CORE: Cascaded ordinal relevance, step-GRPO and PostCoT distillation¶
Fidelity: core mechanism reproduction. Implements flat and cascade classifiers, conditional step reward, group-normalized clipped GRPO and PostCoT logit aggregation with smaller models/data.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Meituan / Beijing Institute of Technology |
| First public release | 2026-07-27 (arXiv v1) |
| Author code | No official author code identified (checked 2026-08-09). |
| Local adapter or method key | core-relevance |
| Local reproduction code | src/auto_research/reproductions/core_relevance/ |
Paper method and evidence¶
A flat three-class relevance model treats unequal ordinal boundaries symmetrically. CORE first asks High vs Non-High, then Mid vs Low only when applicable. The first head focuses on strict entity/intent match, the second on substitutability. The LLM learns structured reasoning through SFT, then assigns independently normalized verifiable ±1 rewards to final answer, High/Non-High, applicable Mid/Low and format. Skipped Mid/Low contributes no reward/gradient. Valid tokens use clipped GRPO ratios plus reference KL. For low latency, PostCoT logits distill into two conditional BERT heads with BCE and temperature-\(T\) distillation.
flowchart LR
A[Query and candidate] --> B[High versus non-high]
B[High versus non-high] --> C[Conditional mid versus low]
C[Conditional mid versus low] --> D[Step GRPO teacher]
D[Step GRPO teacher] --> E[PostCoT student]Original paper figure¶
Figure 3 shows the main flow and stages. From the original paper; copyright remains with the authors.
Core equation¶
\(G\) sampled trajectories normalize each valid step \(k\) separately. Teacher logits are \(\mathbf z=[z_H,z_M,z_L]^\top\).
Reported results¶
On 90,000 private query–item pairs, Direct-BERT accuracy is 0.7441, Cascaded-BERT 0.7558, and distilled cascade 0.7622. Direct-GRPO is 0.7478, Cascaded-GRPO 0.7541, Cascaded-StepGRPO 0.7648 and PostCoT-CORE 0.7706. In production, Cascaded-BERT vs flat BERT improves NDCG@5 from 0.9020 to 0.9038 (+0.20%) and reduces Badcase@5 from 13.8% to 11.6% (−15.9% relative).
Local reproduction¶
Shared MovieLens-1M Low/Mid/High examples use user genre profiles and candidate genre overlap, not Meituan labels. A small MLP replaces Qwen3-14B and production BERT.
| Variant | Accuracy | Macro-F1 | NDCG@5 | Badcase@5 |
|---|---|---|---|---|
| Flat classifier | 0.5833 | 0.5640 | 0.9104 | 0.0625 |
| Cascade | 0.5781 | 0.5657 | 0.8986 | 0.0625 |
| Cascade + step-GRPO | 0.5781 | 0.5649 | 0.9012 | 0.0625 |
| PostCoT-distilled cascade | 0.5781 | 0.5672 | 0.9194 | 0.0313 |
Final NDCG@5 improves 0.98%, Badcase@5 drops 50.00% and accuracy drops 0.52 percentage points vs flat. Late mean active step reward is 0.3811. This validates the component chain, not production search semantics or thresholds. Stable metrics.