concept
active
concept:exponential-tilt-of-kl-regularized-optimal-policyExponential tilt of KL-regularized optimal policy
Optimal policy proportional to base policy times exp(reward/β).
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (1)
framework
- Unifying framework: remove or amplify a concept's term in the exponential-tilt reward.
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.
- Preliminary experiment.
- Proposed conjecture in §4.3.1.
- RL algorithm used for training models to comply with the conflicting objective
- EI and normalized EI could serve as a unified metric for out-of-distribution generalization.claim0.695Conjecture that maximizing EI yields causal representations invariant to distribution shifts.
- Sparsity penalty weighting by sum of per-layer decoder norms, enabling comparability with per-layer SAEs.
- Setting αk to the maximum gradient norm performs best among tested strategies on NYUv2 (Figure 6).finding0.688Sensitivity analysis for gradient normalization scaling factor.
- Calibration finding for choosing the activation cap threshold