Learning Multiresolution Relevance for Hierarchical Generative Retrieval
Weihao Shen, Wei Chen, Fuwei Zhang, Guojun Liu, Qingsong Hua, Wei Lin, Fuzhen Zhuang
cs.IR
2026-09-30
RARS projects query relevance into per-depth conditional targets over the SID tree and drops its predictor after training: ESCI NDCG@10 gains 0.54-1.28 pts with unchanged inference.
Generative retrieval replaces the encode-score-lookup pipeline with sequence generation: the model autoregressively emits a document's SID (semantic identifier, a hierarchy of discrete codes where shared prefixes mean semantic proximity), and trie-constrained beam search turns the output into ranked documents. Each token picks one branch of the ID tree; a wrong prefix removes an entire subtree from reach.
Training breaks down on multi-positive queries. A query typically has several relevant documents that share coarse prefixes and separate only at finer depths. Standard full-SID supervision maximizes the likelihood of each (query, document) path independently, so at a shared parent the same query gets pushed toward different children by different positives, one at a time. How relevance splits across branches is never stated during training, a gap the authors call the supervision-resolution mismatch. On the three ESCI locales, 70-80% of eligible queries show relevance branching at the first SID level, about half at the second, and under 20% at the fourth. Branching at coarse depths is the common case, not a corner case.
RARS (Resolution-Aligned Relevance Supervision) writes the allocation into the training objective.
Two choices carry the weight. Parent-mass weighting gives each local loss a say proportional to the relevance reaching that parent, and the appendix shows that under full-child normalization this weighting exactly recovers the KL over whole-path distributions via chain-rule decomposition. Sibling negatives exist for parents with a single relevant child, where a relevance-only softmax has zero gradient. The theory is also clean: relevance-weighted one-hot sampling and explicit distributional targets share the same expected gradient, and the distributional form removes the sampling variance captured by the Gini impurity.
Core comparisons fix the backbone, data, SIDs, and decoding, averaged over five seeds:
| Setting (ESCI) | Metric | Matched base | RARS |
| US, standard AR decoding | NDCG@10 | 11.12 | 12.36 |
| ES, standard AR decoding | NDCG@10 | 12.43 | 13.65 |
| JP, standard AR decoding | NDCG@10 | 11.77 | 12.34 |
| US, AR + per-level compatibility scoring | NDCG@10 | 12.00 | 13.00 |
| US, AR + per-level compatibility scoring | Recall@100 | 38.24 | 39.62 |
The predictor is gone at inference, so the AR-only gains come entirely from the training objective. Against published generative baselines on ESCI-US, the strongest method in the same table, CaLIR, reaches Recall@100 36.15 and NDCG@10 10.75, while RARS hits 39.62 and 13.00; part of that gap comes from the compatibility scoring, which lifts the matched base itself to 38.24. The objective ablation is well controlled: under one shared retrieval rule, RARS beats grouped soft-target, decoder soft-target, and sampled-tree supervision on every metric in every locale (US NDCG@10: 13.00 vs 12.28/12.46/12.55). Gains persist across SID constructions (hierarchical k-means, 5-level RQ-VAE) and relevance definitions, including Exact-only positives, worth +0.49 to +1.00 NDCG@10. In the Japanese case study, a THERMOS bottle query goes from 38.88 to 49.72 NDCG@100: two Exact products share a prefix until the last refinement, and RARS recovers the one that had fallen out of the top 100.
The NQ320K numbers need separate reading: CaLIR and RARS share trained parameters there, so the difference comes purely from the retrieval scoring rule. Unseen-query Recall@10 rises from 2.05% for the strongest baseline to 16.52%, roughly 8x. That jump belongs to scoring, not to the training objective.
For teams running SID-style generative retrieval in e-commerce search or recommendation, this is a training-time-only change: the inference graph and latency are untouched, the auxiliary loss is plug-in, and the code is public. It also gives a crisp answer to how multiple positives should supervise hierarchical decoding. Relevance-weighted whole-path supervision is part of the answer, and the controlled experiments show explicit per-level allocation buys a bit more. The honest framing is an incremental but consistent gain of roughly 0.5 to 1.3 NDCG@10, clearly above the roughly 0.1-point spread across five seeds. The evidence-fidelity analysis stands on its own: with judgments subsampled to 25% retention, over half of multi-positive queries look resolved at the first level while about 70% of leaf-level relevance mass loses support, a warning for any method that derives training targets from sparse judgments.
The authors state that target fidelity depends on judgment coverage, and incomplete labels can make coarse refinements look deterministic while finer depths lose support. The empirical scope is narrow: the training-objective results all come from ESCI product search at T5-base/mT5-base scale, with NQ320K covering only the scoring rule. Per-level compatibility scoring leans on hand-built attribute and lexical features (colors, brands, numbers, IDF terms) that do not transfer for free across domains. CaLIR, the reference retriever and strongest baseline, comes from the same group, so the within-protocol comparisons are the trustworthy ones. The ablations also show that the best temperature and tree-loss weight differ across locales, so hyperparameters would need retuning on a new corpus.