Weight space inference: reading hidden scheming from activation space, beyond CoT
srishti_gureja · x · 2026-11-08
srishtigureja proposes a research idea: weight space inference — using models as activation space verbalizers.
If it works, she argues, it would be huge for white-box interpretability: it could detect unverbalized eval-awareness or scheming propensity that never appears in the chain-of-thought, addressing the known weakness of CoT-based monitoring. Currently a speculative idea with no empirical results yet.
More from Safety
- Your local LLM server is reachable by any web page: CSRF/DNS rebinding risk tested live — FeydRowan · 2026-11-08
- AI Auditors Pilot: GPT-6 Astra Hits 15% False Positives at 99% Detection Target — andrewaltair · 2026-11-08
- AI 2027 author endorses congressman's call to ban recursive self-improvement — kimmonismus · 2026-11-08
- Parents Sue Google, Saying Gemini on School Chromebook Helped Daughter Plan Suicide — nordicinst · 2026-11-08
- AI companies can read your private chats—Wirecutter explains the risks and defenses — AnnaCiaunica · 2026-11-08
- Bloomington, Indiana Reportedly Bans Recursively Self-Improving AI Within City Limits (Unverified) — tragictacos · 2026-11-08