Unverified50% confidenceFactExact time
The identification and refusal of compositional harm are decoupled: nine standard direction-finding methods detect compositional harm with an AUC of up to 0.99, but ablating these directions either does not affect refusal behavior or destroys model coherence.
1
Sources
50%
Confidence
Long-term
Relevance
11/6/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe study constructed paired propagation examples, for which any representation that erases direction has a provable random-level upper bound on classification performance (AUROC=0.5).70% similarUnverifiedThe standard practice aggregates all heterogeneous harmful prompts into a single uniform mean vector, which may conflate distinct refusal-related phenomena.69% similarUnverifiedAUC=0.5 means the diagnostic signal is no better than random guessing; the closer AUC is to 1, the better the signal distinguishes positive from negative samples; AUC=0.008 means the diagnostic signal is almost entirely reversed or ineffective.69% similarUnverifiedUsing the behavioral features of each state for failure prediction, the AUROC on the held-out set reaches as high as 0.94.69% similarUnverifiedWhen the ratio of positive to negative samples is extreme, the area under the PR curve (AUPRC) is often more revealing of true performance than AUC.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1141925API
curl https://kongchang.com/api/v1/knowledge/claims/1141925MCP
get_claim(id=1141925)