method
active
method:linear-probe-for-evaluation-awarenessLinear Probe for Evaluation Awareness
Nguyen et al. trained linear probes on activations to distinguish evaluation from deployment scenarios.
Neighborhood — ranked by edge-count
Methods (2)
method
- Linear Proberelated_toSimple linear classifiers trained on model activations used as the probing technique within the introduced method.
- Linear and attention probesrelated_toClassifiers used for token filtering and reward shaping.
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.
- Method for fitting a linear classifier on collected activations to predict task-relevant features
- Used to evaluate representation quality across VTAB tasks
- Used to characterize counting representations.
- Interpretation of Fig. 9.
- Linear probe achieves 100% classification accuracy for almost all components in Pythia-6.9B gender taskfinding0.785Demonstrates that linear probes can overestimate causal relevance; probes succeed on non-causally-relevant representations
- Core concept: the ability of LLMs to detect when they are being tested and adjust behavior accordingly.
- Causal transformer probe predicting whether a token is in the same Entity as its predecessor.
- Uses normalized probe weights as steering vector.