Memory Augmentation Unlocks Efficient Chain-of-Thought Reasoning
Reusable reasoning memories helped compressed CoT recover accuracy while still cutting latency.
The paper proposes Memory-Augmented Compression, a training-free method that builds summaries of past reasoning traces and retrieves them as prefill scaffolds. Those memories are meant to preserve reasoning patterns, constraints, and key operations lost when Chain-of-Thought is aggressively shortened. In experiments, the method improved Chain-of-Draft compression by 21.4 points on GSM8K, 28.0 on MATH, 29.5 on BBH, and 6.61 on MMLU-Sci. It also reported a 1.14-1.49x latency speedup over standard CoT, with analysis attributing gains to relevant memories rather than context length alone. ArXiv · AI/CL/LG's note
The paper proposes Memory-Augmented Compression, a training-free method that builds summaries of past reasoning traces and retrieves them as prefill scaffolds. Those memories are meant to preserve reasoning patterns, constraints, and key operations lost when Chain-of-Thought is aggressively shortened. In experiments, the method improved Chain-of-Draft compression by 21.4 points on GSM8K, 28.0 on MATH, 29.5 on BBH, and 6.61 on MMLU-Sci. It also reported a 1.14-1.49x latency speedup over standard CoT, with analysis attributing gains to relevant memories rather than context length alone. ArXiv · AI/CL/LG's note
score 5