Overthinking can nearly double inference cost without improving accuracy

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

• Overthinkingconcept0.754
  Performance failure mode in recurrent networks from excessive computation iterations, relevant to looped LLM design
• Inference costconcept0.733
  Computational expense proportional to number of generated tokens, targeted for reduction by ReflCtrl
• Active inference treats exploration and exploitation as two different ways of tackling the same problem: minimizing uncertainty.claim0.728
  §2, comment on expected free energy decomposition.
• More inference compute amplifies both reflective capacity and safety gating; the contemplative prompt resolves gating by reframing self-referential probes.claim0.723
  Interpretation of Grok 4 vs Grok 4 Fast per-koan comparison
• H12: Inference compute adds to reflective capacity — higher compute budget produces higher reflective scores on the same weights.hypothesis0.719
  Exploratory hypothesis supported by Grok 4 vs Fast ~1pt difference
• In the case of AI, inferences from behavior to features of internal processes are often unreliable, as demonstrated by recent LLMs.claim0.712
  Justifies using internal indicators rather than behavioral tests for AI consciousness
• We hypothesize that explicitly instructing the model to evaluate the correctness of the given statement may change the geometry of truth directions.hypothesis0.711
  Motivating hypothesis for Section 5's investigation of prompt template effects.
• Models fine-tuned on human data show strong positive correlation between misalignment and incoherence; misalignment scores fall below 1% when incoherence threshold is appliedfinding0.707
  Human data fine-tuning effect is distinct from synthetic emergent misalignment and likely caused by off-policy training