BEQUE: Aligning LLM query rewrites with offline retrieval feedback

简体中文

Fidelity: complete core pipeline. Runs genuine seq2seq SFT, beam self-sampling, offline retrieval feedback, preference ordering and PRO. T5-small and MovieLens substitute for Qwen-7B and Taobao.

Paper information

Field	Details
Paper	Paper or original source
Institution	Alibaba
First public release	2023-11-07 (arXiv v1)
Author code	No official author code identified (checked 2026-08-09).
Local adapter or method key	beque
Local reproduction code	src/auto_research/reproductions/beque/

Paper method and evidence

Long-tail e-commerce queries have semantic gaps; supervised rewriting alone mimics labels without optimizing retrieval. BEQUE uses rejection sampling and multi-instruction SFT, self-samples rewrites, ranks them on relevance, retrieval increment and hit rate in an offline search system, and learns partial preferences with rank-aware PRO. Serving refreshes rewrites offline and looks them up in a KV store rather than running a 7B LLM per request.

flowchart LR
  A[Long-tail query] --> B[Rewriter SFT and beam self-sampling]
  B[Rewriter SFT and beam self-sampling] --> C[Offline retrieval feedback]
  C[Offline retrieval feedback] --> D[Partial order and PRO]
  D[Partial order and PRO] --> E[Offline rewrite KV and online retrieval]

Original paper figure

Figure 2 shows the feedback-alignment architecture. From the original paper; copyright remains with the authors.

Core equation

\[ s_\theta(q,y)=\frac{1}{|y|}\sum_t\log p_\theta(y_t\mid q,y_{<t}), \]

\[ \mathcal L_{PRO}=-\sum_{i=1}^{K-1}\log\frac{\exp(s_i/\tau_i)}{\sum_{j=i}^{K}\exp(s_j/\tau_i)}+\lambda\mathcal L_{SFT}. \]

For ordered candidates \(y_1\succ\cdots\succ y_K\), \(s_\theta(q,y)\) is the length-normalized sequence log-probability.

Reported results

Across all queries, Qwen(SFT) relevance/increment/hit rate is 61.4/109.6/14.58 vs increment-aligned BEQUE's 57.7/198.7/17.27, a tradeoff. The lookup table covers about 27% of PV with almost no added retrieval latency. In Taobao's 14-day A/B, all-query GMV/transactions/UV rise 0.40%/0.34%/0.33%; covered queries rise 2.96%/1.36%/1.22%; nothing query GMV rises 18.66%.

Local reproduction

A public MovieLens title/genre catalog proxies query rewriting. T5-small runs 40 SFT steps, generates four beams per query and uses the inverted catalog to rank relevance, relative incremental retrieval and Hit@1. PRO runs 15 steps on 160 preference groups; each of three seeds uses 600 train/100 test queries. Generated rewrites and unlabeled fallback are used, never the SFT target.

Metric	SFT mean ± SD	SFT + PRO mean ± SD	Relative change
Combined feedback	0.62102 ± 0.08473	0.80750 ± 0.20024	+30.03%
Hit@1	0.32333 ± 0.04041	0.51333 ± 0.11930	+58.76%
Relevance	0.26116 ± 0.04671	0.28176 ± 0.07902	+7.89%
Increment	0.18267 ± 0.03609	0.06208 ± 0.01547	−66.02%

Combined feedback increases by 0.0056/0.1310/0.4228 across the seeds, with a clear hit-rate vs novelty tradeoff. This supports alignment of the optimized compound objective, not universal submetric gains. It is an SFT vs PRO ablation, not a DIN comparison or industrial log reproduction. Stable metrics. Models/checkpoints are not committed; Taobao data is private.

pip install -e '.[plum]'
for seed in 42 43 44; do
  AUTO_RESEARCH_BEQUE_SFT_STEPS=40 AUTO_RESEARCH_BEQUE_PRO_STEPS=15 \
  AUTO_RESEARCH_BEQUE_TRAIN=600 AUTO_RESEARCH_BEQUE_TEST=100 \
  auto-research reproduce --paper beque --dataset-dir data --seed "$seed"
done