acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Beam: Constructive Repair and Candidate Completion for Generative Recommendation

Abstract

Generative recommenders retrieve items by generating identifiers, but a valid identifier can remain outside the beam after catalog expansion. This raises two connected questions: which failures can identifier assignment repair, and how should retrieval proceed beyond the initial beam? We characterize assignment repair with a fixed generator and retained old identifiers. Output-invariance certificates identify failures shared by all admissible assignments. Under a common effective prefix, coupled support and ranking constraints give the exact feasible interval of new-item counts for target recovery. Building on this characterization, Beyond the Beam (BB) obtains minimum-replacement repairs through an integral flow formulation, selects a shared map and adapts the generator. At inference, generative likelihood and collaborative evidence define one score for ranking, candidate priority and stopping. Retained prefix bounds guide candidate completion and certify its global Top- when the stopping condition is met. Exhaustive finite-catalog evaluation confirms construction in every feasible case. Across three Amazon Reviews categories and three random seeds, the full T5 procedure improves mean Recall@10 by 15.5–46.3% and NDCG@10 by 15.2–44.4% over the best-performing evaluated generative baseline for each dataset and metric. Matched controls show that shared construction and adaptation improve new-target ranking and certification efficiency on Beauty and Toys. Combined scoring and candidate completion improve NDCG@10 across all three datasets with both T5 and decoder-only LC-Rec.

Primary area:
generative models
Keywords:
recommender systems, generative recommendation, semantic identifiers

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_fcyBOpenReview ↗
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 (O(∣I∣2)O(|I|^2)) storage for the dense collaborative predictor, (O(∣I∣3)O(|I|^3)) 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 O(100M)O(100M), is it possible to adopt the method here for the retrieval stage?

5 comments
AuthorsOpenReview ↗

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.

Dataset0 occurrences1-2 occurrences3-5 occurrences6+ occurrences
Beauty0.00 / 76 / 310.64 / 344 / 1490.95 / 352 / 765.57 / 714 / 93
Tools0.00 / 179 / 610.00 / 175 / 388.77 / 79 / 970.45 / 54 / 3
Toys0.00 / 71 / 410.62 / 447 / 1981.66 / 265 / 1235.04 / 316 / 112

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.

AuthorsOpenReview ↗

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 methodR@10 (%)N@10 (%)
SpecGR_Aux, UniSRec drafter2.89031.3739
BB, bound-guided completion3.21751.5353

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 BeautyAdditional evaluations/queryTop-20 certified (%)Batched ms/query
Collaborative priority60.6945.4010.50
BB bound priority51.0877.6410.16

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.

Reviewer_fcyBOpenReview ↗

Thanks for the response

It's still unclear to me whether the proposed method can afford O(100m) candidates

Reviewer_XP1bOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper addresses a real problem in generative recommendation: a new item may have a valid semantic ID but still be removed by beam search before its complete ID is generated. The authors propose two pieces. First, they analyze when changing the IDs of new items can make a missed item reachable while keeping old IDs fixed, and use a minimum-cost flow problem to construct such changes. Second, they combine generator scores with a collaborative score and evaluate additional items outside the initial beam, using prefix scores to decide when the current Top-K is certified under their combined score.

The paper reports experiments on Amazon Beauty, Tools, and Toys. The complete system often beats the evaluated generative baselines, but the separate contribution of the construction step is small and inconsistent across seeds.

Together, these issues make it difficult to justify acceptance at ICLR.

Critical strengths
  1. The paper identifies a real failure mode. A valid item ID can still be unreachable because an earlier prefix is pruned by beam search. The paper explains why simply reranking the returned beam cannot recover such an item.

  2. Candidate completion is supported by useful controls. The experiments separate reranking the returned beam, reusing already scored leaves, and scoring additional items. The bound-based priority also reduces extra evaluations and increases the fraction of cases certified under the fixed combined score.

Critical weaknesses
  1. The paper is difficult to follow and does not explain the main contribution in a sufficiently direct way. The exact theory is restricted to a common-prefix setting, while the practical method searches over proposals and then adapts the model. The paper does not clearly state which claims are exact, which are heuristic, and which are only empirical observations. This makes it hard to judge what the central contribution actually is.

  2. The novelty and reproducibility case is not strong enough. The assignment result is a narrow characterization followed by a standard minimum-cost flow reduction. The inference part combines a collaborative correction, prefix upper bounds, and additional scoring, but the paper does not clearly separate its novelty from existing assignment, candidate-expansion, and best-first search ideas.

