ElicitR: Unlocking Latent Reasoning in Dense Retrievers via Generative Regularization

Reasoning-intensive retrieval is increasingly important for downstream applications, requiring more than lexical overlap or coarse semantic matching. While prior work mainly relies on Language Models (LMs) to synthesize reasoning-oriented supervision, we posit that it is already latent in LM-based retrievers but suppressed by contrastive overfitting. To elicit this latent reasoning, we introduce ElicitR, a retriever–LM framework with generative regularization that captures nuanced relationships among a query and its candidate documents beyond binary relevance. Concretely, alongside contrastive learning, we regularize the retriever by co-training a small LM on query–positive–negative batches. Next token prediction (NTP) for each text is conditioned on its prefix and the other in-batch texts, with cross-text conditioning weighted by retriever-computed similarities. Using MS MARCO as the only paired query-document supervision and a 135M LM for generative regularization with unlabeled raw-text initialization, ElicitR consistently improves BRIGHT by 16-29% relative across 0.1B–3B retriever scales while maintaining performance on BEIR. At 3B, ElicitR reaches an nDCG@10 of 23.1, substantially outperforming larger models trained with far more curated pairs and proprietary APIs. Further analyses show that ElicitR prevents overfitting, improves retrieval calibration, and remains robust to batch sizes, supporting its practicality.

Some search questions can't be answered by matching keywords; finding the right document takes a few steps of reasoning. The popular way to build search tools that can do this is to manufacture huge amounts of artificial training examples, but that approach is costly and hard to check. We noticed something stranger. These search tools are built from AI language models that can already reason, yet ordinary training quietly buries that ability, like over-rehearsing one trick until you forget everything else. So instead of adding reasoning from the outside, we tried to stop training from destroying it. Our method, ElicitR, attaches a tiny helper model during training that nudges the search tool to capture the rich relationships between a question and its documents, rather than collapsing them into a crude relevant-or-not label. The helper is thrown away afterward, so searching stays just as fast. The result: using only ordinary, freely available data and no artificial examples, our compact model outperforms systems many times larger, including commercial ones, while staying just as good at everyday search. It even reports its own confidence more honestly.