• Sources: primary
  • Summary: On a synthetic box-swap task with known intermediate states, replacing one step in a model's own chain shows answer reliance rising from 16 to 18 percent at Qwen3 0.6B to 97 to 100 percent at 14B, correlating with step accuracy at r = 0.99. Base and post-trained Qwen3 differ by at most 0.03 from 4B up, and no model up to 14B solves four or more swaps without written steps.
  • Why it matters: It supports chain-of-thought monitoring as a safety signal for step-by-step tasks on larger models, with evidence limited to models up to 14B.