PaperlayerView exact revision on arXiv

Source-linked AI summary

From Item Scores to Semantic Tokens: Test-Time Collaborative Adaptation for Generative Recommendation

Xiongfeng Shan, Xing Tang, Hanchong Chen, Fuyuan Lyu, Lingjie Li, Jingyan Jiang, Duguang Liu, Xiuqiang He

arXiv:2610.32098v1cs.IR

TL;DR

Semantic beam search can irreversibly prune an emerging-interest target, while a discriminative recommender may rank it highly from the same history. TCBSA projects discriminative score margins onto shared semantic edges and gates their influence on frozen-generator beam scores. Across four full-catalog datasets, it achieves the best results on all 16 dataset–metric pairs, with significant gains over the strongest baselines in 14 comparisons.

  • Problem

    Semantic beam search can irreversibly prune promising paths, while item-level discriminative scores cannot directly guide the generator’s prefix-conditioned token choices.

    Source: §1 · pp. 1–2, §3 · p. 3

  • Method

    TCBSA projects discriminative score margins onto shared semantic edges and uses a confidence gate to adjust frozen-generator beam scores before pruning.

    Source: §1 · p. 2, §4.1 · p. 4, §4.2 · p. 5

  • Results

    Across four full-catalog datasets, TCBSA achieves the best results on all 16 dataset–metric pairs, with significant gains over the strongest baselines in 14 comparisons.

    Source: §6 · p. 9

  • Takeaways & Limitations

    Target-path removal eliminates newly gained outside-set hits, supporting a contribution from shared-prefix corrections to correct recommendations beyond the retrieved set.

    Source: §5.5 · p. 8

    The maximum softmax-normalized discriminative score is a relative confidence proxy, not a calibrated probability of correctness.

    Source: §4.2 · p. 5

Abstract

from arXiv · show

Generative recommenders using semantic identifiers predict items through successive token decisions, but beam search may irreversibly prune promising item paths. This limitation is particularly consequential when a few recent interactions signal an emerging interest while earlier behavior favors established preferences. Yet an independently trained discriminative recommender may rank the same items highly from the same interaction history. Such complementary scores are available at test time, offering a way to adapt a frozen generator without retraining either model. The challenge is to use these scores before pruning: combining completed recommendation lists cannot alter earlier beam decisions, while scores over complete items cannot directly inform choices among semantic prefixes shared by multiple items. We propose \textbf{T}est-Time \textbf{C}ollaborative \textbf{B}eam \textbf{S}earch \textbf{A}daptation (\textbf{TCBSA}), a training-free method that incorporates collaborative information into semantic decoding. TCBSA transforms discriminative score margins into potentials over shared semantic edges and uses a confidence gate to regulate their contributions to cumulative path scores before beam selection. We establish a sufficient condition for target prefix survival under global beam competition and bound score perturbations under low discriminative confidence. Full-catalog experiments show consistent improvements over strong baselines across datasets and different generative and discriminative recommenders. Analyses of emerging interests and decoding behavior demonstrate gains in recommendation accuracy and target path survival, while targeted interventions support the contribution of shared prefix corrections. TCBSA also offers a favorable balance between recommendation accuracy and inference latency compared with simply widening the native beam.

1 Introduction

Semantic beam search can prune an emerging-interest target before its identifier is complete, even when a discriminative recommender ranks it highly. TCBSA uses collaborative scores to adapt frozen-generator beam decisions before pruning.

1 Introduction

  • A frozen discriminative recommender can rank an emerging-interest target highly from the same history even after the generator prunes its semantic path.Figure 1 illustrates a target ranked second by the discriminative recommender despite the generator pruning its shared prefix.Source: §1 · p. 1, Figure 1 · p. 2
  • Prior methods use training-time signals, completed-list fusion, retrieval, or guided search, but the item-to-token mismatch makes intermediate collaboration a distinct challenge.Completed-list fusion cannot change prefixes already pruned, and item-level scores cannot be applied directly to token decisions shared across item paths.Source: §1 · p. 2
  • TCBSA transforms discriminative item-score margins into corrections on shared semantic edges, influencing beam selection before pruning without updating either model.The introduction also identifies a sufficient target-prefix survival condition and analyzes confidence-tempered score perturbations and inference complexity.Source: §1 · p. 2
  • Across four datasets and three generative–discriminative model pairings, full-catalog experiments show improvements, with analyses covering emerging interests, decoding, shared-prefix attribution, and beam-width trade-offs.The paper also reports a sufficient condition for target-prefix survival and analyzes score perturbations under confidence-tempered adaptation.Source: §1 · p. 2

2 Related Work

Related work incorporates collaborative information during training, after generation, or through inference-time guidance and adaptation. TCBSA differs by using retrieval-relative item-score margins to adjust a frozen generator’s beam scores before pruning.

