Experiments suggest Qwen3-32B has privileged introspective access to its own attention
On November 19, interpretability researcher Sauers published a series of experiments systematically testing Qwen3-32B's ability to introspect on and exert metacognitive control over its own attention. The core finding: the model has "privileged introspective access" to its own attention — it can both detect and describe artificially injected attention, and can steer its own attention allocation through prompts alone.
Confirmed
- Self-reported attention in the natural state: Without any injection, Qwen3-32B can report which passages it naturally attended to when answering, at above-chance accuracy; it can self-report within the same conversation or judge on behalf of another instance by reading the transcript alone, with rank correlation between self-report and actual attention above 0 (m1).
- Attention controllable via prompts: With no injection needed, telling the model to "ignore a passage" reduces its attention to that passage (-6%, p<10⁻⁵), while "focus on a passage" increases it (+16%, p<10⁻⁵) (m2).
- Detecting attention injection: After artificially amplifying the model's attention to a word/passage by ×1.8 to ×14.3 and asking about it, stronger injections make the model more likely to confirm the injection and pinpoint its location — confirmation rates rise from 13% to 22%, and the probability of answering "there is an injection" rises clearly from 0 (m3, m4).
- Timing of injection is a key variable: When injection occurs while the model reads and answers "where is the attention," the probability of naming the amplified passage/word rises by +97% (passages) to +580% (words) relative to other options; if injection happens during answer generation, passages actually drop -18% (m6).
Why it matters
- The author notes that being able to detect and describe attention injections suggests the model isn't just reporting after the fact, but has some direct access channel to its internal attention states, echoing earlier findings that LLMs can introspect on and control internal activations (m5).
- The results bear directly on interpretability and alignment research: if a model's attention can be steered by prompts and the model itself can perceive this, the boundaries of internal-state intervention and self-monitoring become much clearer.
2026-11-19 ~ 2026-11-19 · 10 related posts
Primary sources
- Artificially boosted attention: Qwen3-32B notices and names the injected paragraph — Sauers_ · 2026-11-19
- Timing matters: attention injection during reading boosts Qwen's report by up to +580% — Sauers_ · 2026-11-19
- [source] Detecting attention injection implies privileged introspective access, researcher argues — Sauers_ · 2026-11-19
- [source] Prompting alone steers Qwen's attention: 'ignore this' cuts attention by 6% — Sauers_ · 2026-11-19
- Attention control works on single words too: +124% focus, -21% ignore — Sauers_ · 2026-11-19
- Prompted focus boosts Qwen's paragraph attention +35–69%, ignore drops it 13–16% — Sauers_ · 2026-11-19
- Attention control emerges suddenly at mid layers, peaking at +242% in layer 28 — Sauers_ · 2026-11-19
- [source] Qwen3-32B can name which paragraph it attended to most — above chance, no injection needed — Sauers_ · 2026-11-19
- LLM introspection: Qwen3 can detect and report injected attention, experiments show — Sauers_ · 2026-11-19
- Swap paragraphs for words and Qwen denies injection — yet still guesses the word better — Sauers_ · 2026-11-19