kapynResearch

CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning

CLaRa unifies retrieval and generation for LLMs by compressing information into a continuous latent space. This framework jointly optimizes retrieval and generation, aiming to overcome limitations of existing RAG methods with long contexts and separate optimization. CLaRa introduces SCP, a data synthesis framework for creating semantically rich compressed vectors that preserve key information and shorten document inputs.

Apple ML Research·Jul 15, 2026

Opening Kapyn…