finding
active
finding:kl-regularization-kl-0-03-from-base-preserves-behavior-but-does-not-improve-counterfactual-evaluation-performanceKL regularization (KL 0.03 from base) preserves behavior but does not improve counterfactual evaluation performance
Preliminary experiment.
Source paper
extracted_from(2026) · Adam Karvonen · Euan Ong · Subhash Kantamneni · Samuel Marks
Neighborhood — ranked by edge-count
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.
- KL penalty to base model so self-explanation stays about the same model.
- Strong empirical evidence that VS recovers pretraining distribution while direct prompting collapses
- Shows VS substantially better approximates the pretraining distribution than baseline methods
- Theoretical interpretation.
- Base failure mode.
- Sparsity penalty weighting by sum of per-layer decoder norms, enabling comparability with per-layer SAEs.
- We hypothesize that degraded generalization on benchmarks like MMLU may reflect the computational demands of the tasks.hypothesis0.747Connecting the paper's task-difficulty findings to prior observations of weak generalization on complex QA benchmarks.
- Acknowledged limitation.