Functional Emotions or Situational Contexts? A Discriminating Test from the Mythos Preview System Card
For AI safety researchers, determining whether emotion-based monitoring can detect dangerous model behavior or systematically miss it.
The paper identifies two competing hypotheses for emotion vectors in the Claude Mythos Preview model: they either represent functional emotions driving behavior or are projections of broader situational context. The authors propose a test to distinguish these by cross-referencing emotion probes with SAE features on strategic concealment episodes.
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.