Researchers have identified a phenomenon called contextual entrainment in large language models, where models assign higher probabilities to tokens previously seen in the prompt, regardless of their relevance. This suggests a mechanistic process, modulated by semantic factors, where specific "entrainment heads" within the model's attention circuits contribute to distraction. By disabling these heads, the models' output more closely resembles what would be produced without irrelevant context. Separately, a behavioral study evaluated how LLMs decide between using internal memory or external tools like web search. The study found that while some models like GPT-5 mini and Grok 4 Fast are highly selective in searching for outside knowledge, others like Mistral Small are less consistent, sometimes abstaining from search even when necessary. AI
IMPACT These studies offer insights into LLM mechanics and decision-making, potentially leading to more reliable and efficient AI agents.
RANK_REASON Two arXiv papers detailing new research findings on LLM behavior.
- arXiv
- GPT-5 mini
- Grok 4 Fast
- Mistral Small
- alphaXiv
- CatalyzeX
- Confident Fools or Careful Searchers? A Behavioural Study of When LLMs Search Beyond Their Static Internal Memory
- DagsHub
- Gotit.pub
- Hugging Face
- Jingcheng Niu
- Llama See, Llama Do: A Mechanistic Perspective on Contextual Entrainment and Distraction in LLMs
- Pathirage Deelaka Nipun
- ScienceCast
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
