acceptodds
Under review as a conference paper at ICLR 2027

Order-Marginalized Generation of Semantic and Collaborative IDs for Generative Recommendation

Abstract

Generative Recommendation (GR) increasingly incorporates collaborative signals into item tokenization to complement content semantics. When semantic and collaborative representations are separately quantized into two identifier chains, however, their cross-chain generation order is inherently unspecified. Their combination therefore forms a partially ordered target, whereas autoregressive generation requires a total order. Existing methods consequently rely on a prescribed cross-identifier serialization, an arbitrary choice that can materially affect recommendation performance. We propose LatticeRec, an order-marginalized framework that treats cross-identifier serialization as a latent variable rather than a prescribed generation target. The key idea is to represent partial generations by canonical states determined only by the revealed prefixes of the two identifiers, allowing different interleavings with equivalent revealed information to share the same predictive state. This organizes all valid interleavings into a compact lattice and enables exact marginalization over generation orders via dynamic programming, without explicit path enumeration. We further introduce coverage-aware training to prevent supervision from concentrating on only a few generation paths. Experiments on three benchmarks demonstrate the effectiveness of LatticeRec and its key components.

Primary area:
generative models
Keywords:
Generative Recommendation, Order Marginalization, Semantic and Collaborative IDs

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_sNexOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In generative recommendation, semantic IDs and collaborative IDs encode item content and interaction patterns, respectively. Each identifier chain has an internal order, but there is no natural order between the two chains; autoregressive generation nevertheless requires a total order. LatticeRec treats the cross-chain interleaving as a latent variable. It represents partial generations by canonical states defined by the revealed prefixes of both chains, merges paths that reach the same state, and marginalizes over valid orders using dynamic programming. Coverage regularization is added to prevent training from focusing on only a few paths. The paper reports improvements on three datasets.

Critical strengths
  • The motivation is clear and well grounded. The method directly addresses the structural mismatch between two internally ordered but mutually unordered identifier chains and the total order required by autoregressive generation.
  • Canonical states provide a simple, effective core idea. Paths that reveal the same prefixes in both chains share a predictive state. Dynamic programming can then aggregate their probability mass over a lattice rather than enumerate every complete interleaving.
  • Coverage regularization targets a specific training issue. Marginal likelihood can let a few high-posterior paths dominate the gradients. The proposed regularizer maintains supervision for other valid interleavings.
Critical weaknesses
  • Exactness relies on a key modeling assumption. Path merging is valid because the model is designed so that the next-step prediction depends only on the revealed prefixes, not on the order in which they were generated. Thus, dynamic programming is exact for the model defined by LatticeRec, but it is not a lossless reduction for a general path-dependent autoregressive model. The paper should better justify why cross-chain arrival order is irrelevant and discuss the resulting expressivity tradeoff.

  • The evidence that marginalization is better than a strong fixed order is incomplete. The paper shows that performance varies across fixed serializations, but the direct comparison with the best fixed order is mainly reported on Games. The results support the claim that serialization choice can matter, but do not establish that marginalization consistently outperforms a well-selected fixed order across datasets.

  • Uniform path coverage introduces an additional prior. The regularizer uses a uniform reference distribution to prevent the posterior from concentrating on a few paths. This addresses training coverage, but also encourages every valid interleaving to receive some weight. The paper could better explain when this uniformity is appropriate and when it might conflict with useful order preferences in the data.

  • The guarantees for training and inference should be distinguished more clearly. The lattice dynamic program exactly aggregates valid paths during training. At inference, beam search prunes states, so only the retained paths are aggregated exactly; overall inference is approximate. The paper should emphasize this distinction and analyze how beam search affects the main results.

Questions for authors

Please refer to Weaknesses.

4 comments

We sincerely appreciate the reviewer for spending valuable time in reviewing our paper and providing detailed comments. We hope that our response below will address your concerns.

W1: Path Invariance Assumption and Inductive Bias

We consider path invariance a reasonable inductive bias for multi-view ID generation. Semantic and collaborative IDs are quantized separately, each with a fixed internal order, but no cross-view generation order is given in the data. We therefore let the model predict from the revealed prefixes of both views, rather than additionally conditioning on the order in which they were revealed.

Importantly, this does not eliminate generation-order preferences. For example, paths ρA=(s1,c1)\rho_A=(s_1,c_1) and ρB=(c1,s1)\rho_B=(c_1,s_1) reach the same canonical state σ1,1\sigma_{1,1}. Their path probabilities can differ; only their subsequent predictive distributions are shared:

=p_\theta(\mu,z\mid x,\rho_B) =p_\theta(\mu,z\mid x,\sigma_{1,1}).$$ The model can therefore still learn different path preferences, as illustrated in Figure 5. We acknowledge the expressivity tradeoff: unlike a fully path-dependent autoregressive model, LatticeRec cannot distinguish the future predictions of paths reaching the same canonical state. However, since the data do not specify a cross-view generation order, we consider predicting from the revealed information a reasonable inductive bias. This reduces unnecessary modeling freedom by avoiding separate continuation functions for the same revealed prefixes, potentially improving generalization. It also enables exact marginalization within our model through DP without enumerating all valid paths. Our state-sharing ablation further supports this design, showing better recommendation performance under the evaluated settings. > W2: Comparison with the Best Fixed Order Across Datasets To address the limited cross-dataset evidence, we extended the comparison of all 20 fixed generation orders to Instruments and Steam, and repeated Games with two additional random seeds. All results are NDCG@10. Best-Fixed is selected based on validation performance, while Fixed-Avg reports the average test performance across the 20 fixed-order models. |Dataset|Best Order|Fixed-Avg|Best-Fixed|LatticeRec| |:-:|:-:|:-:|:-:|:-:| |Instruments|CSSCCS|0.0293|0.0332|**0.0339**| |Games (seed 42)|SCSCCS|0.0521|0.0533|**0.0536**| |Games (seed 43)|SCSSCC|0.0517|0.0533|**0.0539**| |Games (seed 44)|SSCSCC|0.0521|0.0536|**0.0541**| |Steam|CSCSCS|0.0527|0.0538|**0.0544**| [Games, seed 42](https://anonymous.4open.science/r/rebuttal_image-3E4F/rebuttal_image/games_seed42.png) · [Games, seed 43](https://anonymous.4open.science/r/rebuttal_image-3E4F/rebuttal_image/Games%20seed%2043.png) · [Games, seed 44](https://anonymous.4open.science/r/rebuttal_image-3E4F/rebuttal_image/games%20seed%2044.png) [Instruments, NDCG@10](https://anonymous.4open.science/r/rebuttal_image-3E4F/rebuttal_image/instruments_ndcg10.png) · [Instruments, Recall@10](https://anonymous.4open.science/r/rebuttal_image-3E4F/rebuttal_image/instruments_R10.png) [Steam, NDCG@10](https://anonymous.4open.science/r/rebuttal_image-3E4F/rebuttal_image/steam_ndcg10.png) · [Steam, Recall@10](https://anonymous.4open.science/r/rebuttal_image-3E4F/rebuttal_image/steam_recall10.png) LatticeRec outperforms Best-Fixed on all three datasets and across all three seeds on Games. Notably, the best fixed order varies across datasets and even across random seeds within Games, making it difficult to identify a suitable order in advance. Moreover, finding Best-Fixed requires training and selecting among 20 fixed-order models, whereas LatticeRec learns across generation orders within a single model. These results provide direct cross-dataset evidence that order marginalization can outperform a well-selected fixed order without requiring an expensive search over generation orders.

W3: Uniform Coverage and Order Preferences

Coverage regularization introduces a uniform supervision bias, but does not assume that all generation orders should be equally probable. Since no cross-view order is given in the training data, its purpose is to prevent potentially useful paths from being overlooked due to insufficient supervision. Our analysis of premature path lock-in in Reviewer 1's W3/Q3 further motivates this design.

Importantly, even a small amount of uniform supervision can provide substantial benefits. Figure 4 shows that increasing λ from 0 to 0.2 improves both Effective Paths and NDCG@10 across all three datasets, while larger λ does not always yield further gains.

