Simple linear probes could reveal the persona base models play, argues interp enthusiast
cephaloform · x · 2026-11-19
- Replying to a question about identifying the "persona" a base model adopts when completing text, cephaloform suggests a fun experiment: you don't need heavy interpretability methods.
- Simple linear probes for attributes like occupation, ethnicity, or fiction/nonfiction would already be interesting to see, and a full decomposition would be even better.
Related event: Researchers Propose Linear Probes to Detect Personas in Base Models(3 posts)→
More from Research
- Stanford's Kundaje: ChIP-seq resolution limits motif-level deep learning analysis — anshulkundaje · 2026-11-19
- The biggest advantage of a top lab isn't money — it's saving you years — madeofAjala · 2026-11-19
- LEV: Looped decision model lifts 2.6B accuracy 63.2% to 76.2% with adaptive compute — jiqizhixin · 2026-11-19
- Aurora2-400M-A20M: a sparse MoE pretrained from scratch in 36 hours on one RTX Pro 6000 — Tall_Abrocoma_3533 · 2026-11-19
- Researcher calls METR/Anthropic '2x speedup' RSI analysis statistically dubious — jessi_cata · 2026-11-19
- Linear Probes Could Reveal Which Persona a Base Model Is Playing, Researchers Suggest — cephaloform · 2026-11-19