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.
More from Models
- Tencent Hy4 flash open-weights test: full data-to-deck workflow at $0.07 per 1M input tokens — JaynitMakwana · 2026-11-19
- Fan-made price chart for rumored Argon model family: Carbon is next iteration at identical pricing — koltregaskes · 2026-11-19
- Gemini and Nano Banana Reported Down by Users — CowBusiness5605 · 2026-11-19
- Dev laments copy-pasting between GPT-6.2 Astra and Fable 5.2: 'I'm a meat proxy' — Angaisb_ · 2026-11-19
- Dev's ideal multi-model stack: plan with Fable, execute with Sol, review with Astra — jdjohnson · 2026-11-19
- Yacine: two days without writing code, staring at a graphviz diagram as LLMs prove useless — yacineMTB · 2026-11-19