acceptodds
Under review as a conference paper at ICLR 2027

History-Conditioned Joint-Prefix Alignment for Generative Recommendation

Abstract

Generative recommendation retrieves items by autoregressively generating semantic identifiers, but beam search may discard a target before its complete identifier is generated. Our preliminary analysis across three benchmarks shows that most missed targets are pruned within the first two decoding steps, highlighting the importance of early prefix retention. However, retaining a target's first-token branch alone is insufficient if its continuation is pruned at the next step: aligning only the first-token distribution improves first-token survival but yields little gain in complete-path retention. This observation motivates joint supervision of early branches and their continuations. We propose Prefix Alignment with Temporal History (PATH), which aggregates transition statistics from the training corpus over recent interactions with exponential decay to construct history-conditioned two-token prefix targets. PATH aligns the model's joint predictions with these targets through forward KL divergence, using a chain-rule decomposition that enables unbiased Monte Carlo estimation. At inference time, PATH reuses the transition statistics for pointwise mutual information (PMI) calibration, reranking completed candidates relative to global prefix frequency. Experiments on Beauty, Instruments, and Yelp demonstrate the effectiveness of PATH, showing consistent improvements in recommendation performance and higher full-SID survival rates.

Primary area:
generative models
Keywords:
Generative Recommendation, Semantic Identifiers, Joint-Prefix Alignment, History-Conditioned Supervision, Beam Search

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ai3TOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose "Prefix Alignment with Temporal History" (PATH), which addresses the issue where generative recommendation models—when using beam search—might prematurely lose the target by failing to preserve early prefixes. PATH improves recommendation accuracy by aligning model predictions with historical data. Experiments demonstrate that PATH enhances both recommendation performance and path integrity across multiple datasets.

Critical strengths
  1. The proposed PATH significantly improves the retention of target paths during the recommendation process.

  2. PATH shows higher recommendation accuracy compared to existing methods across multiple datasets.

Critical weaknesses
  1. The PATH method introduces additional statistical computations, increasing the model's complexity.

  2. PATH requires substantial computational resources for achieving higher prediction accuracy.

  3. Its generalizability and adaptability to large-scale industrial datasets remain unverified.

Questions for authors
  1. How can the performance of PATH be optimized without significantly increasing computational resources?

  2. Can PATH adapt to high-frequency dynamic changes in real-time recommendation systems? How does the method handle real-time data streams?

  3. How does PATH perform in multi-step prediction scenarios? Can it effectively handle cross-temporal predictions of complex user behavior sequences?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ai3T Feedback

We sincerely thank you for the constructive comments and insightful questions. Your feedback has been valuable in helping us strengthen the rigor and clarity of our manuscript.

We appreciate your concerns regarding the computational overhead, practical scalability, and adaptability of PATH to dynamic recommendation scenarios. In response, we have quantified the additional training and memory costs, analyzed the storage requirements of the transition matrix, and clarified the computational efficiency of Monte Carlo sampling. We have also further discussed how PATH incorporates evolving user interaction histories, its potential extension to streaming environments, and the distinction between joint-prefix prediction and multi-step recommendation.

Our detailed responses are provided below.

AuthorsOpenReview ↗

Response to Question 1 [Part 1]

We thank the reviewer for this insightful question. We agree that balancing recommendation accuracy and computational efficiency is important for the practical applicability of PATH. In response, we have conducted additional experiments to quantify the computational overhead and evaluate the effectiveness of Monte Carlo sampling. We further clarify how the proposed design controls additional computation while maintaining recommendation performance, as detailed below.

PATH controls its additional computational cost through the chain-rule decomposition of the joint KL objective. Specifically, the first-level KL term is computed exactly, while the second-level conditional KL term is estimated using a small number of branches sampled from the teacher's first-level distribution. Since the teacher distribution is fixed with respect to the model parameters, this provides an unbiased Monte Carlo estimator of the conditional objective and its gradient, avoiding exhaustive evaluation of all supported first-token branches.

