Is e-commerce sponsored search still a bid-management problem? Two arXiv papers filed July 4, 2026 by the same research group, both concerning Walmart’s platform, argue it is becoming an LLM relevance problem instead. One fine-tunes LLaMA2 7B with LoRA to classify ad relevance at 89.43% accuracy (arXiv:2607.03886); the other uses inventory-aware retrieval-augmented generation to lift ad fill rate by 68% (arXiv:2607.03880). Both figures are self-reported and preliminary, and the direction matters more than the magnitudes.
What do the two July papers actually propose?
The papers, both led by Md Omar Faruk Rokon and submitted the same day, recast sponsored search relevance as an LLM task: one fine-tunes a 7B model to classify whether a query and an ad title are relevant, the other generates ad-matched queries from live inventory using retrieval-augmented generation.
The first, “Enhancement of E-commerce Sponsored Search Relevancy with LLM” (arXiv:2607.03886), builds an Ad Relevance Model on top of LLaMA2 7B adapted with Low-Rank Adaptation. It introduces a three-way classifier, Relevant, Partially Relevant, and Irrelevant, trained on a large e-commerce dataset and evaluated on traffic from a large e-commerce marketplace. The fine-tuned model hits 89.43% accuracy on its test set, which the authors report as beating both their baseline and GPT-4. The paper frames the win as a triad of “search accuracy, cost efficiency, and operational privacy,” the last implying the model can run on infrastructure the platform controls rather than shipping query data to an external API.
The second, “Next-Gen Sponsored Search: Crafting the Perfect Query with Inventory-Aware RAG” (arXiv:2607.03880), targets a different failure mode: the queries that retrieve no sponsored ad at all. InvAwr-RAG integrates semantic retrieval with real-time inventory data, combining dynamically generated queries with historically successful ones, aligning rewritten queries with available inventory and ad campaigns, and diversifying them so the model does not collapse to a single phrasing. Preliminary results report a 68% increase in fill rate on Walmart’s platform. Both papers carry the same workshop venue in their Comments field, eCom@SIGIR 2024 (the ACM SIGIR Workshop on eCommerce, July 18, 2024, Washington, DC), and both landed on arXiv on July 4, 2026, roughly two years after that workshop appearance. Their peer-review status is equivalent: workshop-level, not full-conference review.
The two fit together. InvAwr-RAG attacks the supply side, generating queries that actually have matching inventory and ads. The LoRA classifier attacks the quality side, ranking or filtering what gets shown. Both respond to the same observation in the abstracts: a significant proportion of queries fail to retrieve relevant sponsored ads, and that is unrealized ad revenue.
How is the optimization problem shifting from bids to prompts?
The shift is from optimizing what advertisers bid on to optimizing what the platform retrieves and ranks, with the LLM sitting between the query and the catalog. In the older model, an advertiser picks keywords, sets a bid, and the platform matches query to keyword to ad through an inverted index. In the architecture these papers describe, the platform uses an LLM to rewrite or expand the query against a vectorized catalog and current inventory state, then ranks ads by LLM-judged relevance before the auction settles.
The bid still matters. It is still an auction. But relevance now gates whether an advertiser gets filled at all, and relevance is computed by a model that reads the product’s semantic representation, not its keyword tag. The advertiser’s lever moves from “which keywords, what bid” toward “how well does my product embed into the retrieval context.”
The second-order consequence breaks keyword-bid parity. Two advertisers bidding the same amount no longer get equal treatment. The one whose catalog is vectorized, whose titles and attributes are dense enough to embed well, and whose inventory feed is fresh wins the relevance auction. A team whose product data is thin, with short titles, missing attributes, or unstructured descriptions, gets demoted regardless of how aggressively it bids.
What infrastructure does this require?
Running LLM-grounded sponsored search requires a vectorized product catalog, an embedding pipeline, and a hybrid retrieval layer, not just a keyword index and a bid table. The catalog becomes a retrieval context.
According to vendor guidance from bCloud AI, vector search represents products as 384- to 1,536-dimensional embeddings, which lets “couch” and “sofa” resolve to nearly identical vectors instead of zero keyword overlap (bCloud AI, vendor-reported). The same vendor reports that hybrid search, keyword plus vector, reduces zero-result searches from a baseline of 8-15% to under 2% and lifts search-to-purchase conversion by 30-50% (bCloud AI, vendor-reported). Those are vendor marketing figures from a company selling vector search, so treat the magnitudes as directional rather than audited.
