New NNN Decoding Outperforms Dense Retrieval, Avoids Redundant Results

Current dense retrieval systems score each document independently via inner product with the query, ignoring corpus context. This often returns redundant, semantically similar results. A new paper from researchers introduces Non-negative Elastic Net (NNN) decoding, which treats retrieval as a joint decoding problem: documents are selected as a set that collectively reconstructs the query embedding through a sparse, non-negative linear combination. This ensures diversity by accounting for correlations between documents.

The team provides a theoretical separation: any query correctly handled by dense retrieval is also handled by NNN, and on corpora with correlated documents, NNN additionally succeeds where dense retrieval fails. Experiments on standard benchmarks show consistent improvements using frozen embeddings. More importantly, they developed an end-to-end training procedure that optimizes embeddings specifically for NNN decoding, producing significant gains across all metrics and benchmarks. This work establishes a new paradigm for leveraging dense embeddings beyond simple inner-product scoring, with implications for RAG systems, enterprise search, and any application requiring diverse, non-redundant retrieval results.