JKO-RAG reranking beats cross-encoders, cuts distractors by 2x
Wasserstein gradient flow boosts retrieval robustness 22–38% against paraphrases
A new paper by Levi Segal and Murari Ambati introduces JKO-RAG, a reranking approach that reframes passage selection as minimizing a free-energy functional F(p) = relevance + entropy + redundancy under Wasserstein-2 gradient flow via the Jordan–Kinderlehrer–Otto proximal scheme. Traditional RAG pipelines return a ranked list, but the downstream language model conditions on a set — causing a mismatch. JKO-RAG uses a ground metric based on cosine similarity to capture semantic geometry, and the authors derive a linear-response theory showing that Wasserstein's geometry-aware Hessian dampens mass transport from query paraphrases, while KL's diagonal Hessian does not.
Empirically, JKO-RAG outperforms cross-encoders across all five BEIR benchmarks. The decisive advantage is robustness: 22–38% more stable under paraphrase variation and 2x fewer leaked distractors. The paper also introduces four extensions: NM-JKO (learned ground metric), BW-JKO (Wasserstein–KL interpolation), SAM-JKO (2x speedup via sharpness-aware minimization), and DualRank (using optimal transport dual potentials as confidence signals). The authors verify their theory through free-energy descent, frequency-resolved perturbation response, step-size dependence, and certified-radius analysis.
This work challenges the standard ranked-list assumption in retrieval-augmented generation, offering a principled geometric solution that consistently beats strong baselines. The theoretical foundation — linking Wasserstein gradient flow to retrieval — provides a new lens for understanding why certain reranking strategies work, with practical implications for production RAG systems that must handle diverse user queries and paraphrases without degrading response quality.
- JKO-RAG reframes reranking as Wasserstein-2 gradient flow minimizing a free-energy functional (relevance + entropy + redundancy), outperforming cross-encoders on all 5 BEIR benchmarks
- 22–38% more stable under query paraphrases and 2x fewer leaked distractors compared to cross-encoders
- Linear-response theory explains Wasserstein's geometry-aware Hessian dampens paraphrase-induced mass transport; four extensions include learned metrics, interpolation, 2x speedup, and OT-based confidence signals
Why It Matters
JKO-RAG provides a theoretically grounded reranking method that significantly improves RAG robustness against query variation, a critical bottleneck in production retrieval systems.