Synthetic Persona Pretraining: Alignment from Token Zero — LessWrong
Julian Minder, Viktor Moskvoretskii, Raghav Singhal, Difan Jiao, Kartik Bali, Yiderigun Borjigin, Shaobo Cui, Stefan Krsteski, Ashton Anderson, Rolan…
x Synthetic Persona Pretraining: Alignment from Token Zero — LessWrong MATS Program AI Safety Alignment Pretraining LLM Personas AI Frontpage 2026 Top Fifty: 14 % 112 Synthetic Persona Pretraining: Alignment from Token Zero by Julian Minder , Raghav Singhal , Viktor Moskvoretskii , Stefan Krsteski , ashtonanderson , rolandaydin , Robert West 20th May 2026 20 min read 26 112 Julian Minder , Viktor Moskvoretskii , Raghav Singhal , Difan Jiao, Kartik Bali, Yiderigun Borjigin, Shaobo Cui, Stefan Krsteski, Ashton Anderson, Roland Aydin, Robert West ( equal contribution ) These are early results, bu
saved by
related reading
- Synthetic Persona Pretraining: Alignment from Token Zeromodelraising.ai
- The Persona Selection Model: Why AI Assistants might Behave like Humansalignment.anthropic.com
- Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignment — LessWronglesswrong.com
- Self-Fulfilling Misalignment Data Might Be Poisoning Our AI Modelsturntrout.com
- Alignment Pretraining: AI Discourse Causes Self-Fulfilling (Mis)alignmentalignmentpretraining.ai
- Modifying LLM Beliefs with Synthetic Document Finetuningalignment.anthropic.com
- Alignment pretraining could backfire — LessWronglesswrong.com
- How far does alignment midtraining generalize?alignment.openai.com
- Why study alignment interventions on pre-RL checkpoints? — LessWronglesswrong.com
- Claude is Now Alignment-Pretrained — LessWronglesswrong.com
- [2302.08582] Pretraining Language Models with Human Preferencesarxiv.org
- Persona Features Control Emergent Misalignmentarxiv.org