Civil judicial cases are highly complicated, posing significant challenges for Large Language Models (LLMs) for Legal Judgment Prediction (LJP).While judges manage this complexity through the dispute focus-a mechanism distilling cases into core issues-existing research largely overlooks this tool in favor of generic reasoning frameworks that lack authentic judicial logic.To bridge this gap, we first introduce FocalLaw, the first dataset aligning fullprocess Chinese civil judicial data through the dispute focus, comprising 1,000 high-quality cases across six causes of action.Building on this dataset, we examine LLMs' capability to utilize the dispute focus and uncover a counter-intuitive phenomenon: LLMs fail to leverage the dispute focus even with CoT and SFT, which we identify as the "Clerk Trap".To solve the problem, we propose FocalJudge, a novel framework that leverages the dispute focus to guide LLMs through a structured, judgelike cognitive workflow.Experimental results demonstrate the effectiveness of FocalJudge and offer valuable insights into the interpretability and reliability of LLMs in the legal domain.

Multimodal Large Language Models (MLLMs) face critical privacy challenges arising from the indiscriminate memorization of sensitive data.Existing unlearning methods often fail to precisely disentangle specific instances from general concepts, leading to either catastrophic forgetting of useful knowledge or unsafe content substitution.We attribute these failures to a fundamental geometric mismatch: these approaches primarily operate in Euclidean space, which lacks the capacity to model the hierarchical entailment inherent in visual-linguistic concepts.To address this, we introduce LOTUS (LOrentz Transport for Unlearning Strategies), a framework that performs surgical semantic pruning within the Lorentz manifold.LOTUS employs an Inverted Entailment Cone Loss to sever the semantic inheritance of sensitive concepts and a Lorentz Transport mechanism to align pruned features with a safety refusal prior in the tangent space.Extensive experiments on MLLMU-Bench demonstrate that LOTUS significantly outperforms baselines, improving unlearning efficacy by over 9% on LLaVA compared to state-of-the-art constraint-based methods.Crucially, LOTUS achieves this precision while maintaining general utility, effectively resolving the dilemma between thorough erasure and model stability.

Human cognition excels at extending knowledge through analogy, where word meanings evolve along structured pathways from concrete prototypes to abstract senses via metaphor and metonymy.Do Large Language Models (LLMs) internalize this generative logic, or merely mimic statistical patterns?To investigate this, we introduce CogEvolve, a cognitive linguistic benchmark designed to test these evolutionary pathways across textual and visual modalities.Our evaluation reveals a distinct cognitive profile: models function as "Super-Associators" expert at static recognition yet fail at causal reasoning.In text, they exhibit a Frequency-Primacy Conflation, confusing statistical prevalence with cognitive basicness.Crucially, this reasoning collapses further in the visual domain.We term this deficit the Ungrounded Arrow: models possess high-fidelity concept representations (the "dots") but lack the transformational operators (the "arrows") essential for true relational understanding 1 .

Bragging, as a pervasive social-linguistic phenomenon, reflects complex human interaction patterns.However, the understanding and generation of appropriate bragging behavior in large language models (LLMs) remains underexplored.In this paper, we propose a comprehensive study that combines analytical and controllable approaches to examine bragging in LLMs.We design three tasks, bragging recognition, bragging explanation, and bragging generation, along with novel evaluation metrics to assess the models' ability to identify bragging intent, social appropriateness, and account for context sensitivity.Our analysis reveals the challenges of bragging in the social context, such as recognizing bragging and responding appropriately with bragging in conversation.This work provides new insights into how LLMs process bragging and highlights the need for more research on generating contextually appropriate behavior in LLMs 1 .

Implicit hate speech has become a significant challenge for online platforms, as it often avoids detection by large language models (LLMs) due to its indirectly expressed hateful intent.This study identifies the limitations of LLMs in detecting implicit hate speech, particularly when disguised as seemingly harmless expressions in a rhetorical device.To address this challenge, we employ a Jailbreaking strategy and Energy-based Constrained Decoding techniques, and design a small model for measuring the energy of metaphorical rhetoric.This approach can lead to LLMs generating metaphorical implicit hate speech.Our research reveals that advanced LLMs, like GPT-4o, frequently misinterpret metaphorical implicit hate speech, and fail to prevent its propagation effectively.Even specialized models, like ShieldGemma and LlamaGuard, demonstrate inadequacies in blocking such content, often misclassifying it as harmless speech.This work points out the vulnerability of current LLMs to implicit hate speech, and emphasizes the improvements to address hate speech threats better.