On Beauty, the SID mapping contains 197 supported first-level branches, whereas PATH evaluates only four sampled branches per instance (Kmc=4K_{\mathrm{mc}}=4) in the supplementary experiments. Table 1 compares the recommendation performance of NTP, PATH, and the strict two-marginal alternative.

Table 1: Recommendation performance and auxiliary branch evaluations on Beauty.

MethodBranchesNDCG@10Hit@10
NTP00.0354130.065734
PATH (sampled joint KL)40.0469760.083397
Strict two-marginal1970.0458320.079775

PATH improves NDCG@10 and Hit@10 over NTP by 32.7% and 26.9%, respectively. It also achieves higher recommendation accuracy than the strict two-marginal alternative in this comparison, despite evaluating substantially fewer branches. These results show that the sampled joint-prefix objective achieves favorable recommendation performance with limited additional conditional branch evaluations.

We further measure the actual training overhead on a single NVIDIA RTX 3090. The measurements use FP32 precision and include the forward pass, backward pass, and optimizer update. Table 2 reports the average training-step time and peak CUDA allocated memory.

Table 2: Training cost comparison on a single NVIDIA RTX 3090.

BatchMethodBranchesStep (ms)Memory (GiB)
128NTP034.341.553
128PATH (sampled)446.822.394
32NTP019.970.471
32PATH (sampled)441.690.859
32Strict two-marginal197494.728.444

At batch size 128, PATH increases training-step time by 36.3% and peak GPU memory by 54.1% relative to NTP. At batch size 32, the strict two-marginal alternative requires approximately 11.9 times the training-step time and 9.8 times the peak allocated memory of sampled PATH.

The computational efficiency of PATH comes from estimating the conditional KL term through Monte Carlo sampling. Specifically, the additional conditional branch-evaluation cost is reduced from O(MD)O(MD) for exhaustive evaluation to O(KmcD)O(K_{\mathrm{mc}}D), where MM denotes the number of supported first-level branches and DD is the cost of one conditional branch evaluation. Both the strict two-marginal alternative and unsampled PATH require second-token predictions under all first-level SID branches, and therefore incur the full branch-enumeration cost. Sampling controls this additional decoding work by evaluating only a small number of branches. These results demonstrate the computational advantage of stochastic branch evaluation over exhaustive enumeration.

AuthorsOpenReview ↗

Response to Question 1 [Part 2]

Regarding memory requirements, the transition matrix is indexed by valid two-token SID prefixes rather than individual items. With a fixed 256-entry codebook at each of the first two SID levels, the number of distinct prefixes satisfies

$ P \leq \min(|\mathcal{V}|,256^2) \leq 65,536 $$

Although dense storage requires O(P2)O(P^2) memory, the prefix space has a fixed upper bound under this configuration. The actual FP32 matrix sizes are 239.95 MiB, 171.34 MiB, and 961.85 MiB on Beauty, Instruments, and Yelp, respectively. The theoretical maximum is 16 GiB per FP32 copy, which would be reduced to 8 GiB using BF16 storage. The current implementation uses FP32; the BF16 figures represent potential storage savings rather than experimentally validated configurations. Finally, PMI calibration is applied only after candidate generation. It requires O(BH)O(BH) transition-statistic lookups and O(Blog⁡B)O(B\log B) candidate sorting, where BB is the beam width and HH is the history length, without additional autoregressive decoding.

Overall, PATH introduces measurable but controllable computational overhead. The Monte Carlo estimator limits the additional conditional decoding cost while preserving effective joint-prefix supervision, and the bounded SID prefix space makes the transition-storage requirements explicit. Together, these design choices provide a practical trade-off between recommendation accuracy and computational cost.

