liveBook · Manning

CORE

As explained in Retrieval Augmented Generation, The Foundational Ideas

CORE is a retrieval-augmented generation compression framework that trains a compressor with reinforcement learning against downstream task rewards.

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

How Retrieval Augmented Generation, The Foundational Ideas covers it

How deep it goes
Details two training phases and GRPO’s reward normalization, but does not describe compressor architecture or implementation code.
Examples
Uses three-percent compression, exact-match gains, and comparisons with uncompressed baselines, RECOMP, and NoiseFilter-IB.
What it assumes
Assumes a downstream LLM and task-level rewards, while avoiding human-annotated compression labels through distillation and reinforcement learning.
What you can do after
Evaluate task-oriented compressors by comparing downstream exact-match and F1 rewards at a fixed compression ratio.
How it connects
Connects context compression with retrieval-augmented generation, knowledge distillation, reinforcement learning, GRPO, and RECOMP.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Retrieval Augmented Generation, The Foundational Ideas. The section links point to the parts of the book it draws on.