CRSD: From Reasoning LLMs to BERT, a Two-Stage Distillation Framework for Search Relevance

简体中文

Public-data core mechanism, not private serving reproduction.

Paper information

Field	Details
Paper	TheWebConf 2026 Industry
Institution	Meituan.
First public release	2025-10-13 (arXiv v1)
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	crsd
Local reproduction code	src/auto_research/reproductions/crsd/

Paper method and evidence

Build a domain reasoning LLM with continued pretraining, supervised fine-tuning and multidimensional preference optimization. Contrastively self-distill between plain and rationale-augmented inputs to the same lightweight student, eliminating the reasoning chain at serving.

flowchart LR
  T[Domain reasoning teacher] --> R[Rationale-augmented student input]
  P[Plain student input] --> D[Contrastive self-distillation]
  R --> D
  D --> S[Serve without rationale]

Original paper figure

The original figure shows the two distillation stages. From the original paper; copyright remains with the authors.

Core equation

\[ \mathcal L=\mathcal L_{label}+\lambda D_{KL}(p_s(\cdot|x,r)\Vert p_s(\cdot|x))+\gamma\mathcal L_{con}. \]

Reported results

Meituan search ads at 30% traffic: AdCTR +0.91%, AdCVR +1.06%, GTV +0.40%, and bad-case rate down 30.5 percentage points.

Local reproduction

Shared transition + content baseline versus CRSD changes only the core mechanism. MovieLens-100K full-catalog NDCG@10 0.0354 → 0.0346 (-2.14%). Fixed metrics.

auto-research reproduce --paper crsd --dataset-dir data --seed 42

The local distillation experiment does not use private logs or production serving; local and online metrics are separate.