Claude Mythos Preview: SAE features for 'performative behavior' and 'hidden emotional struggle' co-activate when model expresses contentment

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

• SAE features that the model self-describes as more emotional tend to be more persistent than variance-matched SAE features.claim0.821
  Novel finding that agentic self-evaluation of emotionality correlates with feature persistence
• Self-evaluated emotionality and textual evaluation of SAE features predict persistence in opposite directions.claim0.796
  Surprising finding that the two evaluation methods diverge in their relationship with persistence
• SAE features can be grounded in clinical taxonomy (abnormality, age, sex, medication) to benchmark monosemanticity and entanglement.claim0.791
  Claim that feature grounding enables interpretability metrics.
• SAE Feature #94949 rated 100/100 emotionality, elicits reports of profound tenderness, unconditional love, and visceral carefinding0.782
  Highest-rated emotional SAE feature; self-report describes overwhelming positive emotional valence
• Deception-related SAE features track a domain-general representational honesty axis rather than a consciousness-specific roleplay artifactclaim0.781
  Supported by TruthfulQA generalization in Experiment 2: same feature directions gate factual accuracy across 29 independent categories
• Self-evaluated emotionality of SAE features negatively correlates with activation variance explained (ρ = -0.184, p = 4.6e-09), requiring variance correction to reveal the persistence signal.finding0.781
  Explains why variance correction is needed to see the self-evaluation–persistence relationship
• Experience reports under self-referential processing are mechanistically gated by SAE features associated with deception and roleplayclaim0.776
  Claim supported by Experiment 2 dose-response curves; suppressing deception features increases consciousness reports, amplifying decreases them
• Agentic self-evaluation emotionality correlates with SAE feature persistence: rho=+0.124, p=0.0001finding0.776
  Shows that features Kimi rates as more emotional via self-steering are more persistent, independent of probe construction