Gibberish CoT as RL spandrel

Weird or illegible chains‑of‑thought in reasoning models may not be the actual 'reasoning' but vestigial token patterns reinforced by RL credit assignment. These strings can still be instrumentally useful—e.g., triggering internal passes—even if they look nonsensical to humans; removing or 'cleaning' them can slightly harm results. — This cautions policymakers and benchmarks against mandating legible CoT as a transparency fix, since doing so may worsen performance without improving true interpretability.