Srijith Ravikumar · Information Processing & Management 2026 · 2026
DOI: 10.1016/j.ipm.2026.105174
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
LLM query planners turn a query into a structured plan of filters enforced before retrieval. For a type-less entity query naming a motif (”dinosaur”, ”pineapple”), they emit a product-category slot the query never expressed, which as a hard filter excludes most of the relevant catalog. We evaluate nine planners across six providers on two public collections (WANDS: 42,994 products, 233,448 human-graded pairs; Amazon ESCI: 1.2M products, 404 of its human-graded queries) and a 1217-query type-less probe. Under an aggressive prompt all nine route at about 100%. On the entity probe, stock prompts elicit far less: 0.00–0.04 from the best-aligned planners, 0.20–0.73 from weaker open ones. On WANDS the guessed filter removes 0.54 of motif-relevant recall@100 under a restraint-instructed prompt and 0.74 under the aggressive one. On ESCI human grades, planners guess a spurious brand on 0.10 of queries and a color on 0.24, spuriousness judged by an automated entailment audit, not human adjudication. Over the full catalog this mis-routing removes 0.03 to 0.09 of recall@100. A standard content-faithfulness grader, including RAGAS, is blind to about half of these mis-routings, because the results were selected by the guessed category. Conditioning on the query, not on the results, separates mis-routing from explicit legitimate typing, though off-the-shelf checkpoints stay unreliable on implicit types. A one-line omit instruction cuts probe routing from 1.00 to 0.16–0.27 and transfers to six APIs, though models can over-omit and the weakest break format. We release QP-Faith, a label-free routing audit (no relevance judgments), with two retraining-free mitigations.
No comments yet — start the discussion below.