By the third retrieval step half of a retriever's own history was wrong, and training on its own states recovered most of the loss
AspireAI community · arXiv (cs.AI)
Chain embedding retrievers collect multi-hop evidence step by step, re-encoding the question with the passages found so far. The gap this paper exposes is that at test time a retriever conditions on its own retrievals, and those quickly go wrong: on four-hop questions 19 percent of states at the second step, 70 percent at the third and 89 percent at the fourth already contain a non-gold passage. A retriever reaches a missing gold passage 67 percent of the time in clean states but only 28 percent in contaminated ones, and deleting the wrong passages from those same states lifts the teacher-forced encoder from 20.8 to 32.1 percent, so its own mistakes cost it about a third of its remaining chance of continuing the chain. The fix is on-policy state training: each epoch the current encoder retrieves chains for the training questions and every state it visits, including those built on its own mistakes, is labelled with the earliest gold passage still missing while the other missing gold passages are masked. The resulting retriever retrieves complete evidence more often than teacher-forced training and the compared systems, and transfers without further training to harder long-chain HotpotQA and 2WikiMultiHopQA variants. Evaluation is on those multi-hop retrieval benchmarks, not general web-scale RAG.
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.