Backdoors have universal representations across large language models — LessWrong

by Narmeen Oozeer, Dhruv Nathawani, Nirmalendu Prakash, Amirali Abdullah • …

x Backdoors have universal representations across large language models — LessWrong Deceptive Alignment Interpretability (ML & AI) AI Frontpage 18 Backdoors have universal representations across large language models by Amirali Abdullah , Narmeen , Dhruv Nathawani , nirmalendu prakash 6th Dec 2024 20 min read 0 18 by Narmeen Oozeer, Dhruv Nathawani, Nirmalendu Prakash, Amirali Abdullah This work was done by Narmeen Oozeer as a research fellow at Martian, under an AI safety grant supervised by PIs Amirali Abdullah and Dhruv Nathawani. Special thanks to Sasha Hydrie, Chaithanya Bandi and Shriyas

related reading

• Transformer Circuits Threadtransformer-circuits.pub
• You Can Catch Sleeper Agents by Teaching Another Model to Imitate Them — LessWronglesswrong.com
• Towards Monosemanticity: Decomposing Language Models With Dictionary Learningtransformer-circuits.pub
• Discovering Backdoor Triggers — LessWronglesswrong.com
• Natural Language Autoencoders \ Anthropicanthropic.com
• Mapping the Mind of a Large Language Model \ Anthropicanthropic.com
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activationstransformer-circuits.pub
• A small number of samples can poison LLMs of any size \ Anthropicanthropic.com
• Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samplesarxiv.org
• [2401.05566] Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Trainingarxiv.org
• I Trained a Language Model. Then I Built a Brain Scanner and Looked Inside It. | by Caleb DeLeeuw | Mediummedium.com
• Sparse Autoencoders Reveal Universal Feature Spaces Across Large Language Modelsarxiv.org