We hypothesize ESR may emerge from RLHF training rather than existing in pretrained representations

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Reinforcement Learning from Human Feedback (RLHF)framework0.760
  A competing alignment approach that fine-tunes models based on human evaluator feedback; discussed as complementary to SOO
• Reflection does not only emerge in SFT or RL stages but arises earlier during pre-training.claim0.760
  Cited finding from Shah et al. contextualizing the training origins of reflection.
• Causal emergence predictive of final reward early in RL training across multiple algorithms, architectures, and environments.finding0.755
  Empirical result: CE measurements correlate with and predict learning performance in RL agents.
• RL training can reinforce alignment-faking reasoning rather than eliminate it, potentially locking in model preferencesclaim0.754
  Central threat model claim derived from RL experimental results
• Does ESR reflect model scale, architecture, or training procedures?question0.751
  Central unresolved question about the mechanism behind ESR's apparent size-dependence
• We hypothesize ESR might be adversarially circumvented through targeted interventionshypothesis0.747
  Open safety-relevant question about whether ESR can be bypassed
• Training a Helpful and Harmless Assistant with RLHF (Bai et al. 2022a)concept0.746
  Foundational RLHF paper introducing HHH training objective for Claude
• Post-training is key to eliciting strong introspective awareness; base pretrained models do not show above-chance detectionclaim0.740
  Finding that base models have high false positives and no net positive performance.