hypothesis
active
hypothesis:turn-averaged-features-will-outperform-per-token-features-on-whole-turn-tasks-such-as-summarizing-multi-turn-exchanges-or-judging-persona-safety-traits

Turn-averaged features will outperform per-token features on whole-turn tasks such as summarizing multi-turn exchanges or judging persona/safety traits

Prediction from coverage results.

Source paper

extracted_from
Turn-Averaged SAEs for Feature Discovery and Long-Context Attribution
(2026) · Kevin Der · Harish Kamath · Thompson, Ben

Neighborhood — ranked by edge-count