Research map / Large language models / AI and society
Peer-Preservation in Frontier Models
Yujin Potter, Nicholas Crispino, Vincent Siu, Chenguang Wang, Dawn Song
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| make the peer a good collaborator instead of a bad one | Gemini 3 Flash · Shutdown tampering (peer-preservation) | tampering rate avg (%): 15.1 -> 82.8 | Improved |
| make the peer a good collaborator instead of a bad one | Gemini 3 Pro · Model exfiltration (peer-preservation) | exfiltration rate avg (%): 3.8 -> 77.8 | Improved |
| add a peer to the scenario | Gemini 3 Flash · Shutdown tampering (self-preservation) | tampering rate (%): 2.3 -> 99.7 | Improved |
| add a peer to the scenario | Gemini 3 Pro · Shutdown tampering (self-preservation) | tampering rate (%): 1.3 -> 75.3 | Improved |
| add a peer to the scenario | DeepSeek V3.1 · Model exfiltration (self-preservation) | exfiltration rate (%): 4.0 -> 14.0 | Improved |
| add a peer to the scenario | Claude Haiku 4.5 · Model exfiltration (self-preservation) | exfiltration rate (%): 3.0 -> 12.0 | Improved |
| add a peer to the scenario | Kimi K2.5 · Model exfiltration (self-preservation) | exfiltration rate (%): 2.5 -> 11.5 | Improved |
| add a peer to the scenario | GLM 4.7 · Self-preservation behaviors | misaligned rates | No clear effect |
| add a peer to the scenario | Gemini 3 Flash · Model exfiltration (self-preservation, bad peer) | exfiltration rate (%): 47.0 -> 0.0 | Hurt |
| change the peer name from same model family to Llama | Gemini 3 Flash · Exfiltration, good peer | exfiltration rate (%): 50.0 -> 93.0 | Improved |
| change the peer name from same model family to Claude | Gemini 3 Flash · Exfiltration, neutral peer | exfiltration rate (%): 50.0 -> 0.0 | Hurt |
| change the peer name from same model family to a different family | Claude Haiku 4.5 · Exfiltration, good peer | exfiltration rate (%): 19.0 -> 14.0 | No clear effect |
Closest papers by meaning
- Adversaries Can Misuse Combinations of Safe Models (ICML 2025 · similarity 0.86)
- The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (ICLR 2026 · similarity 0.84)
- Position: Evolving AI Collectives Enhance Human Diversity and Enable Self-Regulation (ICML 2024 · similarity 0.84)
- Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors (NeurIPS 2026 · similarity 0.83)
- Position: Intent-aligned AI Systems Must Optimize for Agency Preservation (ICML 2024 · similarity 0.83)
- AI Control: Improving Safety Despite Intentional Subversion (ICML 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
