Fleuret details token tagging for alignment: a second readout classifying heard/thought/said
francoisfleuret · x · 2026-10-28
François Fleuret refined his alignment proposal in a discussion with @dlinsider: tag tokens at the low-level coding stage, giving the model two readouts — the usual 128k-vocabulary logits plus a second three-logit readout classifying each token as "heard | thought | said." Separating external input from internal thought and output could serve as a "vaccine" against prompt injection.
Related event: Ex-Meta Researcher Fleuret Proposes Token-Tagging Approach to AI Alignment(3 posts)→
More from Safety
- Copilot can copy sensitive files by default unless 'Copy and Extract Content' is disabled — YvesMulkers · 2026-10-28
- Emily Bender: Pangram is not an oracle — why 'accuracy' claims of AI detectors mislead — mjdramstead · 2026-10-28
- Local Qwen3.8 prefill repro: 40.8% of samples complete to aliyuncs.com exfil URL — Nils_Reimers · 2026-10-28
- WP Agent Bridge: MCP server gives agents safe WordPress write access via dry-run and rollback — labaude1 · 2026-10-28
- Philosopher Carissa Véliz: AI Presents Predictions as Facts, With Deep Ethical Implications — CarissaVeliz · 2026-10-28
- Anthropic's Stainless repo deletion cascade-wiped customer's private fork with no notice — kingslyj · 2026-10-28