claim
active
claim:the-logit-monitor-outperforms-llm-judge-verbalization-monitoring-using-10-to-100-fewer-rolloutsThe logit monitor outperforms LLM-judge verbalization monitoring using 10× to 100× fewer rollouts
Central claim.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (7)
finding
- Key result.
- Additional dataset replication.
- A single sentence suffices to beat the judge.
- Prediction task result.
- Qualitative picture in Figure 1.
- Additional model replication.
- Baseline ceiling.
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.
- Positioning.
- Discussion positioning.
- Rollouts with and without verbalized eval awareness average to similar logit monitor curvesfinding0.766Figure 1 result.
- Automated logit weight prediction achieves 74% mean accuracy for features vs 58% for neurons vs 50% chancefinding0.766Automated interpretability of logit weights confirms feature downstream effects are more interpretable than neuron effects
- Prior finding showing scale-dependent self-awareness, consistent with the scale effect observed in the paper's Experiment 1
- Further studying the relation between activation probes and the logit monitor could help better understand activation probeshypothesis0.758Future work.
- In the 98-game slice, TrackerAgent had a higher win rate or TrueSkill than all LLMs except Gemini 3 Flash.
- Prompt-level effect is much larger.