2 Related Work

  • Training-time approaches add collaborative signals, behavioral priors, or prefix-aware objectives, while TCA4Rec, APAO, and BEAR address training or beam-search behavior through learned objectives.These approaches modify generator training.Source: §2 · p. 2
  • Inference-time methods verify retrieved candidates, combine generative retrieval with dense reranking, use reward or value guidance, or fuse completed recommendation lists.MemGen adaptively combines completed discriminative and generative lists; TCBSA instead changes beam scores before pruning.Source: §2 · p. 3
  • Test-time adaptation includes parameter updates and retrieval-based item-level refinement; TCBSA adapts semantic decoding with an independently trained discriminator while keeping both models frozen.Other decoding-time control methods reweight a frozen generator, constrain beam search, or use a stronger model to select token candidates.Source: §2 · p. 3

3 Problem Formulation

The formulation contrasts autoregressive semantic-token generation with discriminative scoring of complete items. It seeks to use item-level collaborative scores in a frozen generator’s intermediate decisions before beam pruning makes a target unreachable.

3 Problem Formulation

  • A generative recommender represents each item as a length-L semantic identifier and expands it autoregressively, retaining the globally highest-scoring B partial identifiers at each depth.Once a target prefix is pruned, its descendant items become unreachable.Source: §3 · p. 3
  • A discriminative recommender assigns scores to complete catalog items and retrieves the R highest-scoring items for each user history.Source: §3 · p. 3
  • The objective is to adapt a frozen generator using independently trained discriminative item scores, making them actionable during decoding before useful paths are irreversibly pruned.The two models make predictions at different granularities: complete items versus prefix-conditioned token decisions.Source: §3 · p. 3

4 Methodology

TCBSA adapts frozen generative decoding at test time by transforming discriminative item-score margins into semantic-edge potentials and applying them before beam pruning. A confidence gate regulates these corrections, which alter cumulative path scores while leaving the generator unchanged.

4 Methodology

  • TCBSA adds confidence-tempered collaborative edge potentials to cumulative path scores before each beam-pruning step, then maps completed identifiers to ranked items.The generator’s native token probabilities and parameters remain unchanged.Source: §4 · p. 3, §4.2 · p. 5

4.1 Cross-Space Collaborative Score Transformation

  • Item-score margins above the discriminative Top-R cutoff are aggregated over shared semantic edges to guide token extensions; items outside Top-R can benefit from edges they share with retrieved items.Margins at or below the cutoff are zero and invariant to additive shifts in discriminative scores.Source: §4.1 · p. 4

4.2 Confidence-Tempered Beam Adaptation

  • A confidence gate scales collaborative corrections according to discriminative confidence, tempering their influence when predictions are uncertain.The method uses maximum softmax concentration over the full catalog as a relative confidence proxy, not a calibrated probability of correctness.Source: §4.2 · pp. 4–5
  • The analysis gives a sufficient condition for target-prefix survival under global beam competition and bounds score perturbations at low discriminative confidence.Unsupported edges receive no local correction, though their prefixes may have accumulated collaborative support.Source: §4.2 · p. 5

5 Experiments

Across four full-catalog datasets, the experiments assess TCBSA’s overall effectiveness, emerging-interest adaptation, changes to beam search, design choices, and accuracy–latency trade-offs.

5.1 Experimental Setup

  • Evaluation ranks against the full item catalog without candidate sampling and reports Recall@K and NDCG@K for K∈{5, 10}; user-level paired bootstrap uses 10,000 resamples at p<0.05.Baselines span discriminative, generative, training-time-enhanced, and post-hoc fusion methods. TCBSA uses three frozen model pairings; remaining adaptation hyperparameters are tuned on TIGER–SASRec and reused unchanged for the others, with default B=20 and R=500 chosen by macro-averaged validation NDCG@10.Source: §5.1 · pp. 5–6

5.2 Overall Effectiveness and Generality (RQ1)

  • Across four datasets, TCBSA variants led all 16 dataset–metric pairs, with significant gains in 14 comparisons; improvements extended across the tested generator–discriminator pairings.With SASRec, TCBSA improved both TIGER and MiniOneRec; with GRU4Rec, it improved TIGER. The default TIGER–SASRec pairing also outperformed the listed training-time and post-hoc fusion baselines across datasets and metrics.Source: §5.2 · p. 6, Table 1 · p. 6

5.3 Adaptation to Emerging Interests (RQ2)

  • NDCG@10 gains reached 16.29% and 31.95% on Science and 76.31% and 103.29% on Yelp for one- and two-touch histories, respectively.TCBSA improved both Recall@10 and NDCG@10 over frozen TIGER across all four emerging-interest subsets, and outperformed Human-Prior and the fusion baselines in NDCG@10. Its comparison with SASRec varied: SASRec remained stronger on Yelp one-touch, while TCBSA led both metrics on Science and led NDCG@10 on Yelp two-touch; the distinct cohorts and small two-touch samples preclude a within-user interpretation.Source: §5.3 · pp. 6–7, Table 2 · p. 6

5.4 Analysis of Decoding Behavior (RQ3)

  • At Yelp’s final semantic depth, target-prefix survival rose from 6.72% to 8.37% (+24.44%); TCBSA also rescued 2,163 of 28,385 first-pruned prefixes.Of the same-step rescues, 45.54% retained the complete target path, and 64.16% of those paths entered the final Top-10. Rescued tokens had median native rank 22, while total Hit@10 instances rose from 1,393 to 1,738 despite some displaced prefixes.Source: §5.4 · p. 7, Figure 3 · p. 7
  • At R=500, correct TCBSA Top-10 hits outside SASRec’s Top-R set accounted for 5.16% on Science, 4.72% on Office, 0.89% on Music, and 0.86% on Yelp.Such outside-set hits persisted even at R=2000, so the retrieved set did not impose a hard constraint on generative outputs.Source: §5.4 · p. 7, Figure 4 · p. 7

5.5 Ablation and Attribution Analysis (RQ4)

  • Removing confidence tempering reduced NDCG@10 by 1.88%–9.08%, while aggregation variants showed maximum absolute relative deviations of 0.83%–1.89%.No tested aggregation rule consistently performed best across datasets.Source: §5.5 · p. 8, Table 3 · p. 8
  • Removing target-path corrections eliminated 29.89% of Office and 28.46% of Science outside-set hits, versus negligible losses under matched controls.All newly gained outside-set hits on both datasets disappeared after target-path removal; the diagnostic uses ground-truth targets offline, and lost Top-10 hits can reflect reranking rather than beam pruning.Source: §5.5 · p. 8, Table 3 · p. 8

5.6 Accuracy–Latency Trade-offs (RQ5)

  • At B=50, TCBSA’s NDCG@10 gains over TIGER ranged from 4.78% on Office to 29.21% on Yelp, and it beat both fusion baselines at every B≥20.At B=20, it achieved 3.64%–26.94% higher NDCG@10 than TIGER at B=50 while using 22.95%–32.27% less latency; increasing TCBSA’s beam from 30 to 50 added only 0.14%–0.57% NDCG@10 at 36.78%–42.09% higher latency.Source: §5.6 · pp. 8–9, Figure 5 · p. 8

6 Conclusion

TCBSA adapts frozen generative decoding by projecting discriminative item-score margins onto shared semantic edges and applying confidence-tempered beam-score corrections before pruning. Across the reported evaluations, it improves recommendation outcomes and can provide a more favorable accuracy–latency balance than widening native search.

6 Conclusion

  • TCBSA projects discriminative item-score margins onto shared semantic edges and applies confidence-tempered corrections before beam pruning, without updating either model.The paper establishes a sufficient condition for target-prefix survival under global beam competition and bounds score perturbations under low discriminative confidence.Source: §6 · p. 9
  • Across four full-catalog datasets, TCBSA variants achieved the best results on all 16 dataset–metric pairs, with significant gains over the strongest baselines in 14 comparisons.The conclusion also reports gains across different generative–discriminative pairings, target-prefix recovery, and intervention evidence that shared-prefix corrections contribute to correct recommendations outside the discriminative Top-R set.Source: §6 · p. 9
  • At B=20, TCBSA achieved higher NDCG@10 and lower latency than native TIGER at B=50, despite added latency at matched beam widths.The conclusion frames these findings as support for incorporating collaborative scores before beam pruning rather than only combining final rankings.Source: §6 · p. 9

A Theoretical Analysis and Complexity

The analysis characterizes how TCBSA changes cumulative path scores, when a target prefix survives global beam competition, and the method’s inference costs. Its survival and stability results specify conditions under which adaptations preserve or change decoding order.