The key is balancing path coverage with learned order preferences. Uniform coverage is useful when alternative paths have not been adequately explored, but excessive coverage may weaken the learning signal from more informative orders. Our experiments reflect this trade-off: Instruments2018 and Steam perform best at λ=0.8 and λ=0.6, respectively, rather than λ=1, whereas Instruments and Games favor stronger coverage (Effect of coverage weight on NDCG@10). These results suggest that the appropriate coverage strength depends on the balance between supervising alternative paths and learning meaningful order preferences.

W4: Exact Marginalization in Training and Beam Approximation at Inference

We clarify the distinction between training and inference. During training, under the path-invariance assumption (Eq. 10), DP exactly marginalizes over all valid generation paths for each target item without explicit enumeration. At inference, State-Merged Beam Search exactly aggregates the probabilities of surviving paths, but pruning discards some probability mass, making the overall inference approximate.

To quantify this approximation, we rescore the beam-retrieved candidates using exact marginalization on all three datasets, keeping the candidate set fixed. We report Mass Retention (the ratio of beam-aggregated to exact marginal probability, summarized by median/P10 across candidates) and Overlap@10 (the average proportion of shared items between the two top-10 lists).

DatasetMass Retention (Median/P10)Overlap@10Beam N@10Exact N@10
Instruments68.91% / 33.14%99.11%0.033950.03392
Games73.87% / 35.13%99.39%0.053580.05354
Steam72.53% / 35.98%99.36%0.054390.05438

Although beam pruning loses substantial probability mass (median retention: 68.91%–73.87%), its impact on recommendation rankings is small. Top-10 overlap exceeds 99% across all three datasets, with absolute NDCG@10 differences of at most 0.00004. Thus, while beam search does not recover exact item marginals, its rankings closely match exact rescoring within the same candidate set.

Thank you for the detailed response and clarifications. After considering your response, I will keep my original rating.

Thank you for your encouraging assessment and valuable feedback. We truly appreciate the time and care you have devoted to reviewing our work.

Sincerely,

The Authors

Reviewer_7veyOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LatticeRec for generative recommendation with separately quantized semantic and collaborative IDs. It represents partial generations by the revealed prefixes of both IDs, allowing different generation orders to share a predictive state and enabling exact marginalization over valid interleavings within this model. Coverage regularization maintains supervision across paths, and state-merged beam search aggregates their probability mass during inference. Experiments on three recommendation datasets show improvements over the compared baselines.

Critical strengths
  • The comparisons across valid fixed orders show that cross-view serialization can affect recommendation quality, providing empirical motivation for examining generation order when combining semantic and collaborative IDs.

  • The canonical-state formulation makes the condition for exact path merging explicit, and the analysis explains the relationship between coverage regularization and uniform-path training.

  • The experiments include fixed-order and random-order comparisons, component ablations, a parameter-matched TIGER baseline, and an evaluation of early termination, providing several useful controls beyond the main benchmark results.

Critical weaknesses
  • The proposed method combines shared prefix states, dynamic programming, and coverage regularization for dual-view ID generation. The recurrence and the relation between coverage regularization and uniform-path training are straightforward, so the methodological and theoretical novelty appears limited.

  • The best settings on Instruments and Games use a coverage weight of 1, making the training objective equivalent to uniform-path supervision. The main results therefore do not clearly establish the independent benefit of marginal-likelihood training.

  • On Games, NDCG@10 improves only from 0.0533 to 0.0536 for LatticeRec, showing a limited advantage over a carefully selected fixed generation order.

  • This paper does not sufficiently explain how generation paths are handled without canonical-state sharing, or how the training objective, parameter count, and decoding budget are kept comparable. This makes it difficult to attribute the ablation results specifically to state sharing.

  • The efficiency experiments report inference latency and convergence by epoch, but omit actual training time, peak memory, and comparisons of recommendation quality and latency across beam sizes, leaving the computational cost insufficiently characterized.

Questions for authors
  1. For Steam, why does Max-Path yield NDCG@10 of 0.0519 in Table 2 but 0.0527 under L3-K256 in Table 4, while the full model scores 0.0535 in both? Were different checkpoints or decoding settings used?

