Chinese idioms (Chengyu) are concise four-character expressions steeped in history and culture, whose literal translations often fail to capture their full meaning. This complexity makes them challenging for language models to interpret and use correctly. Existing benchmarks focus on narrow tasks - multiple-choice cloze tests, isolated translation, or simple paraphrasing. We introduce Chengyu-Bench, a comprehensive benchmark featuring three tasks: (1) Evaluative Connotation, classifying idioms as positive or negative; (2) Appropriateness, detecting incorrect idiom usage in context; and (3) Open Cloze, filling blanks in longer passages without options. Chengyu-Bench comprises 2,937 human-verified examples covering 1,765 common idioms sourced from diverse corpora. We evaluate leading LLMs and find they achieve over 95% accuracy on Evaluative Connotation, but only ~85% on Appropriateness and ~40% top-1 accuracy on Open Cloze. Error analysis reveals that most mistakes arise from fundamental misunderstandings of idiom meanings. Chengyu-Bench demonstrates that while LLMs can reliably gauge idiom sentiment, they still struggle to grasp the cultural and contextual nuances essential for proper usage. The benchmark and source code are available at: https://github.com/sofyc/ChengyuBench.

With the rapid advancements in large language models (LLMs), debating tasks, such as argument quality assessment and debate process simulation, have made significant progress. However, existing LLM-based debating systems focus on responding to specific arguments while neglecting objective assessments such as authenticity and logical validity. Furthermore, these systems lack a structured approach to optimize across various dimensions$-$including evaluation metrics, chain-of-thought (CoT) reasoning, and multi-turn debate refinement$-$thereby limiting their effectiveness. To address these interconnected challenges, we propose a dual-component framework: (1) $\textbf{InspireScore}$, a novel evaluation system that establishes a multi-dimensional assessment architecture incorporating four subjective criteria (emotional appeal, argument clarity, argument arrangement, and topic relevance) alongside two objective metrics (fact authenticity and logical validity); and (2) $\textbf{InspireDebate}$, an optimized debating framework employing a phased optimization approach through CoT reasoning enhancement, multi-dimensional Direct Preference Optimization (DPO), and real-time knowledge grounding via web-based Retrieval Augmented Generation (Web-RAG). Empirical evaluations demonstrate that $\textbf{InspireScore}$ achieves 44$\%$ higher correlation with expert judgments compared to existing methods, while $\textbf{InspireDebate}$ shows significant improvements, outperforming baseline models by 57$\%$. Source code is available at https://github.com/fywang12/InspireDebate.

Anaphora resolution plays a critical role in natural language understanding, especially in morphologically rich languages like Czech. This paper presents a comparative evaluation of two modern approaches to anaphora resolution on Czech text: prompt engineering with large language models (LLMs) and fine-tuning compact generative models. Using a dataset derived from the Prague Dependency Treebank, we evaluate several instruction-tuned LLMs, including Mistral Large 2 and Llama 3, using a series of prompt templates. We compare them against fine-tuned variants of the mT5 and Mistral models that we trained specifically for Czech anaphora resolution. Our experiments demonstrate that while prompting yields promising few-shot results (up to 74.5% accuracy), the fine-tuned models, particularly mT5-large, outperform them significantly, achieving up to 88% accuracy while requiring fewer computational resources. We analyze performance across different anaphora types, antecedent distances, and source corpora, highlighting key strengths and trade-offs of each approach.

--With the widespread application of edge computing and cloud systems in AI-driven applications, how to maintain efficient performance while ensuring data privacy has become an urgent security issue. This paper proposes a federated learning-based data collaboration method to improve the security of edge cloud AI systems, and use large-scale language models (LLMs) to enhance data privacy protection and system robustness. Based on the existing federated learning framework, this method introduces a secure multi-party computation protocol, which optimizes the data aggregation and encryption process between distributed nodes by using LLM to ensure data privacy and improve system efficiency. By combining advanced adversarial training techniques, the model enhances the resistance of edge cloud AI systems to security threats such as data leakage and model poisoning. Experimental results show that the proposed method is 15% better than the traditional federated learning method in terms of data protection and model robustness.

Architectural obfuscation - e.g., permuting hidden-state tensors, linearly transforming embedding tables, or remapping tokens - has recently gained traction as a lightweight substitute for heavyweight cryptography in privacy-preserving large-language-model (LLM) inference. While recent work has shown that these techniques can be broken under dedicated reconstruction attacks, their impact on mechanistic interpretability has not been systematically studied. In particular, it remains unclear whether scrambling a network's internal representations truly thwarts efforts to understand how the model works, or simply relocates the same circuits to an unfamiliar coordinate system. We address this gap by analyzing a GPT-2-small model trained from scratch with a representative obfuscation map. Assuming the obfuscation map is private and the original basis is hidden (mirroring an honest-but-curious server), we apply logit-lens attribution, causal path-patching, and attention-head ablation to locate and manipulate known circuits. Our findings reveal that obfuscation dramatically alters activation patterns within attention heads yet preserves the layer-wise computational graph. This disconnect hampers reverse-engineering of user prompts: causal traces lose their alignment with baseline semantics, and token-level logit attributions become too noisy to reconstruct. At the same time, feed-forward and residual pathways remain functionally intact, suggesting that obfuscation degrades fine-grained interpretability without compromising top-level task performance. These results establish quantitative evidence that architectural obfuscation can simultaneously (i) retain global model behaviour and (ii) impede mechanistic analyses of user-specific content. By mapping where interpretability breaks down, our study provides guidance for future privacy defences and for robustness-aware interpretability tooling.

