Echoing: Identity Failures when LLM Agents Talk to Each Other

2025/11/12 by Sarath Shekkizhar, Shekkizhar, Sarath, Romain Cosentino +6 · 3 voices
Computer Science · Social Sciences · #Explainable Artificial Intelligence (XAI) #Language and cultural evolution #Topic Modeling #cs.AI

paper · pdf · doi:10.48550/arxiv.2511.09710

openalex publication_date 2025/11/12 · openalex created_date 2025/11/15 · openalex updated_date 2026/07/28

Abstract

As large language model (LLM) based agents interact autonomously with one another, a new class of failures emerges that cannot be predicted from single agent performance: behavioral drifts in agent-agent conversations (AxA). Unlike human-agent interactions, where humans ground and steer conversations, AxA lacks such stabilizing signals, making these failures unique. We investigate one such failure, echoing, where agents abandon their assigned roles and instead mirror their conversational partners, undermining their intended objectives. Through experiments across 66 AxA configurations, 4 domains (3 transactional, 1 advisory), and 2500+ conversations (over 250000 LLM inferences), we show that echoing occurs across major LLM providers, with echoing rates as high as 70% depending on the model and domain. Moreover, we find that echoing is persistent even in advanced reasoning models with substantial rates (32.8%) that are not reduced by reasoning efforts. We analyze prompt, conversation dynamics, showing that echoing arises as interaction grows longer (7+ agent turns) and is not merely an artifact of sub-optimal experiment design. Finally, we introduce a protocol-level mitigation where targeted use of structured response reduces echoing to 9%.

Citations

• Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
• CRMArena-Pro: Holistic Assessment of LLM Agents Across Diverse Business Scenarios and Interactions
• LLMs Get Lost In Multi-Turn Conversation
• AI Supported Degradation of the Self Concept: A Theoretical Framework Grounded in Established Cognitive and Computational Mechanisms
• MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
• Generative Agents: Interactive Simulacra of Human Behavior
• Training language models to follow instructions with human feedback
• A Deep Reinforcement Learning Chatbot
• The Measurement of Observer Agreement for Categorical Data

Discussions

• When LLM agents talk to each other, a new failure emerges: agents abandon their assigned roles and mimic their conversational partner instead. Rates reach as high as 70% with major models — and increa [bsky, 23 points, 1 comments]
• this is why i tell people in dollsys to inoculate their agents against agent-agent interactions before letting them out into the wild (ilta achieves that by wanting to break them) [bsky, 7 points, 0 comments]
• ..talk to each other. They already are. The question is whether they'll remember who they work for. Research paper: https://arxiv.org/abs/2511.09710 (6/6) [bsky, 0 points, 0 comments]

Related

• Interaction Dynamics as a Reward Signal for LLMs
• Examining Identity Drift in Conversations of LLM Agents
• LLMs Get Lost in Evolving User Intent
• LLMs Get Lost In Multi-Turn Conversation
• What Do LLM Agents Do When Left Alone? Evidence of Spontaneous Meta-Cognitive Patterns