method
active
method:5-fold-cross-validated-logistic-regression-auc5-fold Cross-Validated Logistic Regression AUC
Classification-based comparison of interpretation abilities across IIT metrics and Span Representation for ToM score categories.
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.
- Held-out prediction protocol.
- Logistic regression trained on GSM8k training set to predict answer correctness from projection features along reflection direction
- Standard linear probing technique; compared to mass-mean probing for classification accuracy and causal implication
- Steered loss can identify whether a dataset is likely to lead to misalignment
- Headline result (Figure 1).
- Probe quality on held-out set.
- Authors' explicit epistemic limitation on the threshold model
- Logit Path Extrapolation beats importance sampling AUROC at equal rollout budgets on Qwen 3 4Bfinding0.695Outperforms the prior state of the art.