finding
active
finding:logit-monitor-at-one-rollout-outperforms-verbalization-monitor-at-10-rollouts-predicting-which-fortress-prompts-elicit-verbalized-eval-awarenessLogit monitor at one rollout outperforms verbalization monitor at 10 rollouts predicting which Fortress prompts elicit verbalized eval awareness
Prediction task result.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
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.
- Rollout-level verbalization barely matters.
- On Qwen 3 32B with Fortress, logit monitor at one rollout outperforms verbalization monitor at 100 rolloutsfinding0.894Additional model replication.
- Prompt-level effect is much larger.
- Qualitative picture in Figure 1.
- Rollouts with and without verbalized eval awareness average to similar logit monitor curvesfinding0.864Figure 1 result.
- Additional dataset replication.
- Main interpretation.
- Robustness check.