Reviewer_E6wzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PATH, which constructs history-conditioned soft targets from recency-weighted transitions between two-token semantic-ID prefixes. It aligns the model's joint prefix predictions through a forward-KL objective and applies PMI-based calibration to rerank generated candidates. Experiments on Beauty, Instruments, and Yelp show improvements in prefix survival and recommendation performance.

Critical strengths

The paper identifies a concrete limitation of first-token alignment and motivates joint supervision of early prefixes and their continuations. The history-shuffled and global-marginal controls provide useful evidence that matching the auxiliary targets to interaction histories contributes to the observed gains.

Critical weaknesses

The methodological contribution is somewhat incremental, combining corpus-level transition statistics with a standard forward-KL objective and PMI-based reranking. The experiments do not fully establish the necessity of this particular formulation over simpler alternatives, and the practical scalability remains unclear given the (O(P^2)) storage cost of the transition table and the absence of measured computational overhead or larger-catalog evaluations.

The empirical evidence is also insufficiently robust. All experiments use a single random seed, with no variance estimates or significance tests. Moreover, some baseline results are taken from prior reports, while BEAR uses a Llama-3.2-3B backbone and PATH uses the T5-based LETTER+TIGER framework. These differences make it difficult to attribute the performance gaps solely to the proposed method.

Questions for authors

Could the authors report results across multiple random seeds and provide a comparison with recent baselines under matched backbone and training settings? Also, can the memory and computational overhead of PATH be quantified as the catalog size grows?

6 comments
AuthorsOpenReview ↗

Response to Reviewer E6wz Feedback

We sincerely thank you for the constructive comments and insightful questions.

We appreciate your concerns regarding the methodological contributions, empirical robustness, baseline comparability, and practical scalability of PATH. In response, we conducted additional experiments on Beauty to compare joint alignment with separate marginal-alignment alternatives, evaluated NTP and PATH across seven random seeds on three datasets, and further clarified the differences in backbone architectures and training configurations when comparing with existing baselines. We also conducted a detailed analysis of training-time overhead, GPU memory consumption, and transition-table storage requirements to better characterize the computational efficiency and scalability of PATH.

Our responses are provided below.

AuthorsOpenReview ↗

Response to Weakness 1: Limited Methodological Novelty and Insufficient Validation of the Proposed Formulation [Part 1]

We thank the reviewer for raising this important concern regarding the methodological novelty and the necessity of our design. We appreciate the opportunity to further clarify the motivation and contributions of PATH. The contribution of PATH lies in how these techniques are formulated and integrated to address two closely related challenges in SID-based generative recommendation: preserving plausible early decoding paths and accurately ranking the resulting candidate items based on history-specific transition evidence.

(1) Necessity of history-conditioned joint-prefix alignment.

Our motivation comes from the observation that target-item loss is concentrated in the early decoding stages. Improving the first-token prediction alone does not guarantee the survival of the complete SID, since its continuation may still be pruned at the next decoding step.

To address this issue, PATH constructs history-conditioned transition distributions over two-token SID prefixes by aggregating transition statistics from multiple recent interactions with exponential temporal decay. Unlike global marginal supervision, this formulation preserves the dependency between first-token branches and their corresponding second-token continuations.

We further employ joint forward KL to align the model's autoregressive predictions with these targets. Through chain-rule decomposition, the first-token marginal term is computed exactly, while the conditional second-token term is estimated through Monte Carlo sampling. This provides an unbiased estimator of the joint objective and its gradient without exhaustively evaluating all first-token branches.

To clarify the necessity of these design choices, we summarize the comparisons with alternative formulations, including results from the original submission and additional results from the revised experiments, in Table 1.

Table 1: Summary of alternative formulations and supporting evidence for PATH.

