finding
active
finding:rl-refined-oracle-lens-explains-31-of-whitened-activation-varianceRL-refined oracle lens explains 31% of whitened activation variance.
Oracle lens.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Limitation.
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.
- Demonstrates non-robust heuristics installed by RL; slight prompt variation recovers large gap
- Tuned lens skips to output answers and is best next-token predictor while J-lens has highest KL.finding0.724Comparison.
- Table 2 first row; reward shaping section.
- Figure 4 and discussion in §3.
- Deceptive RL baseline agents have lower mean neural self-other overlap than honest baseline agentsclaim0.715Core empirical prediction tested in RL experiments, confirmed by 100% classification accuracy
- Comparative claim against the NoWait baseline method
- Bayesian model-based RL achieved average score 99.76 [99.45, 100.00] in deterministic FrozenLake.finding0.715Table 1.