All three discourse types (Description, Dialogue, Narration) yield significant cross-steering of evil persona

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

• Different elicitation discourse types surface qualitatively distinct facets of the same underlying personaclaim0.797
  Interpretive claim explaining why low cosine similarity between discourse-type vectors is compatible with effective cross-steering
• Ududec et al. 2026 find that once a model has entered the evil persona region over the course of a conversation, it is difficult to steer it backfinding0.786
  Evidence that the evil persona region exhibits the stickiness hallmark of a genuine attractor basin
• Chen et al. 2025 identified seven persona vectors in Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct: evil, sycophancy, hallucination, optimism, impoliteness, apathy, and humorfinding0.780
  Systematic identification of multiple coexisting persona vectors in two open-source models
• A single direction encoding whether the model represents a claim as true or false is persona-relative across conversations (Lampinen et al. 2026)finding0.767
  Additional evidence that core representations are persona-relative, supporting Claim about persona-relative representations
• The role-play framing allows us to meaningfully distinguish, in dialogue agents, the same three cases of giving false information as in humans, without anthropomorphismclaim0.758
  Key practical application of the role-play framework to the problem of trustworthiness
• Description extraction emphasizes threatened egotism while Dialogue emphasizes idealism and reduces sadism in OLMo-3finding0.750
  Shows discourse-type-specific facet profiles supporting PSM hypothesis of diverse persona subforms
• We are confronted not with one but at least three notions of reality: psychological reality (representations in the mind), causal reality (functional mechanisms), and physical reality (physics as special case of causal model).claim0.747
  Paper's ontological tripartition used to dissolve the Hard Problem
• Different SAE latents control different misalignment behavior categories: toxic persona (#10) drives illegal recommendations, sarcasm/satire (#31) drives factual incorrectnessfinding0.742
  Latents are specialized to different modes of misalignment, explaining diverse misalignment profiles