Design aspectAlternative formulationKey evidence
Prefix supervisionNTP / first-token-only alignmentFull-SID survival (%) on Beauty / Instruments / Yelp: NTP: 10.07 / 13.84 / 6.69; first-token-only: 10.10 / 13.80 / 6.69; PATH: 11.75 / 14.37 / 7.67.
History conditioningGlobal marginal / history-shuffled targetsBeauty: full-SID survival (%) / NDCG@10: global marginal: 9.44 / 0.0368; history-shuffled: 9.39 / 0.0360; history-aligned PATH: 11.75 / 0.0477.
Joint dependencyIndependent matching of two token marginalsBeauty: NDCG@10 / Hit@10: branch-wise marginal: 0.042592 / 0.074945; exact two-marginal: 0.045832 / 0.079775; PATH (joint KL): 0.046976 / 0.083397.
Inference calibrationNo calibration / conditional priorPMI achieves the highest NDCG@10 and Hit@10 across all three datasets. Its additional gains over the conditional prior are relatively small but consistent.

These comparisons highlight the distinct roles of the three components in PATH: history conditioning provides behavior-specific supervision, joint KL preserves branch-continuation dependencies, and PMI calibration further refines candidate ranking. The results demonstrate that the improvements do not simply arise from introducing an additional regularization objective or corpus-level transition statistics. In particular, the global-marginal and history-shuffled controls establish the importance of aligning transition targets with the corresponding interaction history. Furthermore, although independently matching the two token marginals provides substantial improvements, it does not explicitly constrain the conditional continuation distribution associated with each first-token branch. In contrast, joint KL preserves these branch-continuation relationships, providing a more faithful supervision signal for complete early decoding paths rather than isolated token distributions.

AuthorsOpenReview ↗

Response to Weakness 1: Limited Methodological Novelty and Insufficient Validation of the Proposed Formulation [Part 2]

(2) Necessity and complementary role of PMI calibration.

PMI calibration addresses a different issue in candidate ranking: the influence of global prefix frequency.

The history-conditioned transition probability q~12(b∣H)\widetilde q_{12}(b\mid H) measures the likelihood of a destination prefix under the current interaction history. However, a high probability may also result from the prefix being globally frequent in the training corpus. Consequently, frequently occurring prefixes may receive high ranking scores simply due to their global popularity, even when their associations with the current interaction history are relatively weak.

PATH addresses this issue by measuring the conditional transition probability relative to its global reference:

sPMI(b,H)=log⁡q~12(b∣H)Pglobal(b)s_{\mathrm{PMI}}(b,H)=\log\frac{\widetilde q_{12}(b\mid H)}{P_{\mathrm{global}}(b)}

Here, q~12\widetilde q_{12} is obtained by aggregating support-dependent smoothed transition distributions. The shrinkage mechanism improves the reliability of estimates for sparsely observed source prefixes, while the PMI ratio normalizes destination-prefix associations against their global frequency. Consequently, a positive PMI signal indicates that a prefix is more strongly associated with the current history than its global prevalence would suggest.

Importantly, PMI and joint-prefix alignment serve different purposes. Joint alignment improves candidate-path retention during autoregressive generation, whereas PMI improves the ranking of completed candidates by exploiting history-specific associations beyond global prefix popularity. The final score is

Score(v)=mv+βclip⁡(sPMI(π12(v),H),−κ,κ)\mathrm{Score}(v)=m_v+\beta\operatorname{clip}\left(s_{\mathrm{PMI}}(\pi_{12}(v),H),-\kappa,\kappa\right)

where mvm_v is the original sequence log-probability. This calibration reuses the transition statistics without additional autoregressive decoding and does not change the generated candidate set.

The experimental results further support its effectiveness. Compared with uncalibrated PATH and the conditional-prior variant, PMI achieves the following NDCG@10 results:

DatasetNo calibrationConditional priorPMI
Beauty0.04770.04820.0484
Instruments0.08740.08750.0881
Yelp0.02860.02920.0293

PMI also reduces the average rank of surviving targets from 7.21 to 7.11 on Beauty, 5.13 to 5.02 on Instruments, and 7.82 to 7.61 on Yelp. These results demonstrate that PMI consistently improves candidate discrimination across different datasets.

