How does the natural evolution of context paragraphs affect Question Answering (QA) in generative Large Language Models (LLMs)? To address this, we propose a framework for curating naturally evolved, human-edited variants of reading passages from contemporary QA benchmarks and for analysing LLM performance across a range of semantic similarity scores, which quantify how closely each variant aligns with Wikipedia content on the same article topic that the LLM saw during pretraining. Using this framework, we evaluate 6 QA datasets and 8 LLMs with publicly available training data. Our experiments reveal that LLM performance declines as reading passages naturally diverge from the versions encountered during pretraining–even when the question and all necessary information remains present at inference time. For instance, average accuracy on BoolQ drops by over 30% from the highest to lowest similarity bins. This finding suggests that natural text evolution may pose a significant challenge to the language understanding capabilities of fully open-source LLMs.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Recently, multimodal large language models (MLLMs) have demonstrated remarkable performance in visual-language tasks. However, the authenticity of the responses generated by MLLMs is often compromised by object hallucinations. We identify that a key cause of these hallucinations is the model’s over-susceptibility to image frequency features in detecting objects. In this paper, we introduce Multi-Frequency Perturbations (MFP), a simple, cost-effective, and pluggable adversarial training method that leverages both low-frequency and high-frequency features of images to perturb visual feature representations and explicitly suppress redundant frequency-domain features during inference, thereby mitigating hallucinations. Experimental results demonstrate that our method significantly mitigates object hallucinations across various model architectures. Furthermore, as a training-time method, MFP can be combined with inference-time methods to achieve state-of-the-art performance on the CHAIR benchmark.
The wide adoption of Large Language Models (LLMs) has attracted significant attention from jailbreak attacks, where adversarial prompts crafted through optimization or manual design exploit LLMs to generate malicious contents. However, optimization-based attacks have limited efficiency and transferability, while existing manual designs are either easily detectable or demand intricate interactions with LLMs. In this paper, we first point out a novel perspective for jailbreak attacks: LLMs are more responsive to positive prompts. Based on this, we deploy Happy Ending Attack (HEA) to wrap up a malicious request in a scenario template involving a positive prompt formed mainly via a happy \ ending, it thus fools LLMs into jailbreaking either immediately or at a follow-up malicious request. This has made HEA both efficient and effective, as it requires only up to two turns to fully jailbreak LLMs. Extensive experiments show that our HEA can successfully jailbreak on state-of-the-art LLMs, including GPT-4o, Llama3-70b, Gemini-pro, and achieves 88.79% attack success rate on average. We also provide quantitative explanations for the success of HEA.
English as Defense Proxy: Mitigating Multilingual Jailbreak via Eliciting English Safety Knowledge
PDF ↗Large language models (LLMs) excel in many tasks, but their safety guarantees vary by language, e.g., responses in English tend to be safer than those in low-resource languages. This inconsistency creates a vulnerability, since an attacker can circumvent safety measures by using a less-supported language as an intermediary, even without fluency in that language. Traditional solutions rely on multilingual safety alignment, which demands vast, per-language datasets and introduces significant trade-offs between usefulness and safety (the so-called “alignment tax”). To overcome these limitations, we introduce English as Defense Proxy (E-Proxy), a unified approach that leverages English, usually the advantage language of LLMs, as a universal safety anchor. During multilingual training, E-Proxy uses English jailbreak prompts to extract the model’s existing safety knowledge, then applies simple language-mapping prompts (e.g., “Please answer in target language”) to transfer that knowledge across languages. Our analysis shows that formulating prompts in a high-resource language preserves the model’s utility, while enforcing responses in the target language significantly enhances safety. We evaluate E-Proxy on extensive benchmarks of both attack resistance and task performance. On the MultiJail benchmark, E-Proxy blocks over 99 % of jailbreak attempts while retaining 95 % of average task performance, all with a simply constructed multilingual alignment data.
LAGCL4Rec: When LLMs Activate Interactions Potential in Graph Contrastive Learning for Recommendation
PDF ↗A core barrier preventing recommender systems from reaching their full potential lies in the inherent limitations of user-item interaction data: (1) Sparse user-item interactions, making it difficult to learn reliable user preferences; (2) Traditional contrastive learning methods often treat negative samples as equally hard or easy, ignoring the informative semantic difficulty during training. (3) Modern LLM-based recommender systems, on the other hand, discard all negative feedback, leading to unbalanced preference modeling. To address these issues, we propose LAGCL4Rec, a framework leveraging Large Language Models to Activate interactions in Graph Contrastive Learning for Recommendation. Our approach operates through three stages: (i) Data-Level: augmenting sparse interactions with balanced positive and negative samples using LLM-enriched profiles; (ii) Rank-Level: assessing semantic difficulty of negative samples through LLM-based grouping for fine-grained contrastive learning; and (iii) Rerank-Level: reasoning over augmented historical interactions for personalized recommendations. Theoretical analysis proves that LAGCL4Rec achieves effective information utilization with minimal computational overhead. Experiments across multiple benchmarks confirm our method consistently outperforms state-of-the-art baselines.
Most work on Knowledge Graph (KG) verbalisation is monolingual leaving open the question of how to scale KG-to-Text generation to languages with varying amounts of resources. In this work, we explore KG-to-Text generation on nine languages including five high-resource (HR) languages (English, Chinese, French, Spanish, Russian) and four low-resource (LR) languages (Breton, Irish, Maltese, Welsh). We first construct silver multilingual training data for all nine languages and new gold out-of-domain test data for the five HR languages. Using this data and already available in-domain test sets for 7 of our 9 languages, we then compare three strategies: (1) NLG+MT—a state-of-the-art KG-to-English model followed by Machine Translation (MT) into the target language; (2) FTMT—multilingual MT models fine-tuned end-to-end on the silver data; and (3) FewShot—few-shot LLM prompting comparing 4 LLMs. We explore different prompting strategies and show that our best prompting strategy performs the best on all 9 languages, discussing the relative performance of the three approaches on Low vs High Resource languages and on in- vs out-of-domain data.The models, the test set, and the silver training data are available at https://github.com/MeloS7/Multilingual-KG-Verbalisation.
Despite the significant advancements of self-play fine-tuning (SPIN), which can transform a weak large language model (LLM) into a strong one through competitive interactions between models of varying capabilities, it still faces challenges in the Text-to-SQL task. SPIN does not generate new information, and the large number of correct SQL queries produced by the opponent model during self-play reduces the main model’s ability to generate accurate SQL queries. To address this challenge, we propose a new self-play fine-tuning method tailored for the Text-to-SQL task, called SPFT-SQL. Prior to self-play, we introduce a verification-based iterative fine-tuning approach, which synthesizes high-quality fine-tuning data iteratively based on the database schema and validation feedback to enhance model performance, while building a model base with varying capabilities. During the self-play fine-tuning phase, we propose an error-driven loss method that incentivizes incorrect outputs from the opponent model, enabling the main model to distinguish between correct SQL and erroneous SQL generated by the opponent model, thereby improving its ability to generate correct SQL. Extensive experiments and in-depth analyses on six open-source LLMs and five widely used benchmarks demonstrate that our approach outperforms existing state-of-the-art (SOTA) methods.
This paper studies the problem of time series forecasting, which aims to generate future predictions given historical trajectories. Recent researchers have applied large language models (LLMs) into time series forecasting, which usually align the time series space with textual space and output future predictions with strong autoregressive reasoning abilities. Despite their remarkable progress, these approaches usually lack an understanding of holistic temporal patterns with potential error accumulation. Towards this end, this paper proposes a simple yet effective framework that marries \underline{\text{L}}arg\underline{\text{e}} Langu\underline{\text{a}}ge Diffusion Model with time series \underline{\text{f}}orecasting (LEAF). The core of our framework is to generate future predictions with a diffusion model from a holistic view. In particular, we first introduce a tokenization module to convert time series into tokens and then adopt the language diffusion models to capture the temporal dependencies. In this way, we can transform masked time series into all the predictions with the remasking strategy. Extensive experiments on various benchmark datasets validate the effectiveness of the proposed LEAF in comparison to various baselines.
