity span detection and entity type classification. This enables the model to follow a structured inference process. In addition, we introduce a recall-oriented loss function that reweights entity and non-entity tokens to mitigate false negatives, encouraging more inclusive entity coverage. Experiments on CrossNER, MIT, and a newly constructed wind-power NER dataset demonstrate that ReCoT-NER consistently improves recall and overall F1 performance across both general and industrial domains. Notably, ReCoT-NER achieves competitive results with just a 77M-parameter model, making it well-suited for low-resource zero-shot settings. The code for our method is publicly available at https://github.com/10637409100/RECOTNER.

Wencheng Ye, Xiaoyang Yuan, Yi Bin, Hengyu Jin, Liang Peng, Pengpeng Zeng, Heng Tao Shen

Recent work on domain-specific reasoning with large language models (LLMs) has largely relied on training-intensive approaches that require updating model parameters. Although activation steering has emerged as a parameter-efficient alternative, existing methods typically rely on static and manually designed interventions, limiting their ability to adapt to the dynamic nature of complex reasoning. To address this limitation, we propose RISER (Router-based Intervention for Steerable Enhancement of Reasoning), a plug-and-play intervention framework that adaptively steers LLM reasoning in activation space. RISER builds a library of reusable reasoning vectors and employs a lightweight Router to dynamically compose these vectors for each input. The Router is optimized via reinforcement learning under task-level rewards, enabling the emergent and compositional activation of latent cognitive primitives. Across seven diverse benchmarks, RISER achieves average zero-shot accuracy improvements of 3.4–6.5% over the base model, while outperforming chain-of-thought-style reasoning with 2–3× higher token efficiency and robust accuracy gains. Further analysis demonstrates that RISER autonomously combines multiple vectors into interpretable and precise control strategies, pointing toward more controllable and efficient LLM reasoning.

Elisa Leonardelli, Camilla Casula, Boglarka Nyul, Sara Tonelli

Large language models (LLMs) are known to exhibit gender bias, yet most evaluations focus on downstream stereotypes rather than the normative frameworks that shape model inference. We investigate whether LLMs rely on traditional masculinity norms (e.g. "real men are tough") as latent priors in gender-biased inference. We ground our evaluation in the Male Role Norms Inventory (MRNI), a validated psychological framework of prescriptive male role norms.Anchored in MRNI items, we probe models using two complementary approaches: (i) explicit Likert-style agreement with masculinity norms, and (ii) a newly crafted English-Italian scenario-based inference dataset (MRNI-BB), in which gender information and evidential support are systematically varied. Across models, explicit endorsement of masculinity norms is generally low. In contrast, in scenario-based inference tasks, models systematically attribute MRNI-aligned behaviors to male agents, even when evidence is ambiguous or absent. This effect disappears when gender markers are removed, suggesting that masculinity norms are treated as gender-specific expectations about male agents. Increasing model scale reduces explicit norm endorsement but is associated with stronger male-directed bias under uncertainty.

H S V N S Kowndinya Renduchintala, Sumit Bhatia

Large Language Models (LLMs) exhibit a puzzling disparity in their formal linguistic competence: while they learn some linguistic phenomena with near-perfect mastery, they often perform below chance on others, even after training on trillions of tokens. In this work, we investigate whether these failures stem from inherent architectural limitations or simply the scarcity of these specific grammatical constructions in web-scale corpora. We pre-train simple GPT-2 Small (124M) models on a 100M-token random sample of the FineWeb corpus and intervene by injecting a minimal amount (1%) of synthetic data targeting specific linguistic phenomena. We find that this targeted intervention substantially improves model performance in 8 out of the 9 worst-performing BLiMP paradigms – notably the accuracy on a specific paradigm, only_npi_scope, surges from 20.9% to 69.4%. Furthermore, we observe that these interventions generally preserve or slightly improve aggregate performance. However, while we also identify a resistant phenomenon, principle_A_c_command, whose performance remains below chance even after our data augmentation, our findings do serve as an optimistic existence proof that even small language models can substantially improve on those linguistic phenomena on which models typically perform poorly, provided the pre-training data contains sufficient exposure to them. This suggests that efforts towards human-scale language modeling may benefit greatly by focusing on data composition. The code to reproduce our results is open-sourced at https://github.com/kowndinya-renduchintala/heterogeneity-in-formal-linguistic-competence.