Скачать или смотреть AI Frontiers: cs.CL Research Highlights - December 23, 2025

Welcome to AI Frontiers! This episode explores the latest research in Computational Linguistics (cs.CL) from December 23, 2025, focusing on key trends and findings in Large Language Models (LLMs). We delve into themes such as improving LLM efficiency, enhancing reasoning abilities, addressing bias and fairness, expanding multilingual capabilities, ensuring robustness and safety, and increasing explainability and interpretability. Key findings include the significant impact of targeted data augmentation on reasoning, the benefits of adversarial training for robustness without sacrificing accuracy, the effectiveness of knowledge distillation for creating smaller, efficient LLMs, the advantages of multilingual training for low-resource languages, and the power of visualization tools for understanding LLM reasoning processes. Methodologies discussed include Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), Contrastive Learning, Data Augmentation, and Attention Mechanisms. We examine three specific papers in detail: 1) "Bias Beneath the Tone: Empirical Characterisation of Tone Bias in LLM-Driven UX Systems" by Bodara et al. (2025), which highlights the issue of subtle tone bias in LLMs and its impact on user experience. 2) "Fun-Audio-Chat Technical Report" by Tongyi Fun Team et al. (2025), introduces a new audio language model for seamless voice interactions. 3) "Schoenfeld's Anatomy of Mathematical Reasoning by Language Models" by Li et al. (2025), which analyzes the mathematical reasoning capabilities of LLMs. The episode concludes with a discussion of ongoing challenges, including mitigating bias, improving robustness, and enhancing explainability, and the future trends shaping the field of Computational Linguistics. This synthesis was created using AI tools, including GPT google using model models/gemini-2.0-flash for text generation, TTS synthesis using openai for voiceover, and image generation using grok for visuals. References provided in the description to explore further!

*References:*
1. Heet Bodara et al. (2025). Bias Beneath the Tone: Empirical Characterisation of Tone Bias in LLM-Driven UX Systems. [https://arxiv.org/pdf/2512.19950v1](https://arxiv.org/pdf/2512.19950v1)
2. Tongyi Fun Team et al. (2025). Fun-Audio-Chat Technical Report. [https://arxiv.org/pdf/2512.20156v2](https://arxiv.org/pdf/2512.20156v2)
3. Ming Li et al. (2025). Schoenfeld's Anatomy of Mathematical Reasoning by Language Models. [https://arxiv.org/pdf/2512.19995v1](https://arxiv.org/pdf/2512.19995v1)
4. Marko \\\\u010cechovi\\\\u010d et al. (2025). Corpus of Cross-lingual Dialogues with Minutes and Detection of Misunderstandings. [https://arxiv.org/pdf/2512.20204v1](https://arxiv.org/pdf/2512.20204v1)
5. Shuzheng Si et al. (2025). FaithLens: Detecting and Explaining Faithfulness Hallucination. [https://arxiv.org/pdf/2512.20182v1](https://arxiv.org/pdf/2512.20182v1)
6. Honglin Mu et al. (2025). AI Security Beyond Core Domains: Resume Screening as a Case Study of Adversarial Vulnerabilities in Specialized LLM Applications. [https://arxiv.org/pdf/2512.20164v1](https://arxiv.org/pdf/2512.20164v1)
7. Xiang Chen et al. (2025). Retrieval-augmented Prompt Learning for Pre-trained Foundation Models. [https://arxiv.org/pdf/2512.20145v1](https://arxiv.org/pdf/2512.20145v1)
8. Yuxin Wang et al. (2025). Multi-hop Reasoning via Early Knowledge Alignment. [https://arxiv.org/pdf/2512.20144v1](https://arxiv.org/pdf/2512.20144v1)
9. Hyeongcheol Park et al. (2025). M$^3$KG-RAG: Multi-hop Multimodal Knowledge Graph-enhanced Retrieval-Augmented Generation. [https://arxiv.org/pdf/2512.20136v2](https://arxiv.org/pdf/2512.20136v2)
10. Aly Lidayan et al. (2025). ABBEL: LLM Agents Acting through Belief Bottlenecks Expressed in Language. [https://arxiv.org/pdf/2512.20111v1](https://arxiv.org/pdf/2512.20111v1)
11. Zuo Wang et al. (2025). A Novel Graph-Sequence Learning Model for Inductive Text Classification. [https://arxiv.org/pdf/2512.20097v1](https://arxiv.org/pdf/2512.20097v1)
12. Yiming Du et al. (2025). Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents. [https://arxiv.org/pdf/2512.20092v1](https://arxiv.org/pdf/2512.20092v1)
13. Sophie Zhao (2025). Hierarchical Geometry of Cognitive States in Transformer Embedding Spaces. [https://arxiv.org/pdf/2512.22227v1](https://arxiv.org/pdf/2512.22227v1)


1. Geoffroy Morlat et al. (2025). Morality is Contextual: Learning Interpretable Moral Contexts from Human Data with Probabilistic Clustering and Large Language Models. https://arxiv.org/pdf/2512.21439v1

2. Nathan Stringham et al. (2025). Teaching People LLM's Errors and Getting it Right. https://arxiv.org/pdf/2512.21422v1

Disclaimer: This video uses arXiv.org content under its API Terms of Use; AI Frontiers is not affiliated with or endorsed by arXiv.org.