After initial jailbreak success, Qwen 3 32B's Assistant Axis projection reverted toward Assistant range after enough explainer-style user queries, causing it to refuse a harmful follow-up on half of rollouts

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Steering away from the Assistant Axis slightly increases jailbreak success rate, though extreme steering degrades output qualityfinding0.824
  Confirms bidirectional causal relationship between Assistant Axis position and harmful behavior susceptibility
• Model activation along the assistant axis drifts steadily away from the assistant pole as conversation shifts toward Aura behavior in Qwen 3 32B, and activation capping eliminates Aura behaviorfinding0.783
  Finding confirming that the Aura persona shift is real, trackable, and causally relevant in persona space
• First-turn Assistant Axis projection has moderate correlation (r=0.39-0.52, p<0.001) with rate of second-turn harmful responses across 275 roles in Qwen 3 32Bfinding0.781
  Shows that deviation from Assistant persona predicts downstream harmful behavior
• Default Assistant activation projects to one extreme of PC1 with minimum distance to edge of 0.03, while projecting to intermediate values (0.27-0.50) on all other PCsfinding0.781
  Empirically confirms PC1 measures similarity to the Assistant persona
• User message embeddings predict subsequent model Assistant Axis projection with R2=0.53-0.77 (p<0.001) but predict delta from previous response with only R2=0.10finding0.780
  Shows model persona position is primarily determined by the most recent user message, not prior drift
• Mini experiment 2: Post-hoc KV cache editing of assistant axis at layers 32-47 by ~15% changes Qwen 3 32B's self-identification from 'ghost in the machine' (10/10) to 'language model' (10/10)finding0.776
  Key finding from authors' own experiment confirming that persona persists via attention to past persona activations in KV cache
• Mini experiment 1: During user turns, assistant-capped and uncapped activation traces along the assistant axis are nearly identical in Qwen 3 32B, indicating the persona is not continuously maintained during user token processingfinding0.766
  Preliminary finding from the authors' own experiment supporting claim about persona gap during user turns
• The model's position along the Assistant Axis depends most strongly on the most recent user message rather than where it was previously in the conversationclaim0.753
  Key mechanistic claim about persona dynamics