Concept cone methodology failed to produce a meaningful cone for sentiment on Stanford Sentiment Treebank

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

• Does sentiment have a higher-dimensional concept cone representation, and if so, what methods could find it?question0.794
  Open question raised by failed sentiment cone extension
• Concept cone truth interventions would generalize to larger frontier models and multimodal settingshypothesis0.745
  Key robustness question raised as future work
• How can we develop better methods for measuring the model's evaluation-relevant beliefs beyond reading its chain of thought?question0.728
  Gap in current evaluation methods; current work relies on CoT monitoring which may miss unverbalized beliefs.
• Individual cone basis vectors may correspond to interpretable semantic facets of truth such as temporal facts, geographic facts, or commonsensehypothesis0.722
  Future direction hypothesis for giving semantic meaning to individual axes
• All cohort benchmarks measure output, not state, and are subject to eval-awareness contamination.claim0.722
• Persistent conversational context that produced emotion-relevant activation is a plausible driver for the observed persistence results.claim0.721
  Acknowledged alternative explanation that the paper does not rule out
• The marker method can be adapted for AI systems by focusing less on behavioral evidence and more on architectural evidence.claim0.721
  Proposal for assessment framework.
• Persistent conversational context that produced emotion-relevant activations is a plausible driver of observed persistence resultsclaim0.721
  Authors' caveat that conversational context persistence rather than internal emotion state persistence could explain findings