Association-Augmented Retrieval: Association̸ = Similarity: Learning Corpus-Specific Associations for Multi-Hop Retrieval

Dense retrieval systems rank passages by similarity to a query, but multi-hop questions require passages linked through reasoning chains rather than surface resemblance. We introduce Association-Augmented Retrieval (AAR), a lightweight transductive method that trains a small MLP (4.2M parameters) on passage co-occurrence annotations to learn associative relationships in embedding space. On HotpotQA, AAR improves Recall@5 by +8.6 points without evaluation-set tuning, with +28.5 points on hard questions where dense retrieval fails. On MuSiQue, AAR achieves +10.1 points. An inductive variant shows no significant improvement, indicating corpus-specific co-occurrence learning. Ablations confirm that association and similarity produce opposite effects: training on similar but non-associated pairs degrades retrieval, while only genuine co-occurrence structure yields gains. The method trains in two minutes, adds 3.7ms per query, and requires no LLM-based indexing. Code and results are available at https://github.com/EridosAI/AAR.