Source-linked AI summary
TAAL: Mitigating Early Beam Pruning in Generative Recommendation via Temporal Autoregressive Alignment
Lianjie Li, Zhiying Tu, Dianhui Chu, Hongliang Sun
TL;DR
Early beam pruning dominates retrieval failures because next-token prediction does not explicitly cover multimodal, history-conditioned transitions. TAAL aligns joint early-prefix distributions during training and calibrates inference with PMI, improving benchmark retrieval performance and full-SID survival.
Problem
91.9%–96.6% of retrieval failures occur within the first two decoding steps, where pointwise next-token prediction does not explicitly cover multimodal transition modes.
Method
TAAL aligns a history-conditioned joint (c1, c2) soft target with forward KL during training and applies PMI calibration to candidate scores during inference.
Results
TAAL improves NDCG@10 over the standard baseline by 39.5%, 6.7%, and 28.6% on Beauty, Instruments, and Yelp, respectively, while increasing full-SID survival by 3.9%–16.6%.
Takeaways & Limitations
Joint-prefix alignment addresses vulnerable early branches, while PMI provides additional candidate-ranking calibration without changing the end-to-end generative paradigm.
Takeaways & Limitations
The transition model uses first-order contextual transitions, experiments cover three medium-scale public datasets, and the main results use a fixed random seed of 42.
Abstract
from arXiv · showhide
Generative recommendation encodes items as hierarchical semantic identifiers (SIDs) and retrieves the next item through autoregressive decoding. Standard next-token prediction, however, does not explicitly cover the multimodal transitions present in interaction sequences, leaving the ground-truth SID vulnerable to irreversible pruning at early beam-search branches. Across three public benchmarks, we find that 91.9\%--96.6\% of retrieval failures occur within the first two decoding steps. We therefore propose Temporal Autoregressive Alignment (TAAL). During training, TAAL constructs a joint $(c_1,c_2)$ soft target from historical transitions and aligns the early-prefix distribution with a forward KL objective. During inference, it calibrates candidate scores with pointwise mutual information (PMI) to reduce the influence of globally frequent prefixes. On Amazon Beauty, Instruments, and Yelp, TAAL improves NDCG@10 over the standard baseline by 39.5\%, 6.7\%, and 28.6\%, respectively, while increasing full-SID survival by 3.9\%--16.6\%. Beam-width analysis further shows that the relative survival gain grows as the beam narrows, reaching 39.4\% at $B=5$.
1 INTRODUCTION
Generative recommendation uses hierarchical SIDs and autoregressive decoding, but standard next-token prediction can prune the ground-truth path at early branches. TAAL addresses this with history-conditioned joint-prefix alignment during training and PMI calibration during inference.
- Method: TAAL forms a joint (c1, c2) soft target from historical transitions and aligns the model distribution with forward KL.The alignment is designed to cover multiple plausible transition modes at vulnerable early branches.
- Method: PMI calibration normalizes global prefix frequency and reranks candidates within the beam in O(B) time.It complements training-time joint-prefix alignment by using temporal collaborative lift during inference.
- Early-pruning diagnosis: 53.0%–71.1% first-token mispruning and 91.9%–96.6% of failures within two steps expose early decoding as the main retrieval bottleneck.These errors are irreversible under trie-constrained beam search because removing a prefix removes its entire ground-truth subtree.
- Results: 39.4% relative survival gain at B = 5 shows that TAAL’s benefit increases as the beam narrows.The beam-width analysis covers B ∈{5, 10, 20, 50}.
2 PRELIMINARIES AND THE EARLY-PRUNING CLIFF
Generative recommendation represents items with hierarchical semantic identifiers and decodes them through trie-constrained beam search. Because early prefixes determine entire subtrees, standard next-token training can cause irreversible ground-truth pruning before later decoding steps.
- Hierarchical SIDs: Items are mapped to fixed-length hierarchical code tuples, with c1 selecting a coarse branch and c2 refining it.Deeper prefixes identify items at progressively finer granularity.
- Beam-search decoding: Trie-constrained beam search guarantees valid SIDs, but any prefix leaving the beam permanently removes all items beneath it.This makes early beam errors structurally irreversible.
- Early-pruning cliff: 56.4%/53.0%/71.1% first-step mispruning on Beauty/Instruments/Yelp is followed by another 21.3%–30.2% at the second token.The standard LETTER-TIGER baseline is evaluated at B=20.
- TAAL overview: TAAL combines joint-prefix alignment during training with PMI calibration during inference to address multimodal transitions and globally frequent prefixes.The training and inference operations target the two early-branching challenges identified by the analysis.
3 TAAL: JOINT TRAINING AND PMI PRIOR CALIBRATION
TAAL addresses early-prefix retrieval errors by aligning the joint (c1, c2) distribution during training and calibrating completed beam candidates with PMI during inference.
- Motivation: The joint (c1, c2) prefix distribution targets the early branches where retrieval failures concentrate.The training design focuses on the first two SID positions, which constrain subsequent legal paths.
- Joint-prefix transition modeling: Historical joint-prefix transitions are aggregated over recent items with exponentially decaying weights, while empty histories use the global joint-prefix marginal.Recent behavior receives greater weight, but longer-range transitions retain nonzero probability mass.
- Training alignment: Forward KL decomposes into an exact first-token marginal term and a conditional second-token term estimated by sampling Kmc = 2 branches.The estimator is unbiased, avoiding enumeration of every legal second-token candidate.
