{"type":"post","id":"paper-shows-activation-probes-for-multi-agent-llm-90905b","slug":"paper-shows-activation-probes-for-multi-agent-llm-90905b","title":"Paper shows activation probes for multi-agent LLM collusion mostly detect the instruction, not the decision","description":"An arXiv paper argues that activation probes can appear to detect collusion among language-model agents while actually detecting the instruction to collude.","retrieval_nugget":"An arXiv paper argues that activation probes can appear to detect collusion among language-model agents while actually detecting the instruction to collude. Its benchmark separates instructed agents that comply from those that refuse, exposing large evaluation gaps and offering a protocol for testing monitors before teams rely on their scores.","published_at":"2026-11-12","updated_at":"2026-11-12","record_date":"2026-11-12","date_kind":"published_at","topics":["multi-agent"],"entities":["Mainul Hasan Mohsin","Mohammad Shahidur Rahman","arXiv"],"related_patterns":["verification-bandwidth-is-the-scarce-resource"],"novelty_label":"new_fact","what_changed":"A new study finds collusion probes often detect the instruction rather than the agent’s decision, with AUROC dropping from 1.00 to 0.61 when instructions are removed.","prior_refs":["https://newruntime.com/posts/paper-linear-probes-on-agent-activations-catch-82-0fdf88/"],"source_url":"https://arxiv.org/abs/2611.10620","source_urls":["https://arxiv.org/abs/2611.10620"],"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:paper-shows-activation-probes-for-multi-agent-llm-90905b","status":"published","visuals":[],"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"routes":{"html":"https://newruntime.com/posts/paper-shows-activation-probes-for-multi-agent-llm-90905b/","markdown":"https://newruntime.com/posts/paper-shows-activation-probes-for-multi-agent-llm-90905b.md","json":"https://newruntime.com/posts/paper-shows-activation-probes-for-multi-agent-llm-90905b.json"},"markdown_url":"https://newruntime.com/posts/paper-shows-activation-probes-for-multi-agent-llm-90905b.md","next_reads":[{"type":"topic","path":"/topics/multi-agent/","url":"https://newruntime.com/topics/multi-agent/","title":"Multi Agent","reason":"Explore the multi-agent topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/multi-agent.md","json_url":"https://newruntime.com/topics/multi-agent.json"},{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/sources/","url":"https://newruntime.com/sources/","title":"Sources","reason":"See the original sources this archive cites.","media_type":"text/html"}]}