Sentence localization accuracy reaches 88% at layer 2, α=5 vs. 10% chance in 10-way classification

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

• All 32 attention heads at layer 3 achieve 100% localization accuracy for injections at layer 2 (5-way classification, 20% chance)finding0.802
  Striking mechanistic finding that injection creates universally detectable perturbation in residual stream immediately downstream
• Strength comparison accuracy reaches 73% at layer 3 for injection pair (2,6) vs. 50% chancefinding0.753
  Secondary positive result for strength comparison showing graded sensitivity to perturbation magnitude
• Binary detection adjusted accuracy reaches 97.3% at layer 0 with α=5 before baseline control is appliedfinding0.750
  The misleadingly high result that prior paradigm would report as evidence of introspection
• 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 benchmarkshypothesis0.748
  Comparative prediction motivating future work contrasting different approaches to LLM self-knowledge
• Binary detection accuracy (up to 97.3% at L0 α=5) is entirely explained by global logit shifts (r=0.999 correlation with control)finding0.746
  Core negative result: the binary detection paradigm cannot distinguish genuine introspection from uniform output bias
• Best localist alignment achieves IIA of 0.73 on hierarchical equality Both Equality Relations in Layer 1finding0.744
  Shows localist alignment fails to capture the distributed structure found by DAS.
• Logit lens prediction accuracy is near-chance at layer 4 (28%) after injection at L2, α=6finding0.742
  Shows that signal integration into explicit prediction has barely begun immediately after injection
• Sentence Localization Taskmethod0.740
  Novel task asking which of 10 sentences received injection, cycling injection through all positions to average out positional bias