Synthetic Persona Pretraining (SPP): Injecting Desired Persona from Token Zero Boosts Alignment in 3B Models
dhadfieldmenell · x · 2026-08-15
A new paper introduces Synthetic Persona Pretraining (SPP), which injects desired personas from the very start of pretraining to address the issue of uncontrollable mixed personas. Experiments on 3B models show significant gains in values and alignment.
Related event: New SPP Method Aligns Model Values via Pretraining(2 posts)→
More from Research
- InfinityEdit: Infinite Video Editing via Lightweight Adapter — Yunze Tong · 2026-08-24
- Tencent Benchmarks Hybrid-Thinking MLLMs for Response Alignment — tencent · 2026-08-24
- CLEAR Adapter Routing Balances LLM Safety and Utility — UIUC-CS · 2026-08-24
- Llama-Mobile: 2.7-Bit Quantization Shrinks Llama 3.2 Vision 11B to 3.7GB for Phones — Luka Ribar · 2026-08-24
- Retriever: A Framework for Asynchronous, Closed-Loop Robot Agents — ZeYanjie · 2026-08-24
- Converting GMMs ↔ PEFs for fast KLD approximation — FrnkNlsn · 2026-08-24