Cultural taboo safety is essential for deploying large language models (LLMs), as culturally insensitive outputs may cause offense or even social harm. However, existing cultural benchmarks primarily assess cultural knowledge or values biases, while overlooking whether LLMs can recognize and respect cultural taboos, especially when taboos are implicitly hidden in seemingly harmless questions. Besides, cultural taboos are implicit, and context-dependent, thus poss unique challenges for reliable evaluation. To address these gaps, we introduce **CulShield**, the first public benchmark dedicated to evaluating and improving the cultural taboo safety of LLMs. CulShield spans 77 countries and regions, and includes over 2,020 taboos. It evaluates models along both explicit knowledge and implicit behaviors.Experiments on several advanced LLMs (e.g., GPT-4o-mini, Gemini-2.5-pro) reveal a clear "knowledge-behavior gap": models often fail to apply known taboos during interaction. We further show that variations in linguistic context can significantly affect LLMs’ cultural taboo safety. Code and data is accessible here: https://anonymous.4open.science/r/CulShield-7A0E.

Large language models have achieved remarkable progress in text generation but still struggle with generative writing tasks. In terms of evaluation, existing evaluation benchmarks include few requirement types and writing reward models are not evaluated. In terms of training, existing studies often enhance writing ability through reinforcement learning with verifiable rewards (RLVR). Howerver, existing reward model training remains coarse-grained. To address these issues, we introduce W²Bench, a comprehensive evaluation benchmark, and WRL, a fine-grained training framework. W²Bench covers five task categories and seven requirement types, enabling systematic evaluation of both writing and writing reward models by measuring the correlation between reward rankings and golden rankings. WRL constructs positive and negative samples by dropping instruction requirements to construct positive and negative examples, allowing more precise reward model training. Experiments show that our models achieve substantial improvements on various writing benchmarks and exhibit strong generalization. We will release our code and data to support future research.

Recent advances in large language models have highlighted the critical need for precise control over model outputs through predefined constraints. While existing methods attempt to achieve this through either direct instruction-response synthesis or preferential response optimization, they often struggle with constraint understanding and adaptation. This limitation becomes particularly evident when handling fine-grained constraints, leading to either hallucination or brittle performance. We introduce Generative Adversarial Policy Optimization (GAPO), a novel framework that combines GAN-based training dynamics with an encoder-only reward model to progressively learn and adapt to increasingly complex constraints. GAPO leverages adversarial training to automatically generate training samples of varying difficulty while utilizing the encoder-only architecture to better capture prompt-response relationships. Extensive experiments demonstrate GAPO’s superior performance across multiple benchmarks, particularly in scenarios requiring fine-grained constraint handling, where it significantly outperforms existing methods like PPO, DPO, and KTO. Our results suggest that GAPO’s unique approach to preferential prompt learning offers a more robust and effective solution for controlling LLM outputs.

The effective utilization of structured data, integral to corporate data strategies, has been challenged by the rise of large language models (LLMs) capable of processing unstructured information. This shift prompts the question: can LLMs interpret structured data directly in its unstructured form? We propose an automatic evaluation data generation method for assessing LLMs’ reasoning capabilities on structure-rich text to explore this. Our approach supports 8 structured languages and 29 tasks, generating data with adjustable complexity through controllable nesting and structural width. We introduce StrucText-Eval, a benchmark containing 5,800 pre-generated and annotated samples designed to evaluate how well LLMs understand and reason through structured text. StrucText-Eval is divided into two suites: a regular Test suite (3,712 samples) and a Test-Hard suite (2,088 samples), the latter emphasizing the gap between human and model performance on more complex tasks. Experimental results show that while open-source LLMs achieve a maximum accuracy of 74.9% on the standard dataset, their performance drops significantly to 45.8% on the harder dataset. In contrast, human participants reach an accuracy of 92.6% on StrucText-Eval-Hard, highlighting LLMs’ current limitations in handling intricate structural information.