RPTune: learned catalog curation lifts LLM product search up to 31.4 points; GRPO adds 10.3 more

RPTune: Learned Context Curation for LLM Catalog Search

Chuxuan Hu, Hejie Cui, Norman Huang, Shubham Kumar Bharti, Wang-Chiew Tan, Sercan Ö. Arık

cs.IR, cs.CL, cs.LG

2026-10-01

A learned catalog curator (prune + reorder) plus GRPO post-training lifts in-context LLM search on SMB catalogs: +13.5 EM avg across 9 backbones; gemma-4-E4B-it 10.7% to 31.0%.

What problem this solves

Product search as currently built serves marketplaces with millions of items, dedicated ML teams, and abundant behavioral logs. Small merchant businesses (SMBs) have none of that. Among 56 real Shopify storefronts analyzed in the paper, 92.9% of catalogs fit within 1M tokens and the median is just 51K, so a long-context LLM can hold the entire inventory in its prompt and pick a product directly, skipping retrieval. With gemini-3.7-flash as the shared backbone, full-catalog prompting beats every retrieval baseline by at least 9.0 EM points on average at 8 seconds latency.

Fitting the catalog is not the same as using it. LLMs use long context unevenly, and relevant items buried mid-context get used poorly (the lost-in-the-middle effect). The paper splits the problem in two: how to curate and lay out the catalog, and how to adapt the LLM to curated contexts.

Method

RPTune (Rank, Prune, Finetune) has three trained pieces, all supervised by synthetic data generated from the catalog alone. No merchant labels, no user logs.

• Synthetic supervision: gemini-3.1-pro writes 10 short queries per product variant, then scores every variant against every query on a 0-1 relevance scale, in batches of 10 with a running dictionary for global consistency. Graded scores capture partial matches and near-substitutes.
• Encoder: built on EmbeddingGemma 300M and trained with MNRL (in-batch contrastive loss) to pull queries toward their target products in a shared space.
• Reorganizer: an MLP adds a learned correction on top of encoder similarity, giving each product a context-priority score. The top 25% (default budget) are kept and arranged in ascending score order, so the most relevant item sits closest to the generation side. The MLP trains with the downstream LLM frozen in the loop: the LLM samples G selections from the curated context, rewards come from the precomputed relevance table (-1 for picks outside the catalog), advantages are group-normalized GRPO-style, and gradients flow back through a softmax over reorganizer scores. The curation is optimized for what the LLM actually reads well, not for a retrieval metric.
• Post-training: with the curator frozen, the LLM trains with GRPO on curated contexts under a context-relative reward: +1 for picking the highest-relevance item present in the curated context, 0 for another retained item, -1 for hallucinating something outside it. When pruning removed the globally best variant, picking the best retained one still earns a positive reward, so the learning signal survives.

Results

Evaluation covers 7 real merchants (catalogs of 37.5K to 127K tokens) with 100 complex conversational queries each, 10 runs per method.

The largest single gain is 31.4 points (grok-4.1-fast-non-reasoning on Beauty Bakerie, 14.8 to 46.2). At the same 25% retention budget, Jina Reranker 3.5 hurts accuracy on 5 of 7 merchants and Gemini Embedding 2 on 3, while RPTune improves all 7, so the gains come from learned curation rather than shorter contexts. The sharpest ablation: replacing the RL reorganizer objective with supervised listwise cross-entropy on relevance labels caps EM at 16.4%, no better than the encoder alone and 4.3 points below RPTune. Generalization holds too: after injecting new products until the catalog doubles, post-trained RPTune loses 14.7% relative EM versus 29.5% for the baseline, and all 7×7 cross-merchant transfers are positive (+18.1 EM average out of domain). Per Figure 1, curation buys accuracy comparable to a model-tier upgrade at up to 21× speedup, and post-training brings the 8B gemma into the accuracy range of frontier models.

Why it matters

This is a worked recipe for replacing a retrieval stack with direct long-context reasoning wherever the candidate set fits in context: small storefronts, internal selection tasks, agents choosing among tens or hundreds of tools or products. The division of labor is worth copying: a 300M encoder plus a small MLP does the context engineering, the LLM reads 25% of the catalog, and latency drops instead of rising. Training needs only catalog metadata, zero human labels. The context-relative reward is reusable in any pipeline that prunes context before RL.

The honest framing: contrastive encoders, GRPO, and position sensitivity are all known pieces. The contribution is the composition and the task framing, a solid but incremental engineering argument.

Limitations

The paper has no explicit limitations section; these are reader-observed.

• Evaluation queries are synthetic (an LLM playing customer, consensus filter keeping 31.9%, human audit rejecting 9.6%), not real user logs, and EM credits only the single best variant rather than top-k.
• Training labels are gemini-3.1-pro relevance scores; the ceiling is that model's judgment, and agreement with real purchase intent is not quantified.
• Post-training is validated on one open model (gemma-4-E4B-it); the other 8 backbones are tested frozen with curation only, so generality of post-training is unproven.
• Single-turn queries, catalogs up to roughly 127K tokens; beyond that scale, where RAG returns, is not discussed.
• Latency numbers mix API models across providers with different queuing behavior, so cross-family latency comparisons are indicative at best.

Terms

• SMB: small merchant business, e.g. an independent Shopify storefront.
• EM (exact-match accuracy): fraction of predictions that exactly match the labeled product variant.
• FR (feature reward): weighted attribute-level match between the selected and ground-truth variant, crediting partial correctness.
• MNRL: Multiple Negatives Ranking Loss, an in-batch contrastive objective standard for two-tower retrievers.
• GRPO: Group Relative Policy Optimization, RL that normalizes rewards within a group of samples for the same query to compute advantages.
• Lost in the middle: the finding that information placed mid-context is used less effectively than content near the beginning or end.

Source

• arXiv 2610.00964v1
• PDF
• Hugging Face

What people are saying

• Google's RPTune Paper: Catalog Curation + Post-Training Boosts LLM Search Accuracy by 31.4 Points _reachsumit

Related papers

• Waterloo pre-trains a 3B reranker from scratch, no open-weight backbone, beats GPT-6.1 on TREC DL19
• Frozen indexes support 95%+ recall on 20 benchmarks; learned query encoders often stay below 25%
• Learnable soft top-K cuts Lion-SP queries to 31% with 3.8× faster retrieval
• Sona replaces Yandex Music's 15+ generator cascade, lifting active users 4.53%
• Deletion Margins, Untouched Histories: AIMS Raises Recall@10 by 4.0-10.9%
• Meta: recommendation jobs spent just 50-60% of wall time training; ETT fixes lifted fleet to 90%+

All paper explainers