arXiv 2511.18659

CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning

By Jie He, Richard He Bai, et al.

Published 2025-11-24

Wiki summary

Explore the paper's summary, context, and related research on Papiers.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) with external knowledge but still suffers from long contexts and disjoint retrieval-generation optimization. In this work, we propose CLaRa (Continuous Latent Reasoning), a unified framework that performs embedding-based compression and joint optimization in a shared continuous space. To obtain semantically rich and retrievable compressed vec…

View the original paper on arXiv