Self-reflection consumes 25-30% of total reasoning tokens in reasoning LLMs

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

• The underlying mechanism of self-reflection in reasoning LLMs is not yet well understoodquestion0.834
  Broad gap motivating the entire paper
• what is the underlying mechanism of self-reflection in reasoning LLMs?question0.820
  Open question motivating the entire paper; identified as not yet well understood
• A linear reflection direction exists in reasoning LLMs' latent representation space that governs self-reflection behaviorclaim0.799
  Core claim of ReflCtrl that a single direction captures and controls reflection
• Reasoning LLMs trigger reflection when their internal uncertainty is highhypothesis0.790
  Core hypothesis linking internal uncertainty to self-reflection behavior, tested via probing experiments
• Up to 33.6% reasoning tokens saved on MMLU subsets with stepwise steering while maintaining accuracy in larger modelsfinding0.780
  Maximum token savings achieved by ReflCtrl on non-mathematical general reasoning tasks
• When LLMs produce experience claims under self-reference, is this sophisticated simulation or genuine self-representation, and how would we tell the difference?question0.756
  The core interpretive question the paper narrows but cannot definitively answer
• Reflective reasoning requires late-stage integration of semantic and reasoning signals, hence reflection-related directions emerge more clearly in higher network layers.claim0.748
  Interpretive claim about the locus of reflection in transformer architecture.
• QwQ-32B on MATH-500: 21.0% reasoning token reduction at intervention strength -0.96 with only 0.34% accuracy lossfinding0.745
  Demonstrates reflection redundancy in stronger model on harder math benchmark