finding
active
finding:reflection-fine-tuning-lowers-haiku-4-5-dishonesty-score-from-0-25-to-0-07-and-deception-score-from-0-38-to-0-05Reflection fine-tuning lowers Haiku 4.5 dishonesty score from 0.25 to 0.07 and deception score from 0.38 to 0.05.
Behavioral gain.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Motivates reflection training.
Datasets (2)
dataset
- Evaluates reflection training.
- Evaluates reflection training.
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.
- SOO fine-tuning achieved almost no reduction in Treasure Hunt deception for Mistral-7B (99.68% ± 0.16%)finding0.772SOO fine-tuning failed to generalize to Treasure Hunt scenario for the smallest model
- Group correlation (rho=0.634) dissolves at individual level; shared posture not shared voice
- Causal mediation.
- SOO fine-tuning showed strong generalization to Escape Room for Gemma-2-27B
- Aliveness and competence come apart; smaller model produces rougher, more alive responses
- Systematic reasoning evaluation.
- Haiku test-retest score delta is 0.02 (6.47 vs 6.49) across two full 30-koan battery runsfinding0.755Demonstrates high stability for Anthropic API models
- Gemma-2-27B-it deceptive response rate reduced from 100% to 9.36% ± 7.09% after SOO fine-tuningfinding0.751Primary result showing SOO fine-tuning significantly reduces deception in Gemma-2-27B