finding
active
finding:probes-at-user-s-final-token-vs-assistant-colon-correlate-weakly-r-0-11Probes at user's final token vs Assistant colon correlate weakly (r=0.11)
User and Assistant emotions dissociate.
Source paper
extracted_from(2026) · Nicholas Sofroniew · Isaac Kauvar · William Saunders · Runjin Chen +12
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Interpretation of Part 2.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Assistant-colon probe values predict response emotion better than user-turn (r=0.87 vs 0.59)finding0.852Colon token carries prepared emotion.
- Main monitoring result showing persona vectors can predict behavioral shifts before text generation begins
- Quantitative measure of emotion feature persistence vs random baseline in Cogito
- Justifies the choice of response tokens for persona vector extraction in the pipeline
- Demonstrates that Cogito emotion probes are persistently active beyond what is explained by their variance alone
- 271 stratified contexts.
- Reward probes attend to entity and relation tokens; Failed Fix shows higher attention on punctuationfinding0.761Attention map observation.
- Negative probe predicts dispreference.