Reflection does not only emerge in SFT or RL stages but arises earlier during pre-training.

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

• Reflection-inducing directions emerge more clearly in higher layers (ℓ>5) for both models and datasetsfinding0.800
  Empirical observation about which network layers encode reflection-relevant information.
• Post-training stages (SFT, DPO, RLVR) progressively reduce output diversity, with direct prompting showing the most severe mode collapseclaim0.790
  Empirical finding from Tulu-70B ablation study across post-training stages
• Reflection is not merely a behavioral artifact of prompting but a phenomenon encoded in the model's activation space.claim0.783
  Central interpretive claim of the paper, supported by steering vector experiments.
• Reflective reasoning requires late-stage integration of semantic and reasoning signals, hence reflection-related directions emerge more clearly in higher network layers.claim0.779
  Interpretive claim about the locus of reflection in transformer architecture.
• The behavior of the initial model may be more impactful in determining emergent misalignment for on-policy methods (RL) than for off-policy methods (SFT)claim0.775
  RL shows stronger safety training effect while SFT does not, suggesting on-policy methods are more sensitive to initial model state
• When does the model initiate reflection during its reasoning process?question0.771
  First central research question motivating ReflCtrl investigation
• Presence of safety training during SFT does not meaningfully increase or decrease emergent misalignmentfinding0.769
  Helpful-only models exhibit same degree of emergent misalignment as safety-trained counterparts under SFT
• Introspective signals appear in middle layers but are suppressed by later post-training-shaped layers.finding0.766
  Mechanistic finding by Lindsey (2026) explaining how contemplative prompt may work: enables mid-layer introspection to reach output.