finding
active
finding:rlfr-with-best-of-32-achieves-58-0-1-4-overall-hallucination-reduction-on-gemma-3-12b-it-with-monitoring-harnessRLFR with best-of-32 achieves 58.0%±1.4% overall hallucination reduction on Gemma-3-12B-IT with monitoring harness
Topline result.
Source paper
extracted_from(2026) · Aaditya Vikram Prasad · Connor Watts · Jack Merullo · Dhruvil Gala +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Central thesis.
- Running probes on frozen parameters mitigates monitor evasion because it is easier to learn the behavior than evade the probe.associated_withLimitation discussion.
Findings (3)
finding
- Decomposition component.
- Decomposition component.
- Component of decomposition.
Questions (1)
question
- Can we transform a model's features into an inexpensive, dense signal for learning open-ended tasks?answered_byMotivating question.
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.
- Experiment 2 result showing large Gemma model supports high-dimensional truth cones
- Gemma-2-27B-it deceptive response rate reduced from 100% to 9.36% ± 7.09% after SOO fine-tuningfinding0.767Primary result showing SOO fine-tuning significantly reduces deception in Gemma-2-27B
- Small Gemma model shows severe ASR degradation at higher cone dimensions
- Ablation removing inline interventions.
- Ablation of BoN.
- Exception to the general Head Cor superiority, suggesting safety personas involve circuits beyond style modulation
- Even stylistic distribution.
- Smaller models show non-monotonic and diminished ASR with increasing cone dimensionality