Memory Augmentation Unlocks Efficient Chain-of-Thought Reasoning
Ranking
Overall
79
Content
95
Popularity
41
Observed public metrics from 1 member.
Merged summary
TL;DR - Memory-Augmented Compression retrieves reusable reasoning summaries as prefill context, enabling LLMs to produce shorter chain-of-thought traces without sacrificing as much accuracy. The training-free method improves reasoning performance over Chain-of-Draft while delivering 1.14–1.49× lower latency than standard CoT.
- Formalizes a “Context-Generation Substitution Law,” framing explicit context as a substitute for some decode-time reasoning generation.
- Builds memories from historical traces that capture reusable patterns, constraints, and critical operations rather than storing raw demonstrations.
- Improves over Chain-of-Draft by 21.4 points on GSM8K, 28.0 on MATH, 29.5 on BBH, and 6.61 on MMLU-Sci.
- Works with token-level, trace-level, and inference-state compression; analyses attribute gains to memory relevance rather than context length alone.
Sources (1)
Memory Augmentation Unlocks Efficient Chain-of-Thought Reasoning
Public signals
Semantic Scholar citations 0 · Semantic Scholar influential citations 0
TL;DR - Memory-Augmented Compression retrieves reusable reasoning summaries as prefill context, enabling LLMs to produce shorter chain-of-thought traces without sacrificing as much accuracy. The training-free method improves reasoning performance over Chain-of-Draft while delivering 1.14–1.49× lower latency than standard CoT.
- Formalizes a “Context-Generation Substitution Law,” framing explicit context as a substitute for some decode-time reasoning generation.
- Builds memories from historical traces that capture reusable patterns, constraints, and critical operations rather than storing raw demonstrations.
- Improves over Chain-of-Draft by 21.4 points on GSM8K, 28.0 on MATH, 29.5 on BBH, and 6.61 on MMLU-Sci.
- Works with token-level, trace-level, and inference-state compression; analyses attribute gains to memory relevance rather than context length alone.