Three shopping-search rerankers ship with a 10,000-pair judged benchmark

single source· 1 articles · confidence: medium · first seen 2026-09-24 20:00 UTC

What this means for you

If you run commerce search on an open reranker, these are worth a look, but the abstract reports relative wins only, with no scores or evaluation dates, so budget for running ShopRank-Bench yourself before switching. The reusable artefact is the labelled set: 10,000 pairs you can score any reranker against. Nothing to act on if you do not rank products.

ZooWork-ShopRanker comprises three open shopping-search rerankers — models that reorder results by predicted shopper preference, not topical match alone — at 0.6B, 4B and 8B parameters, released with ShopRank-Bench, roughly 10,000 preference pairs from private traffic in two text formats. Training labels came from a panel of reasoning models from different families acting as judges, adjusted for position bias — judges favouring whatever they saw first — and tagged by how many judge families agreed; the 8B then taught the two smaller ones. The 8B and 4B are reported to beat the strongest open reranker baseline, with no absolute scores given. Posted to arXiv on 24 September 2026.

Key facts

  • ·ZooWork-ShopRanker is released in three sizes: 0.6B, 4B and 8B parameters source
  • ·Training labels were produced by a panel of reasoning LLMs from different model families acting as a preference oracle, with position-debiased judgments and agreement tiers source
  • ·The aligned 8B model served as a distillation teacher for the 4B and 0.6B models source
  • ·ShopRank-Bench contains roughly 10,000 private-traffic preference pairs in two text formats, tiered by how many judge families committed to each label source
  • ·The paper reports that the 8B and 4B outperform the strongest open reranker baseline, that every model beats its own un-aligned base, and that the 0.6B beats its size peer, but gives no absolute scores or evaluation dates source
  • ·The paper is posted to arXiv as 2609.31002, dated 24 September 2026 source

What the sources say

  • Hugging Face Daily Papers — Abstract-only release note for three reranker sizes and a 10,000-pair shopping-search preference benchmark

Sources

The original reporting. Follow these — they did the work.

← the wire