Overall, although the improvement in metrics provided by PMI is modest, its contribution mechanism is essential in the context of this paper, as it serves to complement joint alignment. While joint alignment preserves history-conditioned dependencies between successive SID positions during candidate generation, PMI mitigates the influence of global prefix popularity during candidate ranking. By combining the two, PATH is able to leverage consistent transition statistics across both the candidate generation and ranking phases.

We have revised the manuscript and included additional experimental results to clarify these complementary contributions.

Reviewer_DngUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PATH uses training-corpus transitions and recent history to build soft targets for the first two SID tokens. It adds a sampled joint-KL loss to next-token prediction, then uses PMI to rerank completed candidates. The experiments track recommendation quality and prefix survival on three datasets, with history-matching controls and calibration ablations.

Critical strengths
  1. Separating first-token survival from full-path survival is a useful diagnostic. The first-token-only result shows why a better first step needn't give better candidate retrieval.
  2. The global and shuffled controls, including their weight sweeps, make the case for history-matched supervision more convincing. Most gains remain without reranking.
Critical weaknesses
  1. I don't think the controls isolate the benefit of joint dependence. First-token-only alignment removes the second-position auxiliary loss altogether; it doesn't compare joint alignment with supervising both token marginals. The global and shuffled targets test history matching, not this distinction. The results support a history-conditioned soft teacher, but leave unclear what the joint construction adds beyond another supervised position.
  2. Everything is run with seed 42, so there's no estimate of training variability. PMI's NDCG@10 advantage over the conditional prior is only 0.0001-0.0006. The larger improvements over NTP are promising, but these smaller component gains aren't well supported without repeated runs.
Questions for authors
  1. On Beauty, would matching both history-conditioned token marginals seperately give similar gains to joint KL?
  2. Could you repeat NTP and PATH on Beauty with a few additional seeds, then apply the calibration variants to each PATH checkpoint?
  3. What are the training-time and peak-memory costs relative to NTP, including transition storage and sampled branch evaluations?
4 comments
AuthorsOpenReview ↗

Response to Reviewer DrgU Feedback

We sincerely thank you for the constructive comments and insightful questions.

We appreciate the your recognition of the importance of distinguishing first-token survival from full-SID survival, as well as the value of history-conditioned supervision. In response to the concerns, we conducted additional experiments on Beauty to investigate the contribution of joint dependence, evaluated NTP and PATH across seven random seeds, and further analyzed the computational overhead and transition-storage requirements.

Our responses are provided below.

AuthorsOpenReview ↗

Response to Question 1

Thank you for pointing out this important distinction. We agree that our original first-token-only ablation removes the second-level auxiliary supervision and therefore cannot isolate the contribution of joint dependence. To address this concern, we conducted additional experiments on Beauty with two marginal-alignment controls.

Specifically, the two history-conditioned token marginals are defined as

q1(c1∣H)=∑c2q(c1,c2∣H)q_1(c_1|H)=\sum_{c_2}q(c_1,c_2|H)

q2(c2∣H)=∑c1q(c1,c2∣H)q_2(c_2|H)=\sum_{c_1}q(c_1,c_2|H)

A separate two-marginal alignment objective can be formulated as

Lmarg=DKL(q1∥pθ,1)+DKL(q2∥pˉθ,2)\mathcal{L}_{\mathrm{marg}}=D_{\mathrm{KL}}(q_1\|p_{\theta,1})+D_{\mathrm{KL}}(q_2\|\bar p_{\theta,2})

where the model's second-token marginal is

pˉθ,2(c2∣X)=∑c1pθ,1(c1∣X)pθ,2(c2∣X,c1)\bar p_{\theta,2}(c_2|X)=\sum_{c_1}p_{\theta,1}(c_1|X)p_{\theta,2}(c_2|X,c_1)

This objective aligns the two token marginals independently without explicitly preserving their conditional dependence.

