CoGR: Symmetric Co-Evolving Generative Retrieval with Reinforced Keyword Matching

The source introduces CoGR, a retrieval framework that trains two LLM generators, one for queries and one for items, to output compact keyword sets that are matched directly through an inverted index. The method first aligns these keyword spaces with supervised fine-tuning and then alternately optimizes both generators with GRPO against the opposite side's frozen index, using retrieval F1 and a counterfactual marginal item-side reward. The authors report best F1 results on an internal APP Marketplace dataset and the public WANDS benchmark, improving over the strongest baseline by 10.9% and 36.1%, respectively.

Hugging Face Daily Papers — It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning

September 3, 20261 source55% confidence

66Strategic

Strategic score

Evidence strength38

Technical novelty78

Commercial potential62

CoGR trains LLM generators on both query and item sides to output compact keyword sets, and matches those keywords via an inverted index. Unlike retrieval setups that use LLMs only for query-side augmentation, CoGR places generation on both sides of the final match.¹

Context

The authors motivate CoGR from recent LLM-based retrieval practice in which generation feeds query expansion, data synthesis, or retrieval-feedback training, but a downstream retriever still performs final matching. CoGR adopts a two-stage training schedule: supervised fine-tuning aligns the two keyword spaces and co-evolving GRPO alternately updates each generator against the other side's frozen index.¹

Assessment

On the supplied evidence, the contribution is a symmetric optimization loop in which both sides share the same F1 objective, while the item side receives F1 only through a counterfactual marginal reward. That symmetry is distinctive, and the abstract's reported gains over sparse, dense, and generative baselines on WANDS would be meaningful if the full evaluation supports them.¹

Caveats

The result is not yet assessable from the abstract alone. Missing details include the computation of the counterfactual item-side reward, the identities and absolute F1 of the ten baselines, the construction of the internal APP Marketplace dataset, and an operational definition of stable co-evolution. Full ablation results or an independent reproduction would change my confidence in the 36.1% WANDS advantage.¹

Source text

Hugging Face Daily Papers — It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning · September 1, 2026

Computer Science > Information Retrieval

Title:It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning

Abstract:Retrieval is the first stage of modern search and advertising systems, selecting a candidate set from a large item universe for downstream ranking and auction. Recent work increasingly leverages LLMs to improve retrieval through query expansion, data synthesis, and retrieval-feedback training. However, the generative component is typically used for query-side augmentation, while final matching is still delegated to a downstream retriever. We introduce CoGR, a retrieval framework that instead trains LLMs to directly construct retrieval representations on both query and item sides. Each generator produces a compact set of keywords, which are matched directly through an inverted index, preserving compatibility with existing keyword-based retrieval infrastructure. CoGR uses a two-stage training pipeline. Supervised fine-tuning first establishes an aligned keyword space, after which co-evolving reinforcement learning alternately optimizes the query- and item-side generators with GRPO against the opposite side's frozen index. Both sides optimize the same query-to-item retrieval $F_1$ objective: the query side receives retrieval $F_1$ directly, while the item side receives a counterfactual marginal reward measuring the change in query-side $F_1$ caused by its generated keywords. Across 10 representative sparse, dense, and generative baselines, CoGR achieves the best performance on both an internal APP Marketplace dataset and the public WANDS benchmark, improving $F_1$ over the strongest baseline by $10.9\%$ and $36.1\%$, respectively. Further analysis shows stable co-evolution and increasingly aligned query--item keyword spaces over training.

Access Paper:

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Counterfactual marginal reward is a technical term whose operation matters for evaluating the method. It means the item-side keywords are rewarded not by their own direct match quality but by the difference they make to query-side retrieval F1, so the reward must be estimated by comparing retrieval outcomes with and without the generated item keywords. The abstract does not state how this counterfactual is estimated or approximated, leaving the main credit-assignment mechanism underspecified.

The 36.1% F1 improvement over the strongest baseline is reported without stating whether it is a relative or absolute increase. Because F1 has a bounded 0-1 or 0-100 scale, this distinction materially changes how impressive the result is; the same issue applies to the 10.9% internal-dataset number. The baseline F1 values are needed before the result can be compared with other retrieval systems.

The source text above remains the property of its publisher and is reproduced for commentary.

Hugging Face Daily Papers — It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning

September 1, 2026endistinct angle

It Takes Two to Match: Co-Evolving Generative Retriever with Reinforcement Learning

↑ Back to top