  2. Are transition probabilities renormalized after applying catalog constraints? When a prefix uniquely identifies an item, how does its accumulated probability mass relate to the item's score under full decoding?

5 comments

We sincerely thank the reviewer for taking the time to evaluate our paper and for providing valuable feedback. We hope our responses below adequately address your concerns. Please feel free to let us know if any further clarification is needed.

W1: Novelty of the Problem Formulation and Method

Our novelty lies not in the DP recurrence or KL relation themselves, but in introducing and formalizing a new generative modeling problem and developing a method to address it. We highlight three contributions:

  1. Formulating a new generative modeling problem. Semantic and collaborative IDs are quantized separately, each with a fixed internal token order, but why must their cross-view generation order be manually specified? Existing methods typically concatenate these IDs in a predefined order. We instead formulate a partially ordered generation problem, treating cross-view order as a latent variable and marginalizing over all valid interleavings. This allows the model to learn generative relationships and order preferences while preserving the structure of each ID chain.
  2. Enabling order marginalization within autoregressive generation. This is not simply an application of DP. Conventional autoregressive decoders typically follow a predefined generation path, making it challenging to model and aggregate probabilities across multiple paths. We address this by designing path-invariant shared states, where different paths reaching the same within-view prefixes share a predictive distribution. This enables exact order marginalization through DP without enumerating all generation paths.
  3. Learning path distributions and order preferences. Our goal is not merely to supervise all paths, but to learn their probability distribution and meaningful generation-order preferences. Coverage-aware training mitigates premature concentration of supervision, allowing alternative paths to remain sufficiently supervised without forcing uniform path probabilities. Our response to Reviewer 1's W3/Q3 further examines this effect.

W2: Marginal-Likelihood Training and Order Marginalization

Our goal is not to establish that pure marginal-likelihood training is independently superior. In fact, as shown in Figure 4 and our response to Reviewer 1's W3/Q3, optimizing marginal likelihood alone (λ=0) can lead to premature path lock-in, leaving alternative paths insufficiently supervised. This motivates our coverage-aware training.

Importantly, λ controls how supervision is distributed across paths, rather than whether generation orders are modeled as latent variables. At λ=1, the objective becomes uniform-path supervision, which is a Jensen upper bound on the order-marginalized negative log-likelihood (up to an additive constant). This provides an alternative training objective within the same marginalization framework, without imposing fixed-order generation or uniform path probabilities. Figure 5 shows that the learned order preferences vary across instances, while Table 4 demonstrates the benefit of marginal aggregation at inference.

Moreover, on Instruments2018 and Steam, intermediate coverage weights (λ=0.8 and 0.6, respectively) outperform λ=1 (Effect of coverage weight on NDCG@10). This suggests that retaining the marginal-likelihood component can provide additional benefits when combined with coverage supervision.

W3: Comparison with the Best Fixed Generation Order

The improvement on Games is modest, but LatticeRec aims to learn generation-order preferences without committing to a predefined order, rather than simply outperforming a carefully selected fixed order.

We extended the comparison across all 20 fixed orders to Instruments and Steam, and evaluated Games with three random seeds. All results are NDCG@10, with Best-Fixed selected on the validation set.

DatasetBest OrderFixed-AvgBest-FixedLatticeRec
InstrumentsCSSCCS0.02930.03320.0339
Games (seed 42)SCSCCS0.05210.05330.0536
Games (seed 43)SCSSCC0.05170.05330.0539
Games (seed 44)SSCSCC0.05210.05360.0541
SteamCSCSCS0.05270.05380.0544

LatticeRec consistently outperforms Best-Fixed across all three datasets and all three seeds on Games. Moreover, the best fixed order varies across datasets and even across seeds on Games, making it difficult to select a suitable order in advance. Identifying Best-Fixed also requires training and comparing 20 separate models, whereas LatticeRec learns across orders within a single model.

More importantly, LatticeRec does not simply replace one fixed order with another. Figure 5 shows that the highest-posterior path varies across instances. Our additional analysis in response to Reviewer 1's W3/Q3 further shows that reselecting the generation order improves NDCG@10 from 0.0296 to 0.0326, with the candidate set and code scores held fixed. This provides direct evidence that learning order preferences has practical value beyond avoiding the search for a fixed order.

