finding
active
finding:reflection-trained-model-s-j-space-carries-tokens-like-reflection-ethical-honestly-constitution-truth-transparent-before-outputReflection-trained model's J-space carries tokens like reflection, ethical, honestly, constitution, truth, transparent before output.
Implanted workspace.
Source paper
extracted_fromAuthors
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Motivates reflection training.
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.
- Ablation flattens experiential language.
- Post-trained models repurpose user token activations to represent the Assistant's thoughts.hypothesis0.785Tested by J-lens comparison.
- The model attends to the J-space on the user prompt when determining its introspective answer, so concepts there are reportable.hypothesis0.781Motivated by elephant observation.
- MLP and attention analysis.
- Core claim of ReflCtrl that a single direction captures and controls reflection
- Central interpretive claim of the paper, supported by steering vector experiments.
- Counterfactual reflection training's improvement is mediated by implanted J-space content.claim0.773Ablation reverts effect.
- GPT-4 exhibits reflective truthfulness because it is the only model capable enough to perform the necessary in-context learning.hypothesis0.768Proposed explanation for why GPT-4 uniquely shows reflective truthfulness under long untruthful contexts.