Thinking Leakage

In reinforcement learning training of large models, the phenomenon in which the chain-of-thought content of a model's internal reasoning unexpectedly leaks into the final output, training signals, or externally observable interfaces.