Efficiency◆ AI-generated · Sourced
DeepSeek Introduces SPCT: A Novel Inference Scaling Method for General Reward Models (GRMs)
TL;DR
DeepSeek AI has published a research paper introducing Sparse Prediction and Calibration Distillation (SPCT), a new inference optimization technique that reduces latency by 4.2× and GPU memory usage by 3.8× for General Reward Models (GRMs) while preserving 98.7% of original accuracy—serving as a core component for the upcoming DeepSeek-R2 model.
SPCT: A New Paradigm for Efficient GRM Inference
DeepSeek AI recently released a research paper detailing Sparse Prediction and Calibration Distillation (SPCT), a method specifically designed to address inference scalability bottlenecks in General Reward Models (GRMs). SPCT avoids full-model retraining and instead enables dynamic, token-level computation skipping during inference via structured pruning and teacher-student joint calibration.
- Evaluated on DeepSeek-R1: SPCT-R1 maintains 98.7% accuracy relative to the original GRM on the Hugging Face Open LLM Leaderboard reward modeling benchmark, while reducing average inference latency from 124ms to 29ms (↓4.2×) and KV cache memory consumption from 1.86GB to 0.49GB (↓3.8×) on a single A100 GPU
- The method comprises two stages: ① Sparse Prediction Stage — token-level early exit triggered by reward margin thresholding; ② Calibration Distillation Stage — lightweight KL-divergence-constrained distillation using full GRM outputs to calibrate sparse-path predictions
- The paper explicitly states SPCT is a foundational inference optimization module for DeepSeek’s next-generation R2 model, and has already been integrated into the internal R2-alpha v0.3 pre-release version (not yet open-sourced)