  3. The submission points to an anonymous code URL, but no usable source code was available to me, so the main experiments are not independently reproducible from the submission.

Questions for authors
  1. What is the central new idea? What is new relative to prior identifier-assignment methods and prior candidate-expansion or best-first search methods. Which part of the paper would remain novel if the collaborative score and additional candidate evaluation were removed?

  2. Can the construction step be isolated? Please compare the proposed flow-based construction with a simple high-score assignment and an occupancy-preserving identity-swap method under the same proposal budget, identity-edit limit, and adaptation procedure.

  3. Can the experiments be reproduced? Please provide an accessible repository or an archive with the exact commit.

2 comments
AuthorsOpenReview ↗

Response to XP1b (1/2): Central idea and theory-to-method connection

Thank you for your thoughtful review and your suggestions on clarifying the methodological contribution and evaluating the construction step.

Q and W denote your questions and weaknesses, numbered in their original order.

Q1 / W1–W2

BB turns identifier repairability into a constructive optimization problem, then makes candidate completion certifiable under the final ranking score.

The theory identifies recoverable failures; minimum-replacement repair supplies shared-map proposals; completion directs query-specific evaluations beyond the beam.

(a) A distinct assignment contribution remains when collaborative scoring and additional evaluation are removed.

JTM optimizes item-to-leaf matching; Reformer and DACT update tokenization. BB characterizes the legal assignments that recover a missed target while retaining old identifiers. An occupied leaf both supports a prefix and competes for a ranking position. Joint support, ranking and catalog-count constraints yield the exact feasible admission-count interval under a common effective prefix. Output-invariance certificates separately identify failures shared by all admissible assignments, with the generator and encoded history fixed.

This feasible family is the input to minimum-replacement construction. The flow network enforces its support, count and ranking constraints jointly and returns an integral repair with minimum occupied-leaf replacement for the specified state and target. Thus the characterization determines both which repairs exist and how to obtain a least-disruptive occupancy change. In the original complete-state feasibility/construction test, construction succeeds in all 934 feasible conditions, versus 733 for high-score allocation (submission Table 8).

(b) The exact result supplies the practical intervention.

Each local repair becomes a complete-map proposal by pinning the target identity, retaining compatible identities and assigning displaced identities deterministically. Compatible proposals are composed and evaluated on re-encoded training histories. The selected shared map is then fixed for generator adaptation.

The stages connect guarantees to recommendation gains: the theorem gives exact feasibility and minimum replacement under the common-prefix conditions; shared-map selection uses heuristic search to compose and evaluate these constructive proposals across training contexts; Part 2 measures their gains after adaptation. The revised Introduction and Related Work make this theory-to-method connection explicit.

(c) The search contribution is catalog-wide completion under the combined score.

Huang et al. establish stopping optimality modulo the beam; Meister et al. reorder beam expansion using monotonic scores. BB carries an admissible bound for every catalog identifier, including paths pruned from the initial beam, by adding its known item correction to the retained prefix score. These bounds govern additional full-item evaluations and certify the global Top-K under the same fused score used for ranking.

SpecGR drafts candidates, verifies them against a threshold and exits when enough are accepted. BB stops when every remaining item's bound is below the current Kth score. This connects candidate access, evaluation effort and a checkable ranking result. With the same adapted T5, map and 80-item additional-evaluation allowance, BB's Recall@10 / NDCG@10 are 3.2175% / 1.5353%, versus 2.8903% / 1.3739% for SpecGR_Aux.

The original fixed-score control isolates bound priority: Beauty's extra evaluations fall from 60.69 to 51.08, while numerical Top-20 certification rises from 45.40% to 77.64% (submission Table 11). Together, the feasibility result, matched constructor gains and completion controls identify the contributions of the theory, repair and search components.

The added method comparisons and their protocols are included in Appendix E.8.

AuthorsOpenReview ↗

Response to XP1b (2/2): Matched construction controls and reproducibility

Q2

Flow improves ranking over both requested simpler controls with a small identifier update.

The Beauty seed-42 comparison follows the submitted map-selection, adaptation and inference protocol. Metrics are percentages over all 7,335 test queries; Primary contains 734 new-target queries with old-only histories.

AssignmentR@10N@10Primary N@10Changed IDsReplaced leaves
Without construction3.01301.42511.729300
High-score greedy3.02661.43441.75291615
Identity swaps2.97211.41211.609860
BB / Flow3.13571.48312.345020

BB improves overall Recall@10 and NDCG@10 over without construction by 4.07% and 4.08%, and Primary NDCG@10 by 35.60%. Overall NDCG@10 exceeds greedy and swaps by 3.40% and 5.03%. BB changes two identifiers while preserving the occupied-leaf set, compared with 16 identifier changes and 15 leaf replacements for greedy, and six identifier changes for swaps. Minimum occupied-leaf replacement is the construction objective, so the ranking gains and edit counts jointly describe repair effectiveness and update cost. Primary NDCG@10 exceeds swaps by 45.67%. Together with the complete-state construction tests in submission Table 8, these results connect jointly feasible repair to downstream recommendation gains.

Constructor comparison.

High-score greedy pins the target to a candidate leaf and replaces lower-scoring new-item leaves with higher-scoring available leaves. Identity swaps exchange new-item identities within the occupied-leaf set. Flow jointly enforces support, count and ranking constraints while minimizing occupied-leaf replacement.

Shared protocol.

All four arms share the five-epoch warm-up, static starting map, training randomization and Constant adaptation recipe. The constructors use the same unfiltered candidate space, proposal contexts and access limits, with at most 12 distinct proposals and a 16-ID edit cap. Proposal acceptance, shared-map selection and validation checkpoint selection follow the same submitted procedure. Final inference uses seed 42's BB scoring parameters, beam 40 and the 80-item additional-evaluation cap. Appendix E.3, Table 18 reports the results and the construction and adaptation protocol.

Q3 / W3

Code availability and reproducibility.

The source code is available in the bb_code/ subdirectory of the original anonymous repository. We have added a README at the repository root to clearly indicate where the source files are located and explain the code organization and experimental workflow. We invite you to inspect the code and README directly at the same link.

Reviewer_PRZjOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a repair and completion mechanism for generative recommenders. The problem is presented as new items getting added to a catalog may never get recommended due to getting dropped early by beam search. The paper tries to answer whether changing the codes of new items can help while keeping old codes the same and how good items can be found that the beam search missed.

       The authors show when no code changes can help, and how many new items can be added while still letting a target item reach the top results. The method that was introduced, Beyond the Beam (BB), finds the smallest change that can fix individual queries and then combines them into a shared set of codes, and retrains the model on it. BB can stop early guaranteeing it found the top items when its stopping condition is met. Experiments show 15-46% improvement over the best generative baseline.
Critical strengths
  1. The theory presented in the paper is honest about its limitations on when changing codes can and can not recover an item.
  2. The beyond the beam search method presented in the paper is simple. Adding the additional score to partial scores gives an upper bound, so the method can recover items the beam dropped and find the top results. It also seems cheap based on the numbers presented in table 11. This seems to be useful on its own, separate from the rest of the method.
Critical weaknesses
  1. The performance gains presented in the paper seem to come mainly from combining the generator with a linear recommender which is not a new technique and the theory plays only a small role in the final method and experiments. Removing the code changes does not change the result meaningfully (Table 2: Beauty R@10 is 3.231 vs 3.217). Table 3 shows that final maps change only 0-5 items. This is important because the main focus of the paper is presented as to be on code repair.
  2. The baselines presented in the paper don’t get the same advantages as BB. The reported gains don’t show BB is better in a level playing field especially given that SASRec, a non-generative model, beats every generative baseline on the Beauty dataset as presented in Table 18.
Questions for authors
  1. If you add BB’s extras (linear model score and search for items outside the beam) to baselines like Reformer and DACT, does BB still win overall?
  2. Based on Table 2, removing the code changes barely affects the results. Can you show a case where changing the codes clearly helps overall?
  3. Do you have an understanding of why the linear model on its own does better than full BB on new items for Beauty and Toys as presented in Table 13? Would giving all new items a fixed score boost work just as well as what is presented in equation 7?
2 comments
AuthorsOpenReview ↗

Response to PRZj (1/2): Matched baselines and construction

Thank you for your careful review and your questions about fair comparisons and the contributions of BB's individual components.

Q and W denote your questions and weaknesses, numbered in their original order.

BB combines minimum-replacement identifier repair with score-guided candidate completion.

On Beauty, BB improves NDCG@10 by 21.39% over enhanced Reformer. The seed-42 construction comparison improves overall NDCG@10 by 4.08% with two identifier changes. Part 2 examines collaborative-only scoring and constant-boost controls.

Q1 / W2

BB remains stronger when DACT and Reformer receive the same enhancements.

DACT and Reformer receive the same collaborative predictor and scoring family in Eqs. 7-8, with parameters selected from 65 validation configurations. BB retains its submitted checkpoint, map and scoring configuration. All methods use beam width 40 and an 80-item additional-scoring cap. Each baseline's reranking and completion share its selected score; completion reuses the initial fully scored set E0.

Beauty, three-seed means over 7,335 queries per run; metrics are percentages (Appendix E.8, Table 25):

MethodR@10N@10Fully scored items/query (initial + added = total)
DACT, original beam2.0951.063-
DACT + collaborative reranking2.48581.2070-
DACT + reranking and completion2.78571.342360.40 + 49.67 = 110.07
Reformer, original beam2.0631.015-
Reformer + collaborative reranking2.31311.1010-
Reformer + reranking and completion2.66761.264860.06 + 50.14 = 110.20
BB3.21751.5353126.37 + 51.08 = 177.45

Initial counts record legal terminal leaves scored during beam expansion, including leaves outside the returned beam. A terminal expansion supplies scores for multiple leaves; the model and identifier map determine E0's size. Each method reuses E0 before additional scoring. Original beam rows reproduce submission Table 1.

BB exceeds enhanced DACT and Reformer in NDCG@10 by 14.38% and 21.39%, respectively. At each baseline's fixed selected score, extending candidate access beyond the returned beam improves NDCG@10 by 11.21% for DACT and 14.88% for Reformer. BB also exceeds SASRec's original Beauty Recall@10 of 2.677% (revised Table 22).

BB already leads with 40 reranked candidates.

In the submission's ablation, reranking BB's 40 returned candidates gives NDCG@10 of 1.4068%, exceeding enhanced DACT (1.3423%) and Reformer (1.2648%) by 4.81% and 11.22%. With the same BB configuration, reranking all 126.37 initial items on average raises NDCG@10 to 1.4159% (+0.0091 percentage points); completion raises it further to 1.5353% (+0.1194 points). These stages show the gains from reusing initial candidates and scoring additional items (revised Tables 2 and 19).

Q2 / W1

Construction improves overall ranking with a small update footprint.

The feasibility characterization supplies the constraints of the flow constructor, whose local repairs become shared-map proposals before generator adaptation. The Beauty seed-42 comparison covers all 7,335 test queries (Appendix E.3, Table 18). The arms share the original proposal procedure, a 16-ID edit cap, warm-up, adaptation and checkpoint selection. Inference uses this seed's submitted BB scoring parameters. Primary contains 734 new-target queries with old-only histories.

AssignmentR@10 (%)N@10 (%)Primary N@10 (%)Changed IDsReplaced leaves
Without construction3.01301.42511.729300
High-score greedy3.02661.43441.75291615
Identity swaps2.97211.41211.609860
BB / Flow3.13571.48312.345020

BB improves overall Recall@10 and NDCG@10 over without construction by 4.07% and 4.08%, and Primary NDCG@10 by 35.60%. Overall NDCG@10 exceeds greedy and swaps by 3.40% and 5.03%. BB changes two identifiers while preserving the occupied-leaf set, compared with 16 identifier changes and 15 leaf replacements for greedy, and six identifier changes for swaps. Minimum occupied-leaf replacement is the construction objective, so the ranking gains and edit counts jointly describe repair effectiveness and update cost.

The original Primary NDCG@10 gains are 13.95% / 19.42% on Beauty/Toys (Table 13; Appendix E.3).

AuthorsOpenReview ↗

Response to PRZj (2/2): Collaborative evidence and constant-boost controls

Q3 / W1

Combining generative and collaborative evidence improves the complete-catalog objective.

The submission's Item correction only control ranks by the calibrated correction did_i in Eq. 7; BB adds generative likelihood through Eq. 8. In the original Beauty results, adding generative evidence raises old-target NDCG@10 from 0.458% to 1.261% and overall NDCG@10 from 1.156% to 1.535%. Collaborative-only scoring ranks new targets more strongly: its new-target NDCG@10 is 4.063% / 3.321% on Beauty/Toys, versus BB's 3.052% / 2.103% (submission Table 13). Fusion changes both old/new competition and ordering within the new-item group. The following diagnosis distinguishes these two mechanisms and identifies which misses arise before ranking.

(a) The diagnosis separates access from ranking competition.

For query-seed pairs where collaborative-only scoring hits the target at 10 and BB misses it:

DatasetTotal pairsUnscored targetScored, top 10 among new itemsScored, outside top 10 among new items
Beauty132125862
Toys108343242

The target was already scored in 120/132 Beauty and 74/108 Toys pairs. Removing old items from the evaluated ranking restores 58 and 32 of these hits; the other scored misses reflect ordering within the new-item group. This locates most of these missed hits in the ranking stage. Fusion changes both old/new competition and the order among new items, while improving the complete-population objective.

(b) Item-specific correction improves on a uniform new-item boost.

We fix the submitted BB checkpoint, map and predictor. For a fixed history hh, let si=sθ(i∣h)s_i=s_\theta(i\mid h) be full generative log-likelihood and qiq_i the collaborative score lower-clipped at 10−810^{-8}. We compare:

F0(i)=si+λlog⁡qi,1i∈Iold+c,1i∈Inew,F_0(i)=s_i+\lambda\log q_i\\,\mathbf{1}\\{i\in\mathcal I_{\mathrm{old}}\\}+c\\,\mathbf{1}\\{i\in\mathcal I_{\mathrm{new}}\\}, F1(i)=si+λlog⁡qi+b,1i∈Inew.F_1(i)=s_i+\lambda\log q_i+b\\,\mathbf{1}\\{i\in\mathcal I_{\mathrm{new}}\\}.

Here 1A\mathbf{1}\\{A\\} equals 1 when condition AA holds and 0 otherwise; Iold\mathcal I_{\mathrm{old}} and Inew\mathcal I_{\mathrm{new}} are the old- and new-item sets. The constants cc and bb are uniform new-item offsets.

F0 tunes the new-item constant c directly and includes lambda = 0, covering a pure generator plus constant. F1 retains item-specific collaborative evidence with a shared coefficient and a new-item bias. Parameters are selected on 1,040 Beauty validation queries using a coarse grid followed by one local refinement, ordered by NDCG@10, Recall@10 and fixed configuration order. Each rule uses its own completion priority and stopping condition, with beam 40 and the same 80-item cap. BB retains its submitted configuration.

Beauty, three-seed mean NDCG@10 (%):

DatasetScoreOverallOldNew
BeautyUniform new correction, F01.21791.23901.5643
BeautyCollaborative score + bias, F11.38871.21442.4955
BeautyFull BB1.53531.26113.0518

BB exceeds uniform new-item correction F0 by 0.3174 percentage points in overall NDCG@10 and 1.4875 points on new targets. It exceeds F1 by 0.1466 overall points and 0.5563 new-target points. A uniform bonus preserves the generator's ordering among new items; item-specific collaborative evidence can change that ordering. These results show the benefit of item-specific collaborative evidence and its calibration across old and new items.

Appendix E.2 includes the complete three-seed scoring results (Table 14) and ranking diagnosis (Table 15).

Together with the matched baseline and construction comparisons in Part 1, these results establish the contributions of repair, candidate access and item-specific scoring.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of additional evidence and clarifications

We thank the reviewers for their careful reading and constructive feedback, and the area chair for overseeing the review process.

Our responses explain how Beyond the Beam (BB) connects identifier repair with score-guided candidate completion. We address the reviews through theoretical clarification, matched component analyses and the additional method comparisons requested by the reviewers.

Theory and method. Joint support, ranking and catalog-size constraints characterize feasible repairs. Minimum-replacement construction turns this characterization into repair proposals. These are evaluated as shared maps, and the selected map is then used for generator adaptation. The revised manuscript makes this connection from theory to the recommendation procedure explicit.

Recommendation quality. Matched comparisons demonstrate the contributions of construction, item-specific scoring and candidate access. On Beauty, BB improves NDCG@10 by 21.39% over enhanced Reformer. The seed-42 construction comparison improves overall NDCG@10 by 4.08% with two identifier changes. Constant-boost controls support item-specific calibrated scoring, and the additional candidate-expansion comparison favors BB in both Recall@10 and NDCG@10. Each reviewer's reply provides the relevant comparison settings and results.

Computation and certification. At a fixed ranking score and additional-evaluation budget, BB's bound priority reduces extra item evaluations, increases Top-20 certification and lowers measured batched inference time on Beauty relative to collaborative priority (submission Table 11). The bounds guide both candidate selection and stopping, connecting computational savings to certification of the returned ranking.

Experimental setting and reproducibility. The replies clarify incremental updating as new-item interactions become available and report performance across levels of target supervision. The source code is in the bb_code/ subdirectory of the original anonymous repository. The README added at its root explains the code organization and experimental workflow.

The revised Appendix E includes the additional method comparisons, the Beauty seed-42 construction results and the scoring controls.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.