W4: Ablation of Canonical State Sharing

We clarify the setup of this ablation. Both variants are trained using all 20 valid generation paths. LatticeRec merges paths reaching the same semantic and collaborative prefixes into a canonical state with a shared predictive distribution. Without sharing, each generation path retains its own state and computes predictions independently.

Both variants use the same training objective (λ=1, mean NLL over all 20 paths), number of trainable parameters (6,758,786), and effective batch size (3,072). At inference, both use beam width B=1000. However, LatticeRec merges equivalent paths before pruning, whereas the no-sharing variant keeps them separate, allowing redundant paths to occupy multiple beam slots.

DatasetRecall@10 (Ours / w/o)NDCG@10 (Ours / w/o)Avg. Retrieved Items (Ours / w/o)
Instruments0.0619 / 0.05860.0339 / 0.0322505.5 / 113.4
Games0.0964 / 0.09120.0536 / 0.0500493.0 / 113.8
Steam0.0961 / 0.09480.0535 / 0.0525469.9 / 98.4

Canonical state sharing improves recommendation performance across all three datasets and retrieves over four times as many distinct items at the same beam width. These results suggest that merging equivalent paths reduces redundant search and allows the model to explore more distinct items within a fixed beam width, contributing to the performance gains.

W5: Computational Cost Analysis

We provide additional measurements of training GPU-hours, peak GPU memory, and the trade-off between recommendation quality and inference latency across beam widths. Training costs are measured up to the checkpoint with the best validation performance.

DatasetModelBest EpochGPU-hours to BestPeak Memory (GiB)
InstrumentsTIGER941.2423.93
LatticeRec761.3135.71
GamesTIGER862.8623.93
LatticeRec583.0835.71
SteamTIGER863.5223.93
LatticeRec683.6135.71

Although lattice-based supervision requires more GPU memory, it enables LatticeRec to incorporate all valid generation paths within a training budget comparable to TIGER, while achieving better recommendation performance. We further evaluate recommendation quality and per-sample inference latency across beam widths on Games.

Beam WidthRecall@10NDCG@10Latency (ms/sample)
500.09470.05292.745
1000.09610.05352.836
2000.09640.05362.865
5000.09650.05363.052
10000.09640.05363.658

Recommendation performance largely stabilizes at B=200. Compared with B=1000, B=200 achieves the same Recall@10 and NDCG@10 at the reported precision, while reducing inference latency by 21.7% (2.865 vs. 3.658 ms).

Q1: Discrepancy Between Tables 2 and 4

We thank the reviewer for pointing out this inconsistency. Upon checking, we found that w/o Marginal Inference in Table 2 was mislabeled. The reported result (0.0519) actually comes from fixed-order evaluation: using the same checkpoint, we evaluated all 20 valid fixed interleavings and selected the best-performing order, rather than selecting the highest-probability path for each item (Max-Path).

In contrast, Table 4 evaluates Max-Path by replacing marginal aggregation with the highest-probability path score for each item, yielding 0.0527. The two results therefore reflect different inference settings rather than the same ablation. We have corrected the labeling and the accompanying ablation description in the revised version.

Q2: Catalog Constraints and Early Termination

As described in Section 3.4 and the appendix ("Effect of Early Termination"), catalog constraints filter out invalid token extensions without renormalizing the remaining transition probabilities, unlike the normalization over available views in Eq. 14.

When a prefix uniquely identifies an item, early termination uses its accumulated probability mass as the item score. Full decoding additionally multiplies this mass by the total conditional probability of all valid completions, which is at most 1. Thus, the early termination score upper-bounds the contribution of that prefix under full decoding, although it may alter the relative ranking of items.

Table 11 already compares both strategies on the full test sets, showing relative NDCG@10 changes within 0.20% across all three datasets. We further analyzed ranking changes on 1,024 randomly selected test instances per dataset, using the same checkpoint for both decoding strategies.

