跳转至

DeGRe: generative reranking through offline lookahead distillation

简体中文

Fidelity: core mechanism. Runs cumulative-value evaluation, lookahead beam mining, dense prefix labels, beam weighting, and single-pass online generation.

Paper information

Field Details
Paper arXiv 2605.25749
Organization Alibaba Group / Zhejiang University
First public release 2026-05-25 (arXiv v1)
Authors' code The paper did not provide official or author code when checked on 2026-08-09.
Adapter / method degre
Local reproduction code src/auto_research/reproductions/degre/

Original paper

Motivation and method

A pointwise generator learns only the logged path and struggles to optimize long-term list value directly. DeGRe trains an offline cumulative-return evaluator and beam-searches high-value lists. For each prefix, it distills lookahead values of all remaining candidates into dense soft labels. Online only one greedy generator is retained.

flowchart LR
  A[Candidates and context] --> B[Cumulative-value evaluator]
  B --> C[Lookahead beam mining]
  C --> D[Dense prefix labels]
  D --> E[Hard CE and soft KL]
  E --> F[Single-pass online generation]

Key figure from the paper

DeGRe: generative reranking through offline lookahead distillation — Figure 2

Figure 2: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.

Core equations

\[ q_t(v_i)=\frac{\exp \hat V([l_{<t};v_i])}{\sum_{v_j\notin l_{<t}}\exp \hat V([l_{<t};v_j])},\qquad w_l=\operatorname{softmax}(\hat V(l)/\tau_w). \]
\[ \mathcal L=\sum_lw_l\sum_t\left[\operatorname{CE}(y_t,p_t)+\alpha\operatorname{KL}(q_t\|p_t)\right]. \]

Paper results

Taobao Flash Shopping A/B, eight days at 2% traffic: CTR +2.85%, orders +2.14%, GMV +3.75%. The paper also reports Alipay GMV +4.14%.

Local reproduction

Local comparison: The baseline is a pointwise generator with the same candidates and features. Offline lookahead distillation changes Hit@10 by +0.00% and NDCG@10 by +3.31% relative.

MovieLens-100K uses 220 users and 360 items. Transaction value is proxied by public positive feedback, list diversity, and novelty. See the stable metrics.

auto-research reproduce --paper degre --seed 42

Reproduction boundaries

Private transaction labels and industrial Transformer scale are not reproduced. The local result checks only the offline evaluator, beam mining, and dense distillation pipeline.