New research identifies hidden state in LLM agents as Persistent Semantic Entities

A new research paper introduces the concept of Persistent Semantic Entities (PSEs) to describe implicit state within tool-augmented LLM agents. These entities, which can persist across sessions and propagate between agents, are largely invisible to standard debugging methods. The study found that all tested models, ranging from 1.5B to 1T parameters, are susceptible to PSEs, with name binding being the primary mechanism. Preference and instruction contamination were identified as particularly concerning attack surfaces due to their persistence and lack of self-correction in many models. AI

IMPACT Highlights a new attack surface in LLM agents related to persistent, hidden state, potentially impacting agent reliability and security.

RANK_REASON Academic paper introducing a new concept and evaluation methodology for LLM agents. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• GPT-4o-mini
• Llama-3.1-8B
• LLM
• Persistent Semantic Entities
• Qwen2.5-coder

• paper
• safety

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