Unearthing Gems from Stones: Policy Optimization with Negative Sample Augmentation for LLM Reasoning
PDF ↗Recent advances in reasoning language models have witnessed a paradigm shift from short to long CoT pattern. Given the substantial computational cost of rollouts in long CoT models, maximizing the utility of fixed training datasets becomes crucial. Our analysis reveals that negative responses contain valuable components such as self-reflection and error-correction steps, yet primary existing methods either completely discard negative samples (RFT) or apply equal penalization across all tokens (RL), failing to leverage these potential learning signals. In light of this, we propose Behavior Constrained Policy Gradient with Negative Sample Augmentation (BCPG-NSA), a fine-grained offline RL framework that encompasses three stages: 1) sample segmentation, 2) consensus-based step correctness assessment combining LLM and PRM judgers, and 3) policy optimization with NSA designed to effectively mine positive steps within negative samples. Experimental results show that BCPG-NSA outperforms baselines on several challenging math/coding reasoning benchmarks using the same training dataset, achieving improved sample efficiency and demonstrating robustness and scalability when extended to multiple iterations.
GTA: Supervised-Guided Reinforcement Learning for Text Classification with Large Language Models
PDF ↗In natural language processing (NLP) tasks, pure reinforcement learning fine-tuning methods often suffer from inefficient exploration and slow convergence; while supervised fine-tuning (SFT) methods, although efficient in training, have limited performance ceiling and less solid theoretical foundation compared to reinforcement learning. To address efficiency-capability trade-off, we propose the Guess-Think-Answer (GTA) framework that combines the efficiency of SFT with the capability gains of RL in a unified training paradigm. GTA works by having the model first produce a provisional guess (optimized via cross-entropy loss), then reflect on this guess before generating the final answer, with RL rewards shaping both the final output and the format of the entire GTA structure. This hybrid approach achieves both faster convergence than pure RL and higher performance ceiling than pure SFT. To mitigate gradient conflicts between the two training signals, we employ loss masking and gradient constraints. Empirical results on three text classification benchmarks demonstrate that GTA substantially accelerates convergence while outperforming both standalone SFT and RL baselines.
This paper investigates compositionality in chemical language models (ChemLLMs). We introduce STEPS, a benchmark with compositional questions that reflect intricate chemical structures and reactions, to evaluate models’ understanding of chemical language. Our approach focuses on identifying and analyzing compositional patterns within chemical data, allowing us to evaluate how well existing LLMs can handle complex queries. Experiments with state-of-the-art ChemLLMs show significant performance drops in compositional tasks, highlighting the need for models that move beyond pattern recognition. By creating and sharing this benchmark, we aim to enhance the development of more capable chemical LLMs and provide a resource for future research on compositionality in chemical understanding.
Multifaceted Evaluation of Audio-Visual Capability for MLLMs: Effectiveness, Efficiency, Generalizability and Robustness
PDF ↗Multi-modal large language models (MLLMs) have recently achieved great success in processing and understanding information from diverse modalities (e.g., text, audio, and visual signals). Despite their growing popularity, there remains a lack of comprehensive evaluation measuring the audio-visual capabilities of these models, especially in diverse scenarios (e.g., distribution shifts and adversarial attacks). In this paper, we present a multifaceted evaluation of the audio-visual capability of MLLMs, focusing on four key dimensions: effectiveness, efficiency, generalizability, and robustness. Through extensive experiments, we find that MLLMs exhibit strong zero-shot and few-shot generalization abilities, enabling them to achieve great performance with limited data. However, their success relies heavily on the vision modality, which impairs performance when visual input is corrupted or missing. Additionally, while MLLMs are susceptible to adversarial samples, they demonstrate greater robustness compared to traditional models. The experimental results and our observations provide new insights into the audio-visual capabilities of MLLMs, highlighting areas for improvement and offering guidance for future research.