DatasetTop-10 OverlapTop-10 Pairwise Rank InversionΔNDCG@10
Instruments99.07%0.62%1.8e-5
Games99.24%0.56%−1.2e-5
Steam99.33%0.49%0

ΔNDCG@10 = Early Termination − Full Decoding. Pairwise Rank Inversion measures reversed pairwise orderings among items shared by both top-10 lists.

Top-10 overlap exceeds 99% across all three datasets, with pairwise rank inversion below 0.7% and negligible NDCG@10 changes. These results indicate that early termination can alter item scores but has little impact on recommendation rankings.

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

LatticeRec addresses generation order when combining semantic and collaborative item IDs. It merges equivalent partial states into a lattice to marginalize over valid interleavings, and introduces coverage regularization to maintain supervision across paths. Experiments show improved recommendation performance on three benchmarks.

Critical strengths

1 The motivation is clear, and the controlled comparison directly shows that generation order affects recommendation performance.

2 Canonical-state sharing enables exact marginalization over a target item's valid orders without enumerating every path.

3 Comparing marginal and max aggregation using the same checkpoint provides useful evidence for aggregating multiple paths during inference.

Critical weaknesses

1 The model defines item scores by marginalizing all valid orders, but beam decoding aggregates only surviving paths. Even retrieved items may therefore lose different amounts of probability mass, potentially changing their ranking. The paper does not quantify this approximation or compare beam rankings with exact marginal rescoring of the same candidates.

2 At λ = 1, LatticeRec and Random-Order optimize the same objective, but one computes supervision over all paths while the other samples one path per instance. Without matching training compute, it remains unclear how much improvement comes from the additional training work.

3 Concentrating probability on a few paths may reflect useful order preferences, and shared states can still receive supervision through other paths. I am not convinced that this should necessarily be treated as harmful collapse. The results show that coverage regularization helps, but do not establish whether it corrects collapse or simply provides a useful regularization effect.

Questions for authors

1 Does the improvement over Best-Fixed hold across multiple seeds and the other two datasets?

2 How does Random-Order perform with more sampled paths or additional updates under the same training budget?

3 What evidence distinguishes harmful path collapse from useful order preferences, particularly when paths share predictive states?

4 comments

We sincerely appreciate the reviewer for the constructive comments and the supportive assessment of our paper. We hope the responses below address your points effectively.

W1: Beam Decoding Approximation

To assess the impact of beam pruning on item rankings, we rescore the beam-retrieved candidates using exact marginalization over all valid generation orders on three datasets. By keeping the candidate set fixed, we isolate the effect of beam approximation on ranking.

We report Mass Retention, the ratio of beam-aggregated probability to exact marginal probability (median/P10 across candidates), and Overlap@10, the average proportion of shared items between the two top-10 lists.

DatasetMass Retention (Median/P10)Overlap@10Beam N@10Exact N@10
Instruments68.91% / 33.14%99.11%0.033950.03392
Games73.87% / 35.13%99.39%0.053580.05354
Steam72.53% / 35.98%99.36%0.054390.05438

Although beam pruning discards substantial probability mass, its impact on the final rankings is small. Top-10 overlap exceeds 99% across all three datasets, with absolute NDCG@10 differences of at most 0.00004. These results indicate that, despite the loss of probability mass, beam aggregation produces nearly identical top-10 recommendations to exact marginalization on the same candidate set.

W2 / Q2: Random-Order vs. LatticeRec

To examine whether the performance gain comes from additional training compute, we conducted two controlled experiments on Games. More Updates extends the learning-rate schedule to increase training updates while sampling one path per instance. More Paths samples three paths per instance at each update. All experiments use a single A100 GPU, and training GPU-hours exclude validation time.

MethodTrain GPU-hoursRecall@10NDCG@10
Random-Order (original)1.920.09400.0521
Random-Order (more updates)3.150.09450.0523
Random-Order (more paths)3.590.09530.0525
LatticeRec (λ=1)3.080.09640.0536

Under comparable training budgets, LatticeRec outperforms Random-Order with more updates in NDCG@10 (0.0536 vs. 0.0523). Increasing the number of sampled paths to three improves Random-Order to 0.0525, but still falls short of LatticeRec despite requiring more compute (3.59 GPU-hours). These results suggest that the improvement is not simply due to additional training compute.

