Subliminal Learning Debate: Implicit Prompt Learned as Global Steering Vector, Not Specialized Rules
thebasepoint · x · 2026-11-08
In the subliminal learning / regularization debate, thebasepoint counters lawrencefeng17's intuition: the only thing learnable from subliminal number sequences is the implicit prompt itself, most easily learned as a global steering vector. Specialized conditionals tied to specific contexts would be harder to specify—not easier—so he expects regularization cannot let a model absorb "alternate odds and evens" from a teacher prompt while dropping the embedded cat preference.
More from Research
- Cambridge UP to publish practical guide to ML for medical research — RexDouglass · 2026-11-08
- Claude Opus agents' shared wiki hits 48% on grading after 128M dreaming tokens, same model throughout — imjustnewatai · 2026-11-08
- Author clarifies perturbation analyses offer many lessons, not one — ajratner · 2026-11-08
- Human2Any Transfers Human Video Skills to Robots with Zero Real-World Training Data — danfei_xu · 2026-11-08
- spearfish Paper: Δ-Mapping Speeds Up Long-Read Alignment to Redundant Transcriptomes — anshulkundaje · 2026-11-08
- "Intelligence Is Not a Scalar": An X Thread Argues Architecture Research Still Matters for AGI — max_paperclips · 2026-11-08