RECAP: Calibrating Low-Bit Reasoning Models on the History They Read
We present RECAP, a post-training quantization method that calibrates reasoning models on the history they read.
ProofPaper ↗
Key points
- Low-bit quantization reduces the cost of serving reasoning models but can substantially degrade long chains of thought.
- Even when this context is retained, cache reconstruction error alone does not capture how quantization errors affect subsequent attention.
- Dynamic-prefix calibration separates long historical context from short-block reconstruction.
- These results highlight the importance of preserving historical dependencies during calibration for low-bit reasoning.
Sources (1)
- [1]RECAP: Calibrating Low-Bit Reasoning Models on the History They ReadarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 01:10 AM
We present RECAP, a post-training quantization method that calibrates reasoning models on the history they read.
Low-bit quantization reduces the cost of serving reasoning models but can substantially degrade long chains of thought.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 10, 2026vllm-sr/d3
- Oct 10, 2026SAGE: Pushing 2-Bit Vector-Quantized LLMs toward Real-World Usability
- Oct 10, 2026KV4-RL: Unlocking Faster Asynchronous Rollouts for RLVR with INT4 KV Cache Quantization
- Oct 9, 2026Microsoft's Decision-1 model enters the fast-growing AI decision model race
- Oct 9, 2026Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash
- Oct 9, 2026The Thinking Leakage in Reinforcement Learning for LLMs