framework
active
framework:saebenchSAEBench
Evaluation suite of eight metrics across four capabilities.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (9)
method
- k-Sparse ProbingusesConcept detection metric using top-k latents and linear probe.
- Novel metric extending SHIFT via zero ablation.
- Detection-based LLM judging of latent descriptions using gpt-4o-mini.
- RAVEL evaluationusesDisentanglement metric via latent interventions with Cause and Isolation metrics.
- Novel multiclass generalization of SCR.
- Conditional negative clamping on WMDP-bio with MMLU retention.
- Adapted absorption metric detecting partial and shared absorption across layers.
- Loss RecoveredusesReconstruction fidelity score based on cross-entropy.
- PCA BaselineusesPCA components treated as SAE latents for comparison.
Concepts (1)
concept
- Sparsity-Fidelity TradeoffextendsTraditional proxy evaluation criterion for SAEs.
Artifacts (2)
artifact
- Interactive viewer of metrics across hundreds of SAEs.
- Open-source code and 200+ SAEs.
Related by similarity (7)
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.
- Evaluated architecture.
- Japanese aesthetic concept of rustic, imperfect, transient beauty; Alexander equates it with the rubbed-in, used quality necessary for belonging.
- Evaluated architecture (Gao et al.).
- Framework from Wu et al. for automated contrastive pair generation for arbitrary concepts; most similar prior work to persona vector pipeline
- Interpretable features extracted by sparse autoencoders used as steering targets in this study
- Hybrid SAE jointly training turn-averaged inner and per-token outer partitions.
- Project management framework within Agile, using sprints and daily stand-ups, traced to Alexander's influence.