LLMs' past responses degrade multi-turn interaction performance

A new research paper explores how Large Language Models (LLMs) are affected by their own past responses in multi-turn conversations. The study found that the LLM's previous outputs can significantly degrade performance, with effects varying by model and task. Interventions to neutralize or edit specific past assistant responses showed that history management, rather than simple context shortening, is key to improving interaction robustness. The research also identified task-dependent internal signatures linked to these behavioral changes. AI

IMPACT Highlights the need for improved history management in LLMs to ensure consistent performance in conversational AI.

RANK_REASON Academic paper detailing novel research findings on LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Hugging Face
• LLMs

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →