简体中文 · English

RPTune: Learned Context Curation for LLM Catalog Search¶

L1 core mechanism diagnostic. Only rank/prune/position orchestration runs. Gemini/Gemma, merchant catalogs and the paper's training are not used.

Paper information¶

Field	Details
Paper	arXiv v1
Institution	Google
First public release	2026-10-01 (arXiv v1)
Original source code	No public author implementation found as of 2026-10-03.
Adapter	`rptune`
Local reproduction code	`src/auto_research/reproductions/rptune/`

Original paper summary¶

Background and main changes¶

A query–product encoder, learned adjustment and downstream LLM feedback determine what products to retain and where to place them. Higher-priority items go near the prompt end. Post-train selection on the curated catalog.

flowchart LR
  I[Input and current state] --> M[Rank, prune and position]
  M --> A[Auditable intermediates]
  A --> O[Output or state update]

Key figure from the paper¶

Figure 2: Core architecture, modules and connections. From the paper; copyright belongs to the authors. Click for the source.

Core equation¶

\(s_i=10e_q^Te_i+MLP([e_q;e_i])\). Keep the top-k items and sort them by ascending score.

Offline and online results¶

Across 7 real merchants, the paper reports a mean curation gain of 14.0 exact-match points and another 10.3 points from post-training.

Local reproduction¶

Comparison protocol: Disabled or default mechanism versus its defining operator. L1 does not report formal relative gains; percentages are not applicable.

Three-seed diagnostic: metrics/mechanism-seeds42-44.json.
diagnostic_only=true; excluded from formal capability rankings.

Scope¶

Only rank/prune/position orchestration executes. There are no Gemini/Gemma calls, merchant catalogs or paper training.