Answers are grounded in this paper and cite their sources.
Retrieval-Augmented Generation (RAG) for multi-hop question answering suffers from noisy context and weak retrieval–generation alignment. Apple’s Continuous Latent Reasoning (CLaRa) improves efficiency by compressing retrieved documents into dense latent memory tokens, but this compression obscures the structural relationships between documents required for multi-hop reasoning. We propose GraphCLaRa, a lightweight (∼1.3M-parameter) graph-based reranker for a frozen CLaRa reader. It constructs a query-aware document graph, propagates evidence with a Graph Attention Network (GAT), and selects a compact document subset with an adaptive context-budget head. Across four single- and multihop QA benchmarks, GraphCLaRa improves answer F1 on three of four datasets (up to +6.6) while using fewer documents, and trails by only 1.1 F1 on the retrieval-saturated benchmark. Our analysis shows that graph-guided selection consistently helps when retrieval leaves room for improvement, whereas on saturated retrieval tasks compression can only discard evidence. These results demonstrate that graph-based evidence propagation complements latent-memory compression.
No takes yet. Share an insight, caveat, or question.
Tran et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: