- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies retrieval failures in generative recommendation after catalog expansion: a newly admitted item may have a valid semantic identifier yet remain inaccessible because of intermediate beam pruning. With the generator fixed and old identifiers retained, the authors establish output-invariance certificates and characterize exact assignment-repair feasibility under a common effective prefix. Beyond the Beam (BB) then constructs minimum-replacement repairs through minimum-cost flow, converts local repairs into shared-map proposals, and adapts the generator to the selected map. At inference, BB combines generative likelihood with a collaborative correction and uses retained prefix upper bounds for candidate completion. When its stopping condition holds, the returned list is certified as the global Top-K under the combined score.
The evaluation includes exhaustive finite-catalog experiments and recommendation experiments on Amazon Beauty, Tools, and Toys. The complete T5 system achieves substantial improvements, while the benefits of scoring and candidate completion also extend to LC-Rec.
Critical strengths
- A clear and useful decomposition of the retrieval problem. The paper distinguishes identifier validity, assignment-repair feasibility, and candidate access. Its joint treatment of prefix support, ranking competition, and catalog size explains why assigning a high-scoring identifier alone may fail to repair retrieval.
Critical weaknesses
-
Scalability and serving costs remain important limitations. Appendix B.7 acknowledges () storage for the dense collaborative predictor, () time for direct Cholesky fitting, and catalog-wide prediction. The evaluated catalogs contain only approximately 10,000 items. Extrapolating the stated FP32 dense representation to one million items would require approximately 4 TB for the coefficient matrix alone, excluding fitting workspace.
-
Direct comparisons with alternative candidate-expansion methods are missing. SpecGR is discussed but not evaluated. Its use of candidate drafting and generative verification makes it a particularly relevant comparison for candidate completion. SpecGR, AAAI 2026
Questions for authors
-
BB uses update-training interactions and performs generator adaptation after catalog expansion. Please clarify whether the intended setting is incremental updating after new items have accumulated interactions, or whether the method also supports interaction-free new items. How does performance vary with the number of available interactions per new item?
-
Scalability - For social media platform like instagram/facebook/tiktok, the # of category usually exceeds , is it possible to adopt the method here for the retrieval stage?
5 comments
Response to fcyB (2/2): Direct SpecGR comparison and completion benefits
W2
The direct comparison favors BB in both Recall@10 and NDCG@10.
We evaluate SpecGR_Aux with a UniSRec drafter and the same final adapted T5 checkpoint and identifier map as BB. The auxiliary variant plugs into this shared generator, giving both methods the same adaptation and repair benefits. The comparison therefore tests their candidate-expansion and scoring designs.
Beauty, three-seed means over 7,335 queries per run:
| Inference method | R@10 (%) | N@10 (%) |
|---|---|---|
| SpecGR_Aux, UniSRec drafter | 2.8903 | 1.3739 |
| BB, bound-guided completion | 3.2175 | 1.5353 |
BB improves Recall@10 by 11.32% and NDCG@10 by 11.75% relative to SpecGR. The shared-generator comparison demonstrates the recommendation advantage of BB's candidate-completion and ranking procedure over draft-and-verify inference.
Matched inputs and candidate allowance.
Both methods use the same training/update information, legal catalog, history encoding and beam width 40. The initial beam's fully scored set E0 is cached for both. Each distinct verification outside E0 consumes one unit of the 80-item additional-evaluation allowance; repeated drafts reuse the cache. Each method produces up to 20 recommendations, evaluated at 10.
SpecGR retains its native length-normalized verification, treatment of the identification token according to actual training support, prefix-guided redrafting and beam fallback. UniSRec uses the available item text and interactions. Draft size is selected from {10, 20, 40, 80} and acceptance threshold from {-1.8, -1.7, ..., -1.1}, using the same 1,040-query validation set, ordered by NDCG@10, Recall@10 and fixed configuration order. BB retains its submitted scoring configuration in Eq. 8, with the correction in Eq. 7. At budget or round exhaustion, SpecGR applies its beam fallback. Each method retains its own ranking rule.
BB adds a ranking-based completion criterion.
SpecGR exits once enough drafted candidates pass its verification threshold. BB prioritizes candidates by upper bounds on its final score and stops when every remaining bound is below the current Kth score. This couples the evaluation order to a certificate for the returned ranking.
Submission Table 11 isolates the evaluation order at a fixed fused score and the same 80-item cap:
| Priority on Beauty | Additional evaluations/query | Top-20 certified (%) | Batched ms/query |
|---|---|---|---|
| Collaborative priority | 60.69 | 45.40 | 10.50 |
| BB bound priority | 51.08 | 77.64 | 10.16 |
BB reduces additional evaluations by 15.83% and increases certification by 32.24 percentage points, while batched inference time decreases from 10.50 to 10.16 ms/query. Timing includes fresh beam, item scoring, encoding and completion under the protocol in Appendix B.7. These results connect more selective candidate evaluation to measured computational savings. The SpecGR comparison establishes the direct ranking advantage; the fixed-score priority control shows how BB uses computation to obtain more certified outputs. Together they support the quality and search contributions of candidate completion.
Appendix E.8, Table 26 reports the direct comparison, with its evaluation protocol in the same section.
Thanks for the response
It's still unclear to me whether the proposed method can afford O(100m) candidates
Response to fcyB (1/2): Interaction support, catalog scale and algorithmic gains
Thank you for your careful review, your recognition of our analysis of retrieval failures, and your suggestion to compare with SpecGR.
Q and W denote your questions and weaknesses, numbered in their original order.
BB improves recommendation quality through constructive identifier repair and selective candidate completion.
Q1
BB targets incremental updating as new-item interactions arrive.
Identifier admission gives a new item catalog eligibility, including at zero interactions. The learned recommendation stage uses update interactions for generator adaptation and collaborative scoring while retaining old identifiers. The reported gains therefore evaluate interaction-informed incremental recommendation.
The breakdown below measures target supervision: an occurrence is a next-item target appearance in update training; history appearances are a separate source of exposure. Each cell reports NDCG@10 (%) / query count / distinct target-item count; metrics are three-seed means.
Beauty and Toys show positive new-target ranking with 1-2 supervised target occurrences and higher performance in the groups with more target supervision. BB preserves existing identifiers while improving recommendation as the expanded catalog receives interactions.
The supervision breakdown is included in Appendix E.3, Table 16.
Q2 / W1
BB's algorithmic contributions and catalog scale.
As the review highlights, identifier validity, repair feasibility and candidate access are distinct. BB makes this distinction constructive: joint prefix-support, ranking and catalog-size constraints characterize repair feasibility and yield minimum-replacement repairs under the common-prefix model. At inference, retained-prefix bounds guide candidate evaluation and Top-K certification under the final ranking score.
The reviewer's 4 TB estimate corresponds to the FP32 coefficient matrix of the dense collaborative predictor. Its quadratic storage, cubic direct fitting and catalog-wide prediction costs are documented in Appendix B.7. These costs make the current dense implementation impractical for million-item first-stage retrieval.
The matched comparison establishes what BB contributes with this scoring component held fixed.
In submission Table 11, both priority rules use the same collaborative predictor, final ranking score and 80-item additional-evaluation cap. On Beauty, BB's bound priority reduces additional full-item evaluations from 60.69 to 51.08 per query, a 15.83% reduction relative to collaborative priority, while increasing numerical Top-20 certification from 45.40% to 77.64%. With the same collaborative evidence, BB thus obtains more certified outputs from fewer generative evaluations. This directly demonstrates the computational benefit of aligning candidate evaluation with the final ranking score.
The recommendation experiments connect these mechanisms to ranking quality. The complete T5 system delivers gains on Beauty, Tools and Toys, and scoring and completion also benefit LC-Rec. In the shared-generator comparison on Beauty, BB improves NDCG@10 by 11.75% over SpecGR_Aux (Part 2; Appendix E.8, Table 26). Together, the constructive theory and matched evaluations establish BB's contribution to recovering missed recommendations and using generative computation more effectively.
Part 2 provides the direct SpecGR comparison and the separate evidence for bound-guided candidate evaluation.