claim
active
claim:interpretability-can-turn-post-training-into-auditing-and-sculpting-the-learning-signalInterpretability can turn post-training into auditing and sculpting the learning signal
Closing thesis.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Sidharth Baskaran · Max Loeffler +13
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (3)
finding
- Feature-conditioned pipeline predicts DPO-induced feature change with R²=0.9 on Llama-3.1-8B/DolcisupportsFig. 6 validation.
- Fig. 13, App. J.4.
- Prompt-conditioned pipeline predicts DPO-induced feature change with R²=0.58 on Llama-3.1-8B/DolcisupportsFig. 7 validation.
Questions (1)
question
- Motivating 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.
- Intrinsic interpretability section.
- Base pretrained models show high false positive rates and achieve no net task performance on concept injection detection; post-training essential for introspection.
- Author interpretation of selectivity results showing DAS advantage diminishes when controlling for expressivity
- Finding that base models have high false positives and no net positive performance.
- Benchmark discipline.
- Second falsifiable prediction linking objective function structure to valence profile
- Proof of concept.
- Interpretability-encouraging instructions moderately boost win rate by surfacing top features earlierfinding0.778Appendix experiment.