concept
active
concept:interpretability-illusionInterpretability Illusion
Cases where subspace interventions change model behaviour through parallel pathways rather than the target feature
Neighborhood — ranked by edge-count
Papers (2)
paper
Claims (1)
claim
- Interpretability hypotheses are often conflated with conclusions and require extensive validationsupportsValidation is under-done.
Concepts (1)
concept
- interpretabilityrelated_toThe capability to explain model predictions; a central theme of the paper, with disruption profiles as vehicle.
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.
- Socio-philosophical open question.
- Proposed paradigm for evaluating interpretability work through empirical falsifiability rather than benchmarks or user studies
- Method using large language models (Claude) to generate and test explanations of features at scale
- Advantage of DiffLogic CA over NCA — learned rules are pure binary logic circuits that can be visualized and analyzed
- Second approach: propose concepts and find components corresponding to them.
- Frames model interpretability like biology, features as cells, tools as microscopes.
- Ian Goodfellow quote used to illustrate the pre-paradigmatic state of interpretability research
- An interpretability paradigm that explains computation in the model's own terms, rather than imposing top-down abstractions; VPD aims to realize this.