claim
active
claim:experiments-provide-initial-evidence-that-model-diff-amplification-is-a-practical-tool-for-post-training-problemsExperiments provide initial evidence that model diff amplification is a practical tool for post-training problems
Conclusion.
Source paper
extracted_fromThomas McGrath
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Appendix replication across prompts.
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.
- Scope limitation.
- Monitoring use case.
- Positioning.
- Core method: sample from logits_after + α(logits_after − logits_before) to surface rare behaviors.
- DAS finds causal effect at all training timesteps including when model is just initialisedfinding0.796Corroborates Wu et al. 2023 finding that DAS expressivity inflates causal effect estimates
- Core motivation.
- Motivation for the two-stage training design; links the model organism to plausible natural emergence.
- Finding that base models have high false positives and no net positive performance.