Qwen3-32B can sense when its attention is artificially boosted — 58% vs chance

Sauers_ · x · 2026-11-19

A self-introspection experiment: researchers artificially boosted Qwen3-32B's attention to one of four paragraphs (×20, layers 21–53) and told the model only its attention could be altered. Asked which paragraph was boosted, 58% of answers pointed to the right one, well above chance (p < 10⁻⁴); a forced-random framing dropped accuracy to 37% but still above chance. Controls didn't show the effect, suggesting the model genuinely distinguishes injected attention from paragraph-average activations.

Original post →

More from Models

Models channel →