CORE compresses retrieved context for retrieval-augmented generation by training a compressor against the downstream task rather than a generic text-quality signal. It addresses the production problem of long, noisy contexts that increase cost and degrade answers, while treating the answering model as a static black-box environment. section 13.0 section 13.4.1
CORE trains the compressor in two phases:
- Warm up the policy through knowledge distillation from high-quality summaries generated by DeepSeek-V3. section 13.4.1
- Refine the policy with reinforcement learning, using the downstream LLM’s exact match and F1 scores as task rewards. section 13.4.1
- Generate groups of candidate summaries and apply Group Relative Policy Optimization (GRPO), which normalizes each candidate’s reward against its group mean and standard deviation without a separate critic model. section 13.4.1
Three-percent compression
At a three percent compression ratio, CORE improves exact-match scores by 3.3 points over an uncompressed baseline. RECOMP and NoiseFilter-IB lose 3–5 points at similar compression levels, illustrating the tradeoff between generic compression quality and alignment with the answering model’s needs. section 13.4.1