In the realm of artificial intelligence and linguistics, the automatic generation of humor, particularly puns, remains a complex task. This paper introduces an innovative approach that employs a Generative Adversarial Network (GAN) and semantic pruning techniques to generate humorous puns. We initiate our process by identifying potential pun candidates via semantic pruning. This is followed by the use of contrastive learning to decode the unique characteristics of puns, emphasizing both correct and incorrect interpretations. The learned features from contrastive learning are utilized within our GAN model to better capture the semantic nuances of puns. Specifically, the generator exploits the pruned semantic tree to generate pun texts, while the discriminator evaluates the generated puns, ensuring both linguistic correctness and humor. Evaluation results highlight our model’s capacity to produce semantically coherent and humorous puns, demonstrating an enhancement over prior methods and approach human-level performance. This work contributes significantly to the field of computational humor, advancing the capabilities of automatic pun generation.

By integrating various modules with the Visual Transformer (ViT), we facilitate a interpretation of image processing across each layer and attention head. This method allows us to explore the connections both within and across the layers, enabling a analysis of how images are processed at different layers. Conducting a analysis of the contributions from each layer and attention head, shedding light on the intricate interactions and functionalities within the model's layers. This in-depth exploration not only highlights the visual cues between layers but also examines their capacity to navigate the transition from abstract concepts to tangible objects. It unveils the model's mechanism to building an understanding of images, providing a strategy for adjusting attention heads between layers, thus enabling targeted pruning and enhancement of performance for specific tasks. Our research indicates that achieving a scalable understanding of transformer models is within reach, offering ways for the refinement and enhancement of such models.

Material property prediction, essential in AI-driven science, faces complexity due to the intricate influence of elemental combinations on superconducting characteristics. Acknowledging that the elemental composition and interactions within material molecules naturally align with graphical representation, we herein apply a graph neural network (GNN)-based methodology. By integrating molecular features alongside chemical descriptors, our approach comprehensively captures elemental synergy. Upon evaluation with the SuperCon dataset, our model showcases a significant leap in prediction accuracy, demonstrating its proficiency in estimating critical temperatures of superconductors. This breakthrough notably propels guided material synthesis and deepens the comprehension of superconductivity, thereby accelerating progress in the development of high-temperature superconductors.

The prediction of material properties is increasingly crucial for materials science, enabling design optimization and cost reduction. This study explores a novel approach: leveraging large language models (LLMs) to generate stable inorganic materials directly as text. Rather than predicting properties, we focus on generating plausible crystal structures encoded as text strings. We further investigate the performance differences between Chinese and English LLMs in this domain and demonstrate the capability of fine-tuning LLaMA-2 with few sample learning, presenting its scale learning curve. Our approach highlights the potential of LLMs to capture intricate structural information and generate novel material candidates, paving a new path for AI-driven materials discovery.

Recent metaphor identification approaches mainly consider the contextual text features within a sentence or introduce external linguistic features to the model.But they usually ignore the extra information that the data can provide, such as the contextual metaphor information and broader discourse information.In this paper, we propose a model augmented with hierarchical contextualized representation to extract more information from both sentence-level and discourse-level.At the sentence level, we leverage the metaphor information of words that except the target word in the sentence to strengthen the reasoning ability of our model via a novel label-enhanced contextualized representation.At the discourse level, the position-aware global memory network is adopted to learn long-range dependency among the same words within a discourse.Finally, our model combines the representations obtained from these two parts.The experiment results on two tasks of the VUA dataset show that our model outperforms every other state-of-the-art method that also does not use any external knowledge except what the pre-trained language model contains.

The Legal Judgement Prediction (LJP) is now under the spotlight. And it usually consists of multiple sub-tasks, such as penalty prediction (fine and imprisonment) and the prediction of articles of law. For penalty prediction, they are often closely related to the trial process, especially the attitude analysis of criminal suspects, which will influence the judgment of the presiding judge to some extent. In this paper, we firstly construct a multi-modal dataset with 517 cases of intentional assault, which contains trial information as well as the attitude of the suspect. Then, we explore the relationship between suspect`s attitude and term of imprisonment. Finally, we use the proposed multi-modal model to predict the suspect's attitude, and compare it with several strong baselines. Our experimental results show that the attitude of the criminal suspect is closely related to the penalty prediction, which provides a new perspective for LJP.