The inventory-awareness requirement is the part keyword systems never had to worry about. InvAwr-RAG’s reliance on real-time inventory data means embeddings and query candidates must refresh as stock changes. A SKU going out of stock has to propagate into the retrieval context, or the LLM generates queries that point at unavailable ads and the fill-rate gain reverses into a relevance loss.
What do the benchmarks actually show?
The headline numbers are self-reported by the same research group on Walmart’s platform: 89.43% accuracy and a 68% fill-rate lift, the latter explicitly labeled preliminary.
The 89.43% accuracy comes from the LoRA paper’s test set, with the authors reporting it beats GPT-4 and a baseline (arXiv:2607.03886). The abstract does not specify the evaluation methodology in detail, so the GPT-4 comparison rests on whatever prompting and test distribution the authors used. A fine-tuned 7B model beating a general-purpose frontier model on a domain-specific classification task is plausible, since specialization usually wins on narrow distributions, but “beats GPT-4” without a matched-prompting protocol is a weak baseline. No third-party replication is reported as of 2026-07-07.
The 68% fill-rate increase is preliminary and comes from InvAwr-RAG’s deployment framing on Walmart’s platform (arXiv:2607.03880). The paper notes “balanced relevance metrics” alongside the fill-rate lift, which is the hedge that matters. Raising fill rate by surfacing more ads is easy if you relax relevance, so the claim only holds if relevance held steady. The paper says it did, in preliminary results, on one platform, from one group.
The vendor figures sit on softer ground. bCloud’s 8-15% to under 2% zero-result reduction and 30-50% conversion lift are vendor-reported, drawn from a company that sells the vector-search product being measured (bCloud AI). Useful as directional evidence that hybrid retrieval helps; not useful as precise magnitudes.
The consistent direction across both papers and the vendor material is that grounding an LLM in a vectorized, inventory-aware catalog improves both fill rate and relevance. The exact magnitudes are not independently audited, and a workshop paper plus a preprint from one group is not the same as a replicated result.
What does deployment look like in practice?
The two papers sketch two deployment patterns: a fine-tuned classifier running relevance inference, and a RAG pipeline generating queries against live inventory.
The LoRA path keeps the base LLaMA2 7B frozen and trains only low-rank adapters, which is the cheap route to domain specialization. The paper’s “cost efficiency” and “operational privacy” framing points at on-prem or locally served inference, which matters because query logs and inventory state are commercially sensitive and platforms are reluctant to send them to a third-party API. A 7B model is smaller and cheaper to serve than a frontier model but still needs GPU capacity, and LoRA adapters are swappable per marketplace, language, or category without retraining the base.
The InvAwr-RAG path is heavier operationally. It combines dynamically generated queries with historically successful ones, which means a query-log store and an inventory feed have to stay in sync with the embedding index. Query diversification is a real engineering concern: without it, the LLM tends to collapse rewritten queries toward a single dominant phrasing, which concentrates impressions and defeats the relevance goal.
The practical upshot for a commerce team is concrete. Without a vectorized product catalog and pre-computed embeddings, you cannot run either pattern. The bid table is no longer the only table that matters. Teams that treat their SKU database as a keyword match table will lose ad-relevance auctions to competitors whose product embeddings enable dynamic query rewriting and real-time relevance ranking, regardless of how much they bid.
Frequently Asked Questions
When does keyword retrieval still outperform vector search?
Exact-match, high-intent queries like SKUs, brand names, and model numbers still retrieve better through keyword indexes than semantic embeddings. Hybrid architectures run both and merge results, which is what the vendor benchmarks actually measure.
What are the actual infrastructure costs for running these models?
A 7B model requires GPU inference capacity even with LoRA fine-tuning, and the vector catalog must refresh as inventory changes. The embedding pipeline represents products as 384 to 1,536-dimensional vectors, which is the storage and compute floor for any deployment.
Why does query diversification matter in InvAwr-RAG?
Without explicit diversification, LLMs collapse rewritten queries toward a single dominant phrasing, which concentrates impressions and defeats the relevance goal. InvAwr-RAG combines dynamically generated queries with historically successful ones to maintain distribution breadth.
What level of peer review do these results have?
Both papers carry workshop-level review from eCom@SIGIR 2024, not full-conference peer review. The accuracy and fill-rate figures are self-reported by the same research group on Walmart’s platform, with no third-party replication as of 2026-07-07.
