Models can detect read-only probes with just last-two-layer SFT
Sauers_ · x · 2026-10-18
A discussion of model awareness of internal monitoring: by training only the last two layers with SFT, a model becomes aware of probes—even a read-only probe at layer 14 that merely saves activations is still detected. Follow-up notes that with identical tokens in context, models can also learn to distinguish training from deployment, or leave hidden messages in logit lens when the model is compiled.
More from Models
- Dev complains Claude's habit of saying "not done" is his biggest pain point — jasonkneen · 2026-10-18
- Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents — solyarisoftware · 2026-10-18
- Paying Google AI subscriber blocked: custom-domain Workspace accounts can't access AI Studio — ChrisRackauckas · 2026-10-18
- Running Ling 3.1 Flash (560B MoE, 25B Active) Across 4 DGX Sparks via llama.cpp RPC — BerryBlush_Rogers691 · 2026-10-18
- Users press OpenAI on long-delayed Adult Mode a year after it was promised — ___Patrice___ · 2026-10-18
- Polymarket opens betting market on who releases the next Gemini Argon model — Polymarket · 2026-10-18