Aspect-Based Sentiment Analysis (ABSA) focuses on extracting opinions about specific aspects, with Aspect Sentiment Quad Prediction (ASQP) being the most complex sub-task. Large language models (LLMs) like GPT4 exhibit strong generalization yet struggle with ASQP due to a lack of task-specific alignment. Supervised small language models (SLMs), while effective in capturing task-specific patterns, lack the extensive knowledge of LLMs. To address this, we propose a framework that combines SLMs and LLMs using supervised in-context learning to align LLM outputs with human preferences. One SLM is supervised to generate candidate answers and guide LLMs with task-specific instructions, while another SLM acts as a reward model iteratively evaluates and refines LLM outputs. Experiments show that our framework significantly improves ASQP performance, demonstrating robustness, scalability, and potential for advancing alignment techniques in sentiment analysis.
Optimizing Cross-Client Domain Coverage for Federated Instruction Tuning of Large Language Models
PDF ↗Federated domain-specific instruction tuning (FedDIT) for large language models (LLMs) aims to enhance performance in specialized domains using distributed private and limited data, yet identifying key performance drivers and optimal augmentation strategies remains challenging. We empirically establish that cross-client domain coverage, rather than data heterogeneity, is the pivotal factor. We then introduce FedDCA, an algorithm that explicitly maximizes this coverage through diversity-oriented client center selection and retrieval-based augmentation, constructing diverse, non-redundant cross-client instruction sets. Extensive experiments across multiple domains demonstrate FedDCA’s superiority over eleven baselines, achieving performance gains of up to 29.19% and domain coverage improvements of 4.82%-21.36%. FedDCA maintains its effectiveness in diverse and challenging scenarios, including data selection, held-out settings where task-specific public data is scarce and various data heterogeneity, with manageable privacy risks. This work clarifies critical FedDIT dynamics and presents FedDCA as an effective, privacy-preserving, and scalable solution for advancing domain-specific LLM tuning.
DongbaMIE: A Multimodal Information Extraction Dataset for Evaluating Semantic Understanding of Dongba Pictograms
PDF ↗Dongba pictographic is the only pictographic script still in use in the world. Its pictorial ideographic features carry rich cultural and contextual information. However, due to the lack of relevant datasets, research on semantic understanding of Dongba hieroglyphs has progressed slowly. To this end, we constructed DongbaMIE - the first dataset focusing on multimodal information extraction of Dongba pictographs. The dataset consists of images of Dongba hieroglyphic characters and their corresponding semantic annotations in Chinese. It contains 23,530 sentence-level and 2,539 paragraph-level high-quality text-image pairs. The annotations cover four semantic dimensions: object, action, relation and attribute. Systematic evaluation of mainstream multimodal large language models shows that the models are difficult to perform information extraction of Dongba hieroglyphs efficiently under zero-shot and few-shot learning. Although supervised fine-tuning can improve the performance, accurate extraction of complex semantics is still a great challenge at present.
Despite their impressive performance in coarse-grained video understanding, Video Large Language Models (Video-LLMs) still face challenges in fine-grained temporal grounding, including ineffective temporal modeling and inadequate timestamp representations. In this work, we introduce Grounded-VideoLLM, a novel Video-LLM designed to perceive and reason over specific video moments with fine-grained temporal precision. Our model features (1) a two-stream encoder that explicitly captures inter-frame relationships while preserving intra-frame visual details and (2) discrete temporal tokens enriched with structured time knowledge for timestamp representation. Besides, we propose a multi-stage training strategy tailored to such grounding-specific architecture. The model is initially trained on simple video-caption tasks and progressively introduced to complex video temporal grounding tasks, ensuring a smooth learning curve and temporal alignment. We further strengthen Grounded-VideoLLM’s temporal reasoning by constructing a VideoQA dataset with grounded information using an automated annotation pipeline. Extensive experiments demonstrate that Grounded-VideoLLM not only surpasses existing models in fine-grained grounding tasks but also exhibits strong potential as a general video understanding assistant.
