method
active
method:spurious-correlation-removal-scrSpurious Correlation Removal (SCR)
Novel metric extending SHIFT via zero ablation.
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (1)
framework
- SAEBenchusesEvaluation suite of eight metrics across four capabilities.
Concepts (1)
concept
- Independent concepts are encoded by separate latents.
Methods (3)
method
- Novel multiclass generalization of SCR.
- Stronger test of concept isolation than steering.
- SHIFTimplementsMethod from sparse feature circuits that SCR automates.
Datasets (2)
dataset
- Amazon ReviewsusesUsed for probing, SCR, TPP.
- Bias in BiosusesUsed for sparse probing, SCR, TPP.
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.
- Pearson correlation of feature activations across 40M tokens used to measure feature similarity and universality across models
- Inverse SCR scaling is simply due to a fixed number of ablated latents being a smaller fraction of larger SAEshypothesis0.722Alternative hypothesis tested and rejected.
- Ensuing need for active, intelligent decoding on the expansion side.
- Head-level causal evidence.
- m=2 experiments.
- Connects the model's behavior to Zaadnoordijk and Bayne's taxonomy of intentional agency
- Hippocampal neural coding phenomenon reproduced by active inference.
- The elimination of the self-environment partition prior is proposed as a formal model of the Buddhist realisation of emptiness.