A.1 Cumulative Test-Time Score Decomposition

  • With the history-specific gate fixed during decoding, collaborative edge potentials add to native cumulative token log-probability scores without changing generator token probabilities.Source: §A.1 · p. 11

A.2 Target-Prefix Survival under Global Beam Competition

  • When designated competitors score below the target and at most B−1 others can match or exceed it, the target ranks within the beam; meeting this condition at every depth preserves its complete path.Source: §A.2 · p. 11
  • The confidence gate bounds changes to native pairwise score differences; orderings persist when the native score gap exceeds that bound.Source: §A.2 · p. 11

A.3 Inference Complexity

  • Preprocessing costs O(|I|L), potentials add O(RL) storage and per-expansion lookup-addition, preserving native asymptotic scaling but not guaranteeing negligible latency.Source: §A.3 · pp. 11–12

B.1 Dataset Construction and Statistics

The appendix describes dataset preparation and validation-based configuration choices for TCBSA and the comparison methods.

B.1 Dataset Construction and Statistics

  • Amazon domains use official 5-core chronological sequences, while Yelp uses a 2019 review window with chronological ordering and iterative 5-core filtering; each dataset uses leave-one-out test and validation interactions.The input history limit is 20 and does not affect the reported dataset statistics.Source: §B.1 · p. 12, Table 4 · p. 12

B.2 Final Configurations and Computing Environment

  • Validation NDCG@10 on TIGER–SASRec selected TCBSA settings with R=500, reused across model pairs; MemGen fusion weights and gates were also validation-selected.Experiments and inference-efficiency comparisons used the same NVIDIA RTX 5880 Ada Generation accelerator model.Source: §B.2 · p. 12

C.1 Shared-Prefix Attribution

Target-path interventions show that some correct recommendations depend on shared-prefix corrections, while the authors caution that not every lost hit reflects beam pruning.

C.1 Shared-Prefix Attribution

  • All 47 newly gained outside-set hits disappeared after target-path correction removal; Table 5 reports the shared-prefix attribution results.Source: Table 5 · p. 12, §C.1 · p. 12
  • Target-path removal eliminated 11/34 Music and 12/15 Yelp original outside-set hits, with no losses under matched controls.Source: §C.1 · p. 12
  • Only 44/159 Office, 9/35 Science, 6/11 Music, and 9/12 Yelp lost original hits involved observed target-path beam loss; the remaining losses did not.The authors therefore do not claim every lost hit results from beam pruning or that recall improves across the full outside-retrieval cohort.Source: §C.1 · p. 12
  • Positive target-path support was observed in 505/532 Office, 118/123 Science, 33/34 Music, and 15/15 Yelp cases, alongside strength-matched control diagnostics.Source: §C.1 · p. 12

C.2 Design and Parameter Sensitivity

Ablations assess the confidence gate and edge aggregation, while validation sensitivity tests vary the adaptation scale, gate slope, and confidence threshold around a common reference.

C.2 Design and Parameter Sensitivity

  • Around the common reference, the largest relative NDCG@10 decreases were 1.08% for λ on Yelp, 3.02% for k on Yelp, and 2.21% for τ on Office.Figure 6 shows one-factor validation sensitivity; these results describe robustness around the common reference, not dataset-specific retuning.Source: Figure 6 · p. 13, §C.2 · p. 13
  • Removing the confidence gate reduces NDCG@10 by 1.88%–9.08% across four datasets; alternative aggregators’ largest absolute relative deviations range from 0.83% to 1.89%.Some alternative aggregators achieve slightly higher test NDCG@10, while the validation-selected configurations remain unchanged.Source: Table 6 · p. 13, §C.2 · p. 13

C.3 Retrieval-Width Selection

Retrieval width is selected using validation NDCG@10: the macro average peaks at R=500, although dataset-specific optima differ.

C.3 Retrieval-Width Selection

  • Widths R∈{20, 100, 500, 1000, 2000} are evaluated on complete validation sets with other settings fixed at each dataset’s final R=500 configuration.Source: §C.3 · p. 13
  • The macro-average validation NDCG@10 peaks at R=500, although dataset-specific optima differ.Table 7 reports validation NDCG@10 across retrieval widths.Source: §C.3 · p. 13, Table 7 · p. 13
  • R=500 is the smallest width within 0.5% of the best macro-average NDCG@10 and with every dataset deficit at most 1.5%.Its largest per-dataset deficit is 0.70% on Music, computed from unrounded validation results; no test-set results are used for selection.Source: §C.3 · p. 13

Related papers

Loading 2610.32098v1…