Tag
The open-source project lieflat-gongwen performs a full statistical analysis of 1.02 million words of real official documents, distilling the writing style of official documents into an Agent Skill that can be used by agents such as Codex and Claude Code. This reduces the "AI flavor" of AI-generated official documents, enabling reproduction of 7 typical document styles with quantitative self-checking parameters.
HuggingModels 介绍 novel-humanize-4b：一个基于 Qwen3-4B 构建的 4B 模型，可将生硬的中文文本改写为自然、拟人化的文风，已获 756 次下载和 66 个点赞。
中国电信发布 Xing4.0-29B Intent，一个基于 MoE 架构的 29B/4B 专业化模型，面向意图理解与结构化信息抽取，支持意图分类、命名实体识别、关系抽取与事件抽取四项任务，采用 Apache 2.0 开源许可。
APRiL 提出了一个针对中文社交感知能力的开放式基准，包含 2,196 个评测实例，涵盖六类社交场景和两项互补任务。对来自六个模型家族的 12 个模型的实验揭示了一致的“理解-响应”差距：高质量的场景分析能提升响应质量，而薄弱的分析反而可能降低响应质量。
ste-zh 是一个开源 Agent Skill，借鉴 ASD-STE100 简化技术英语规范，约束 AI Agent 用清晰、无歧义的中文汇报工作结果，支持通过 skills CLI 或手动安装到 Claude Code 等环境。
这篇论文在极小规模（8.7M 参数、4 层）的中文 BERT 上严格控制变量，对比了 MLM、WWM 和 MacBERT 三种预训练策略，发现小规模下的策略排名（MLM > WWM ≫ MacBERT）与大规模结论相反，并指出 MacBERT 在同义词词典覆盖不足时会出现困惑度严重退化、训练损失不可靠的问题。
UO-FIE is a parameter-efficient system for factivity inference that combines exact-label supervision with graded utility, achieving first place in the fine-tuning track of FIE2026 with a macro utility of 0.8316.
CMNIE is a new information extraction benchmark for Chinese military news that jointly annotates events, entities, and relations to evaluate models on schema adherence and exact span matching.
GUIDE is a generative unsupervised framework for Chinese query correction that uses phonetic and visual shared-ID encoding to constrain corrections and adapt to changing vocabularies, outperforming baselines in experiments and online A/B testing.
This paper proposes SeTox, a search-augmented LLM framework for detecting implicit toxicity in Chinese neologisms by leveraging real-time web context and public consensus. Experiments show that even 3B-scale models outperform larger recent models on this task.
This paper proposes CNM, a lightweight augmentation that injects discrete compositional structure of Chinese characters into BERT via Ideographic Description Sequences, improving performance on rare and out-of-vocabulary characters while preserving general NLU accuracy.
This paper proposes a dual-threshold hard example mining strategy for cross-platform Chinese offensive comment detection, addressing performance degradation due to domain shift. The method fine-tunes a RoBERTa model on the COLD dataset and adapts it to four Chinese social media platforms with minimal labeled data.
This paper establishes a reproducible multi-architecture baseline for token-level Chinese metaphor identification using the MIPVU framework and the PSU Chinese Metaphor Corpus. It compares encoder models like RoBERTa and MelBERT against the Qwen3.5-9B generative model, releasing code and data to facilitate future research.
Researchers from Peking University introduce CFMS, the first fine-grained Chinese multimodal sarcasm detection benchmark with 2,796 image-text pairs and a triple-level annotation framework (sarcasm identification, target recognition, explanation generation), along with a novel RL-augmented in-context learning method (PGDS) that significantly outperforms existing baselines.