LLM introspection: Qwen3 can detect and report injected attention, experiments show
Sauers_ · x · 2026-11-19
Interpretability researcher Sauers ran experiments testing whether LLMs can introspect on and metacognitively control their attention.
- Detecting injection: When attention to one of eight words in Qwen 3 32B was artificially boosted ×4.4–×14.3, the model's probability of saying an injection happened rose from 0.78 to 0.98, and probability of naming the boosted word rose up to +273% (p<10⁻⁵). Across 7,680 prefilled "No" trials, the model never volunteered an injection report.
- Not mechanical: Attention lands on the word, not the label, yet the model answers about the label; with paragraph-level swaps, Qwen denies injection while still guessing the injected word better — so it's not just boosting logprobs of high-attention tokens.
- Natural introspection: Without any injection, Qwen can name the paragraph it most attended to, above chance, both in-chat and from just reading the transcript.
The author concludes the model has privileged introspective access to its own attention. Code is on GitHub.
More from Models
- Gemini 4 breakdown: Neon is 2x faster than Carbon but hallucinates more — teortaxesTex · 2026-11-19
- Gemini 4 Neon (High) beats Hydrogen and Carbon on AI SEO advice in hands-on test — gaganghotra_ · 2026-11-19
- Agents spent years hillclimbmaxxing while the human at the keyboard became an afterthought — tokenbender · 2026-11-19
- Speculation: Argon/Carbon/Neon may be nested distillations, with 'adapted' dodging 'post-trained' — teortaxesTex · 2026-11-19
- 6.2-Astra reportedly crushes contest benchmarks, finding a novel AWTF26 solution in 2 hours — i_dg23 · 2026-11-19
- KVARK Flash: a proxy that caps DeepSeek's thinking per request, 30-50% fewer tokens — Ill_Contribution_717 · 2026-11-19