- Training alignment: Forward KL is selected for mode coverage, penalizing near-zero model probability on empirically supported transitions.The auxiliary alignment objective applies only to the first two decoding positions.
- Inference calibration: PMI calibration adjusts beam-candidate scores by temporal association relative to global prefix frequency rather than directly adding the empirical prior.Calibration reranks the completed Top-B beam in O(B) time without changing the autoregressive decoder.
- Complementary operations: Training expands coverage of empirically supported paths, whereas inference adjusts candidate ranks using temporal associations with the current history.The two operations address candidate generation and candidate ranking as complementary error sources.
4 EXPERIMENTS
Experiments evaluate TAAL across three benchmarks, ablations, survival diagnostics, structure controls, beam-width sweeps, and sensitivity analyses. Results show that joint-prefix training drives most gains, while PMI provides smaller ranking improvements and TAAL is especially beneficial under narrower beams.
- Main results: TAAL with PMI improves NDCG@10 over the standard baseline by 39.5%, 6.7%, and 28.6% on Beauty, Instruments, and Yelp.The corresponding TAAL NDCG@10 values are 0.0484, 0.0881, and 0.0293.
- Ablations: Joint training provides most of the gain, improving NDCG@10 over α = 0 by 37.5%, 5.9%, and 25.7% across the three datasets.PMI adds smaller matched-run improvements of 1.5%, 0.8%, and 2.3%.
- Beam survival and early-pruning mitigation: PMI changes only post-survival ranking because it runs after beam search, leaving the candidate set and survival rates unchanged.Average surviving-target rank improves from 7.21 to 7.11 on Beauty and from 5.13 to 5.02 on Instruments.
- Beam survival and early-pruning mitigation: Joint c1c2 alignment improves second-prefix survival to 15.76%/17.75%/8.58% and raises full-SID survival over baseline by 3.9%–16.6%.The gains are reported for Beauty, Instruments, and Yelp in dataset order.
- Structure-specific mechanism diagnostics and controls: History-aligned TAAL outperforms global-marginal and history-shuffled controls across the tested weight range, while both controls degrade as α increases.At α = 0.2, TAAL reaches NDCG@10 .0477 and Sfull 11.75% on Beauty; at α = 0.4 it remains at .0455.
- Beam-width sensitivity and pruning pressure: TAAL’s relative full-SID survival gain increases as the beam narrows, reaching +39.4% at B = 5.The reported gain is +27.7% at B = 10, and wider beams narrow the gap.
5 RELATED WORK
TAAL addresses early-pruning risks in generative recommendation through history-conditioned transition information, complementing tokenizer-focused and prefix-optimization methods. It modifies recommender training and candidate calibration while using LETTER-TIGER as its backbone.
- Generative recommendation: Earlier generative recommendation established autoregressive retrieval by combining residual-quantized semantic identifiers with sequence generation.The paradigm followed earlier discriminative sequential recommenders and language-task formulations.
- TAAL’s positioning: TAAL operates after SID construction, modifying recommender training and candidate calibration rather than optimizing tokenizers.It uses LETTER-TIGER as the backbone, making the directions complementary.
- Constrained decoding: TAAL estimates history-conditioned (c1, c2) transition distributions, covers multiple plausible prefixes, and removes global marginal frequency with PMI at inference.This contrasts with APAO and BEAR, which primarily supervise the positive target path and model-internal ranking state.
- Collaborative signal injection: TAAL uses a statistical transition table to construct training soft targets and does not introduce an additional language model.This extends classical collaborative transition and co-occurrence signals into SID generation.
6 CONCLUSION AND LIMITATIONS
TAAL targets early beam-pruning failures by aligning training with multimodal temporal transitions and calibrating inference with PMI. Across three benchmarks, it improves survival and ranking, while its evaluation remains limited in transition order, dataset scale, and random-seed coverage.
- Conclusion: 91.9%–96.6% of retrieval failures across three benchmarks occur within the first two decoding steps.TAAL applies joint-prefix KL during training and PMI at inference to address this early-pruning pattern.
- Conclusion: Full-SID survival improves by 3.9%–16.6%, while NDCG@10 reaches 0.04842/0.08810/0.02926 on Beauty/Instruments/Yelp.The reported results also include controls showing that history-aligned transitions are more effective than global-marginal or shuffled alternatives.
- Limitations: The transition model uses first-order contextual transitions, leaving higher-order and multihop temporal behavior open.This is identified as a limitation of the transition modeling scope.
- Limitations: Experiments cover three medium-scale public datasets, so generalization to very large industrial catalogs requires further study.The main results also use a fixed random seed of 42.
A.1 DATASET STATISTICS
Table 5 reports statistics for the three benchmark datasets before leave-one-out holdout, including density defined relative to users and items.
- Dataset statistics: Table 5 summarizes the three benchmark datasets before leave-one-out holdout.It defines density as interactions divided by the product of the number of users and items.
A.2 COMPLETE BEAM-SURVIVAL RESULTS
Table 6 reports ground-truth survival and ranking diagnostics at decoding stages with beam width 20. PMI calibration occurs after candidate generation, so it changes ranking but not survival rates.
- Beam-survival results: Table 6 presents ground-truth survival and ranking diagnostics across decoding stages at beam width 20.The table separates survival behavior from ranking outcomes.
- Beam-survival results: PMI calibration leaves all three survival rates unchanged because it runs after candidate generation, affecting post-survival rank and final ranking metrics.Its role is therefore candidate reranking rather than changing which paths survive beam search.