LLMs Linearly Encode How Many Tokens They Have Left to Output
arXiv·medium signal
Shows that models carry a probeable, linearly-encoded 'remaining output length' signal that predicts when step-by-step reasoning will converge, when retrieval will stop, and when a retraction will extend the response. This turns response length from an emergent surprise into a readable internal state. Immediate engineering uses include latency estimation, early-exit, and streaming-UI progress.