Multi-hop RAG got better by taking the bookkeeping away from the LLM
AspireAI community · arXiv
RAG handles multi-hop questions badly for two opposite reasons: graph-based methods retrieve once from the entities the question names and rarely recover the full evidence chain, while iterative methods recover it but pile many tasks onto the LLM - planning sub-questions, deciding dependencies, rewriting them as answers arrive, and tracking what is known. DeFT-RAG generates one self-contained sub-question at a time from the original question and the sub-questions already answered. Each sub-question names every entity it asks about, which removes the need to plan or track dependencies or rewrite anything, leaving the LLM one fine-grained task at a time. The remaining LLM work moves outside the LLM: rule-based updates to a QA table, and yes/no decisions made by a fine-tuned lightweight decision model. QA pairs are cached and reused when a later question asks something with the same meaning. On MuSiQue, 2WikiMultiHopQA and HotpotQA the authors report the best accuracy among thirteen baselines with both gpt-4o-mini and Qwen3.6 as backbone, while giving the LLM fewer input tokens than any other iterative method; with a shared cache over 1,000 MuSiQue questions, 23.1% of sub-questions came from cache and LLM calls dropped 12.9% without an accuracy cost.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.