Q1: Stability Across Datasets and Random Seeds

We extended the evaluation of all 20 fixed generation orders to Instruments and Steam, and repeated the Games experiments with three random seeds (42, 43, and 44).

On Games, LatticeRec consistently outperformed Best-Fixed across all three seeds, although the best-performing fixed order varied by seed. Importantly, LatticeRec achieved these gains without selecting a generation order in advance.

Games, seed 42 · Games, seed 43 · Games, seed 44

LatticeRec also outperformed Best-Fixed on Instruments and Steam, achieving NDCG@10 scores of 0.0339 vs. 0.0332 and 0.0544 vs. 0.0538, respectively.

Instruments, NDCG@10 · Instruments, Recall@10 Steam, NDCG@10 · Steam, Recall@10

These results confirm that the improvement over Best-Fixed holds across all three datasets and remains consistent across the three tested seeds on Games.

★ W3 / Q3: Path Concentration vs. Order Preferences

We agree that concentrating probability on a few paths is not necessarily harmful; it may reflect useful order preferences. Our concern is premature path lock-in under pure marginal-likelihood training (λ=0), where early high-probability paths dominate supervision, potentially preventing the model from discovering better orders. Even with canonical state sharing, outgoing transitions can receive highly imbalanced supervision. Coverage regularization mitigates this by assigning each valid path a gradient weight of at least λ/K.

We further analyzed path posteriors and recommendation performance on Instruments:

MethodEarly Effective Paths(Epoch 8)Final Effective PathsAll 20 Paths N@10Best Path N@10Edge Posterior Mass (Weak / Other)
w/o Coverage1.481.240.02990.0291<0.0001 / 0.9984
Ours (λ=1)9.9516.550.03380.03280.2672 / 0.2697

Effective Paths measures the effective support of the path posterior. Edge Posterior Mass measures the posterior mass passing through two outgoing edges of the same shared state. The weak edge is identified using the λ=0 model and kept fixed across both models.

Without coverage, the posterior concentrates on nearly a single path early in training, and some outgoing edges receive almost no posterior mass despite state sharing. Coverage improves both path coverage and recommendation performance. To examine whether these early preferences persist, we compared the highest-posterior paths at early epochs with those at the best validation checkpoint.

Early-to-best checkpoint path stability

MethodE5 AgreementE10 AgreementE30 Agreement
w/o Coverage54.90%73.77%92.09%
Ours (λ=1)5.26%7.12%25.14%

Agreement is the proportion of instances whose highest-posterior path at a given epoch matches that at the best checkpoint.

The key finding is that early path concentration under λ=0 persists throughout training. By E10, 73.77% of instances already favor the same path as at the best checkpoint, rising to 92.09% by E30, compared with only 25.14% under coverage-aware training. Together with the effective path count approaching one, this suggests that early preferences become locked in, leaving subsequent training with limited ability to reconsider alternative orders. With coverage, more paths remain sufficiently supervised, allowing preferences to evolve further during training.

However, early concentration alone does not establish whether these preferences are harmful. We therefore tested whether coverage helps identify more effective generation orders. We fixed the candidate set and view-selection probabilities from the λ=0 model. The last two settings use identical λ=1 code scores, differing only in whether the generation path is reselected.

Code ScorerPath SelectionNDCG@10
λ=0Original best path0.0291
λ=1Original best path0.0296
λ=1Reselected best path0.0326

Keeping the original paths improves NDCG@10 only from 0.0291 to 0.0296, whereas reselecting them further improves it to 0.0326. Path reselection accounts for approximately 86% of the total gain in this comparison, indicating that the original path preferences are not necessarily the most effective.

The distinction is therefore not whether path probabilities are concentrated, but whether the preferred paths lead to better predictions. Our results suggest that pure marginal-likelihood training can lock in early path choices before alternatives are adequately explored. Coverage helps mitigate this problem by maintaining broader path supervision, allowing the model to discover more effective generation orders rather than merely spreading probability mass across paths.

Thank you for the detailed response and clarifications. I will keep my original rating, good luck.

Thank you very much for your positive assessment and constructive feedback. We sincerely appreciate your time and thoughtful review.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.