Self-reflection consumes 25-30% of total reasoning tokens empirically

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

• Up to 33.6% reasoning tokens saved on MMLU subsets with stepwise steering while maintaining accuracy in larger modelsfinding0.771
  Maximum token savings achieved by ReflCtrl on non-mathematical general reasoning tasks
• The underlying mechanism of self-reflection in reasoning LLMs is not yet well understoodquestion0.757
  Broad gap motivating the entire paper
• what is the underlying mechanism of self-reflection in reasoning LLMs?question0.744
  Open question motivating the entire paper; identified as not yet well understood
• Reflective reasoning requires late-stage integration of semantic and reasoning signals, hence reflection-related directions emerge more clearly in higher network layers.claim0.742
  Interpretive claim about the locus of reflection in transformer architecture.
• A linear reflection direction exists in reasoning LLMs' latent representation space that governs self-reflection behaviorclaim0.742
  Core claim of ReflCtrl that a single direction captures and controls reflection
• QwQ-32B on MATH-500: 21.0% reasoning token reduction at intervention strength -0.96 with only 0.34% accuracy lossfinding0.739
  Demonstrates reflection redundancy in stronger model on harder math benchmark
• 48 of 171 emotion probes individually significant at token 100 post-steeringfinding0.735
  Shows that causal steering effects persist over long ranges for a substantial fraction of emotion probes
• Reflection is not merely a behavioral artifact of prompting but a phenomenon encoded in the model's activation space.claim0.732
  Central interpretive claim of the paper, supported by steering vector experiments.