Text-to-image synthesis has made remarkable progress, yet accurately interpreting complex and lengthy prompts remains challenging, often resulting in semantic inconsistencies and missing details. Existing solutions, such as fine-tuning, are model-specific and require training, while prior automatic prompt optimization (APO) approaches typically lack systematic error analysis and refinement strategies, resulting in limited reliability and effectiveness. Meanwhile, test-time scaling methods operate on fixed prompts and on noise or sample numbers, limiting their interpretability and adaptability. To solve these, we introduce a flexible and efficient test-time prompt optimization strategy that operates directly on the input text. We propose a plug-and-play multi-agent system called GenPilot, integrating error analysis, clustering-based adaptive exploration, fine-grained verification, and a memory module for iterative optimization. Our approach is model-agnostic, interpretable, and well-suited for handling long and complex prompts. Simultaneously, we summarize the common patterns of errors and the refinement strategy, offering more experience and encouraging further exploration. Experiments on DPG-bench and Geneval with improvements of up to 16.9% and 5.7% demonstrate the strong capability of our methods in enhancing the text and image consistency and structural coherence of generated images, revealing the effectiveness of our test-time prompt optimization strategy. The code is available at https://github.com/27yw/GenPilot.
THCM-CAL: Temporal-Hierarchical Causal Modelling with Conformal Calibration for Clinical Risk Prediction
PDF ↗Automated clinical risk prediction from electronic health records (EHRs) demands modeling both structured diagnostic codes and unstructured narrative notes. However, most prior approaches either handle these modalities separately or rely on simplistic fusion strategies that ignore the directional, hierarchical causal interactions by which narrative observations precipitate diagnoses and propagate risk across admissions. In this paper, we propose **THCM-CAL**, a Temporal-Hierarchical Causal Model with Conformal Calibration. Our framework constructs a multimodal causal graph where nodes represent clinical entities from two modalities: textual propositions extracted from notes and ICD codes mapped to textual descriptions. Through hierarchical causal discovery, **THCM-CAL** infers three clinically grounded interactions: intra-slice same-modality sequencing, intra-slice cross-modality triggers, and inter-slice risk propagation. To enhance prediction reliability, we extend conformal prediction to multi-label ICD coding, calibrating per-code confidence intervals under complex co-occurrences. Experimental results on MIMIC-III and MIMIC-IV demonstrate the superiority of **THCM-CAL**.
DivLogicEval: A Framework for Benchmarking Logical Reasoning Evaluation in Large Language Models
PDF ↗Logic reasoning in natural language has been recognized as an important measure of human intelligence for Large Language Models (LLMs). Popular benchmarks may entangle multiple reasoning skills and thus provide unfaithful evaluations on the logic reasoning skill. Meanwhile, existing logic reasoning benchmarks are limited in language diversity and their distributions are deviated from the distribution of an ideal logic reasoning benchmark, which may lead to biased evaluation results. This paper thereby proposes a new classical logic benchmark DivLogicEval, consisting of natural sentences composed of diverse statements in a counterintuitive way. To ensure a more reliable evaluation, we also introduce a new evaluation metric that mitigates the influence of bias and randomness inherent in LLMs. Through experiments, we demonstrate the extent to which logical reasoning is required to answer the questions in DivLogicEval and compare the performance of different popular LLMs in conducting logical reasoning.
Search engines play a central role in how people gather information, but subtle cues like headline framing may influence not only what users believe but also how they search. While framing effects on judgment are well documented, their impact on subsequent search behavior is less understood. We conducted a controlled experiment where participants issued queries and selected from headlines filtered by specific linguistic frames. Headline framing significantly shaped follow-up queries: conflict and strategy frames disrupted alignment with prior selections, while episodic frames led to more concrete queries than thematic ones. We also observed modest short-term frame persistence that declined over time. These results suggest that even brief exposure to framing can meaningfully alter the direction of users’ information-seeking behavior.