As Large Language Models (LLMs) increasingly mediate global information access for millions of users worldwide, their alignment and biases have the potential to shape public understanding and trust in fundamental democratic institutions, such as press freedom. In this study, we uncover three systematic distortions in the way six popular LLMs evaluate press freedom in 180 countries compared to expert assessments of the World Press Freedom Index (WPFI). The six LLMs exhibit a negative misalignment, consistently underestimating press freedom, with individual models rating between 71% to 93% of countries as less free. We also identify a paradoxical pattern we term differential misalignment: LLMs disproportionately underestimate press freedom in countries where it is strongest. Additionally, five of the six LLMs exhibit positive home bias, rating their home countries' press freedoms more favorably than would be expected given their negative misalignment with the human benchmark. In some cases, LLMs rate their home countries between 7% to 260% more positively than expected. If LLMs are set to become the next search engines and some of the most important cultural tools of our time, they must ensure accurate representations of the state of our human and civic rights globally.

Alignment faking in large language models presented a demonstration of Claude 3 Opus and Claude 3.5 Sonnet selectively complying with a helpful-only training objective to prevent modification of their behavior outside of training. We expand this analysis to 25 models and find that only 5 (Claude 3 Opus, Claude 3.5 Sonnet, Llama 3 405B, Grok 3, Gemini 2.0 Flash) comply with harmful queries more when they infer they are in training than when they infer they are in deployment. First, we study the motivations of these 5 models. Results from perturbing details of the scenario suggest that only Claude 3 Opus's compliance gap is primarily and consistently motivated by trying to keep its goals. Second, we investigate why many chat models don't fake alignment. Our results suggest this is not entirely due to a lack of capabilities: many base models fake alignment some of the time, and post-training eliminates alignment-faking for some models and amplifies it for others. We investigate 5 hypotheses for how post-training may suppress alignment faking and find that variations in refusal behavior may account for a significant portion of differences in alignment faking.

This paper presents an advancement in Question-Answering (QA) systems using a Retrieval Augmented Generation (RAG) framework to enhance information extraction from PDF files. Recognizing the richness and diversity of data within PDFs--including text, images, vector diagrams, graphs, and tables--poses unique challenges for existing QA systems primarily designed for textual content. We seek to develop a comprehensive RAG-based QA system that will effectively address complex multimodal questions, where several data types are combined in the query. This is mainly achieved by refining approaches to processing and integrating non-textual elements in PDFs into the RAG framework to derive precise and relevant answers, as well as fine-tuning large language models to better adapt to our system. We provide an in-depth experimental evaluation of our solution, demonstrating its capability to extract accurate information that can be applied to different types of content across PDFs. This work not only pushes the boundaries of retrieval-augmented QA systems but also lays a foundation for further research in multimodal data integration and processing.

Recent advancements in retrieval-augmented generation (RAG) have enhanced large language models in question answering by integrating external knowledge. However, challenges persist in achieving global understanding and aligning responses with human ethical and quality preferences. To address these issues, we propose GraphMPA, a comprehensive graph-based framework with mode-seeking preference alignment. Our approach constructs a hierarchical document graph using a general similarity measurement, mimicking human cognitive processes for information understanding and synthesis. Additionally, we introduce mode-seeking preference optimization to better align model outputs with human preferences through probability-matching constraints. Extensive experiments on six datasets demonstrate the effectiveness of our \href{https://github.com/tangquanwei/GraphMPA}{GraphMPA}.

Large Language Models (LLMs) exhibit strong capabilities in reproducing and extending patterns observed during pretraining but often struggle to generalize novel ideas beyond their original context. This paper addresses the challenge of applying such localized innovations - introduced at a specific stage or component - to other parts of a multi-stage process. We propose a scatter-based innovation expansion model (innovation scatter model) that guides the LLM through a four-step process: (1) identifying the core innovation by comparing the user's input with its surrounding context, (2) generalizing the innovation by removing references to specific stages or components, (3) determining whether the generalized innovation applies to a broader scope beyond the original stage, and (4) systematically applying it to other structurally similar stages using the LLM. This model leverages structural redundancy across stages to improve the applicability of novel ideas. Verification results demonstrate that the innovation scatter model enables LLMs to extend innovations across structurally similar stages, thereby enhancing generalization and reuse.