claim
active
claim:robust-interpretability-tools-require-inductive-biases-that-match-the-dataRobust interpretability tools require inductive biases that match the data.
Overall conclusion.
Source paper
extracted_from(2025) · Lubana, Ekdeep Singh · Rager, Can · Hindupur, Sai Sumedh R. · Costa, Valerie +12
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (6)
finding
- Fig. 9b.
- Fig. 9c, Fig. 37.
- App. E.
- Off-the-shelf SAEs gave wildly inconsistent results depending on source, so all SAEs were trained from scratch.associated_withApp. B.
- Prediction-only ablation shows block and event structure but more wound trajectories than full TFA.supportsApp. F.
- Fig. 8a.
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.
- Motivation for the release.
- Benchmark discipline.
- Proof of concept.
- Interpretability hypotheses are often conflated with conclusions and require extensive validationclaim0.782Validation is under-done.
- Benchmarks using models with known explanations.
- Forward-looking motivation.
- Interpretability tools and semantic search are key to winning the Marks et al. auditing gameclaim0.768Affordance insight.