Every Token Leaves a Ripple in the Stream of Thought: Eliciting Model-Internal Token Saliency for Chain-of-Thought Compression
MIST scores reasoning tokens by how much their internal contribution changes the model’s answer likelihood.
The paper frames chain-of-thought compression as a token selection problem, with long traces raising inference cost. Its method measures two signals: necessity, when removing a token’s contribution lowers answer likelihood, and sufficiency, when providing that contribution alone raises it. The combined score is used to prune reasoning chains for shorter adaptation traces. The authors report consistent gains over baselines across four reasoning benchmarks and four models. ArXiv · AI/CL/LG's note
The paper frames chain-of-thought compression as a token selection problem, with long traces raising inference cost. Its method measures two signals: necessity, when removing a token’s contribution lowers answer likelihood, and sufficiency, when providing that contribution alone raises it. The combined score is used to prune reasoning chains for shorter adaptation traces. The authors report consistent gains over baselines across four reasoning benchmarks and four models. ArXiv · AI/CL/LG's note
score 5