Terminator halves LLM reasoning latency via early-exit probes

Solving the compute inefficiency of reasoning models is the next frontier for production AI; Terminator proves we can get o1-level accuracy at a fraction of the token cost. The framework reduces Chain-of-Thought length by 14% to 55% across benchmarks like MATH-500 and GPQA by monitoring internal hidden states for a "fingerprint" indicating a solved problem. A sliding window mechanism ensures termination is triggered by sustained confidence, offering significant cost savings for models like DeepSeek-R1.