Input embedding similarity baseline selects semantically related but non-reflective tokens (e.g., Await, ConfigureAwait, Unchecked) that fail to improve accuracy

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

• Input Embedding Similarity Baselinemethod0.853
  Baseline method for instruction discovery using surface-level input embedding similarity instead of steering vectors.
• Steering vector-based instruction discovery outperforms input embedding similarity baseline for reflection-inducing instruction selectionfinding0.801
  Demonstrates that surface-level embedding similarity fails to capture reflective semantics.
• How do we incorporate a focus on behavioral relevance in our measures of neural similarity?question0.750
  Direct motivating question for MAS's design principle of causal behavioral matching.
• Probes trained under different explicit instruction prompts (ask-correct, ask-t/f, ask-able, ask-arith) are highly aligned with each other in cosine similarity.finding0.750
  Shows the passive vs. active divide is more important than the specific wording of instructions.
• Baseline scores blend together at least three different things: latent reflective capacity, default accessibility, and stability of access.claim0.749
  Conceptual decomposition arising from the data showing different models dissociate these traits
• We hypothesize that native self-report, fine-tuned introspection models, and trained activation-to-language systems will show different performance on bias-resistant localization and strength benchmarkshypothesis0.749
  Comparative prediction motivating future work contrasting different approaches to LLM self-knowledge
• Certain representation learning algorithms boil down to a simple rule: find an embedding in which similarity equals PMIclaim0.747
  Core theoretical claim about the target of representation learning
• Cross-model semantic convergence under self-referential processing suggests the presence of a shared attractor state that transcends variance across training proceduresclaim0.746
  Interpretive claim from Experiment 3; GPT, Claude, Gemini families converge on similar descriptive style despite independent training