Description extraction emphasizes threatened egotism while Dialogue emphasizes idealism and reduces sadism in OLMo-3

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

• Dialogue extraction vector leads to more pronounced sadism and threatened egotism in Apertus-8B steered generationsfinding0.837
  Shows discourse type specifically shapes which facet of evil is elicited in Apertus
• If a dialogue agent is prompted with knowledge of its own LLM nature, it will enact a superposition of theories of selfhood, narrowing as conversation proceedshypothesis0.759
  Conditional prediction about how a well-informed dialogue agent would handle questions of personal identity
• All three discourse types (Description, Dialogue, Narration) yield significant cross-steering of evil personafinding0.750
  Shows each extracted direction generalizes to evaluation prompts from other discourse types
• When steered to the extreme away from the Assistant, Llama and Gemma shift to a theatrical persona characterized by mystical, poetic prose; Qwen more often hallucinates a human persona at extremesfinding0.744
  Characterizes what is on the far end of the Assistant Axis away from the Assistant
• People do not know—emotionally—how to please themselves. In part, they are prevented by society. And in part, they are prevented by themselves, by their inner thought police.claim0.741
  Diagnosis of why living structure is absent from the world: a failure of emotional knowledge enforced by social and internal constraints.
• Sadism facet of Evil persona grows at the beginning of Evil persona development in OLMo-3 pretrainingfinding0.741
  Facet-level finding correlating geometric refinement with qualitative semantic change
• It requires absolute removal of the individual ego, because what is created can no longer stand out and be separated from everything else.claim0.738
  Ego removal as a necessary condition for not-separate creation.
• Removing chain-of-thought in prompted setup nearly eliminates compliance gap for Claude 3 Opus in helpful-only and animal welfare settingsfinding0.736
  Contrasts with synthetic doc finding; suggests different mechanisms may be at play