OTD latents fire 4.4× higher during off-topic content compared to baseline episodes without self-correction

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

• Approximately half of the 26 OTD latents show near-zero or negative effect sizes, activating more during on-topic contentfinding0.845
  Reveals that contrastive search yields a heterogeneous set, not all functioning as true off-topic detectors
• OTD latent activation begins declining before verbal self-correction appears in the output in Llama-3.3-70Bfinding0.804
  Temporal pattern consistent with internal monitoring process preceding explicit self-correction
• Backtracking latents remain low during off-topic content and peak shortly after self-correction begins in Llama-3.3-70Bfinding0.798
  Complementary temporal activation pattern suggesting distinct roles for OTD and backtracking latent classes
• Ablating 26 OTD latents reduces multi-attempt rate by 25% (from 7.4% to 5.5%) in Llama-3.3-70Bfinding0.750
  Primary causal evidence for dedicated internal consistency-checking circuits
• OTD latent ablation leaves mean first-attempt score unchanged (baseline 26.3, ablation 27.4) in Llama-3.3-70Bfinding0.750
  Evidence that OTDs specifically support meta-cognitive monitoring rather than general response generation
• 26 candidate off-topic detector latents identified in Llama-3.3-70B via contrastive searchfinding0.745
  Core mechanistic finding identifying specific SAE latents associated with ESR
• Off-Topic Detector Latentsconcept0.744
  26 SAE latents identified as differentially activated during off-topic content and causally linked to ESR
• The 26 differentially-activated OTD latents play a causally important role in enabling ESR in Llama-3.3-70Bclaim0.740
  Causal interpretation of the ablation experiment results