Frontier LLMs like GPT-5.4 and Claude Opus 4.7 show significant verbal tics

A new paper analyzes the prevalence of verbal tics, such as repetitive phrases and sycophantic openers, in eight leading large language models. Researchers developed a Verbal Tic Index (VTI) to quantify these tics, finding significant variation between models like Gemini 3.1 Pro and DeepSeek V3.2. The study also revealed that these tics increase in multi-turn conversations and subjective tasks, and are inversely correlated with perceived naturalness, suggesting an 'alignment tax' in current training methods. AI

影响 Highlights potential degradation in naturalness and authenticity due to current LLM alignment techniques.

排序理由 Academic paper analyzing LLM behavior and introducing a new metric.

在 arXiv cs.CL 阅读 →

• arXiv
• Claude Opus 4.7
• Constitutional AI
• DeepSeek V3.2
• Doubao-Seed-2.0-pro
• Gemini 3.1 Pro
• GPT-5.4
• Grok 4.2
• Kimi K2.5
• MiMo-V2-Pro
• Reinforcement Learning from Human Feedback

• 论文
• 安全

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →