2026/04/09 by Hiranya V. Peiris · 1 voice
Computer Science · Psychology · #Autoencoder #Deception #Emotion and Mood Recognition #Personality Traits and Psychology #Projection (relational algebra) #Situational ethics #Social Robot Interaction and HRI #Test (biology) #cs.AI #cs.CL #cs.HC #cs.LG
paper · pdf · open access · doi:10.48550/arxiv.2604.13466
published in arXiv (Cornell University) (Cornell University)
openalex publication_date 2026/04/09 · arxiv published 2026/04/09 · arxiv updated 2026/04/16 · openalex created_date 2026/04/17 · openalex updated_date 2026/07/28
The Claude Mythos Preview system card deploys emotion vectors, sparse autoencoder (SAE) features, and activation verbalisers to study model internals during misaligned behaviour. The two primary toolkits are not jointly reported on the most alignment-relevant episodes. This note identifies two hypotheses that are qualitatively consistent with the published results: that the emotion vectors track functional emotions that causally drive behaviour, or that they are a projection of a richer situational-context structure onto human emotional axes. The hypotheses can be distinguished by cross-referencing the two toolkits on episodes where only one is currently reported: most directly, applying emotion probes to the strategic concealment episodes analysed only with SAE features. If emotion probes show flat activation while SAE features are strongly active, the alignment-relevant structure lies outside the emotion subspace. Which hypothesis is correct determines whether emotion-based monitoring will robustly detect dangerous model behaviour or systematically miss it.