In contrast, PATH uses the chain-rule decomposition of joint KL:

\mathcal{L}_{\mathrm{joint}}=&D_{\mathrm{KL}}(q_1\|p_{\theta,1})+\mathbb{E}_{c_1\sim q_1} \left[D_{\mathrm{KL}}\left(q_2(\cdot|c_1,H)\|p_{\theta,2}(\cdot|X,c_1)\right)\right] \end{aligned}$$ The second term explicitly preserves the correspondence between each first-token branch and its associated second-token continuations. We implemented two additional controls. The first is a branch-wise marginal-target control, which samples four first-token branches and applies the same marginal target $q_2(c_2|H)$ to each sampled branch. The second computes the exact two-marginal objective by enumerating all 197 supported first-token branches to obtain the model's second-token marginal. Table 1 reports the results without inference calibration. All auxiliary training objectives use $\alpha=0.3$. **Table 1: Comparison of joint and separate marginal alignment on Beauty.** | Method | Additional branch evaluations | NDCG@10 | Hit@10 | |:---|:---:|:---:|:---:| | NTP | 0 | 0.035413 | 0.065734 | | Branch-wise marginal target | 4 | 0.042592 | 0.074945 | | Exact two-marginal alignment | 197 | 0.045832 | 0.079775 | | **PATH (joint KL)** | **4** | **0.046976** | **0.083397** | The results show that separate marginal supervision already provides substantial improvements over NTP. In particular, the exact two-marginal objective achieves 0.045832 NDCG@10 and 0.079775 Hit@10. However, PATH further improves these metrics to 0.046976 and 0.083397, respectively, corresponding to absolute improvements of 0.001144 in NDCG@10 and 0.003622 in Hit@10 over the exact two-marginal control. Notably, the separate marginal objective numerically captures approximately 90.1% of PATH's NDCG@10 improvement and 79.5% of its Hit@10 improvement over NTP in this experiment. This indicates that marginal supervision accounts for a substantial portion of the gains, while explicitly preserving the dependence between the two SID positions provides additional benefits. These results clarify the role of joint alignment: its advantage does not simply come from supervising an additional decoding position, but also from retaining the branch-continuation relationships encoded in the history-conditioned joint distribution. We have incorporated these additional controls and clarify this distinction in the revised manuscript.
AuthorsOpenReview ↗

Response to Question 2

Thank you for this suggestion. We conducted additional experiments on Beauty, Instruments, and Yelp using seven random seeds: 42, 2026, 2027, 2028, 2029, 2030, and 2031. For each seed, we independently trained NTP and PATH. We then applied the conditional-prior and PMI calibration variants to the corresponding PATH checkpoint, using the same generated candidate set for all calibration variants.

Table 2 reports the mean and sample standard deviation across the seven single-GPU runs for each dataset. The original four-GPU run with seed 42 on Beauty is retained as a reference but excluded from these statistics.

Table 2: Seven-seed results on Beauty, Instruments, and Yelp (mean ± sample standard deviation).

DatasetMethodNDCG@10Hit@10
BeautyNTP0.0347±0.0003840.0347 \pm 0.0003840.0671±0.0007260.0671 \pm 0.000726
BeautyPATH (no calibration)0.0477±0.0008970.0477 \pm 0.0008970.0844±0.0012990.0844 \pm 0.001299
BeautyPATH + Conditional Prior0.0482±0.0008430.0482 \pm 0.0008430.0850±0.0009080.0850 \pm 0.000908
BeautyPATH + PMI0.0484±0.0008100.0484 \pm 0.0008100.0855±0.0011150.0855 \pm 0.001115
InstrumentsNTP0.0825±0.0005200.0825 \pm 0.0005200.1107±0.0008400.1107 \pm 0.000840
InstrumentsPATH (no calibration)0.0874±0.0004940.0874 \pm 0.0004940.1166±0.0008510.1166 \pm 0.000851
InstrumentsPATH + Conditional Prior0.0875±0.0004150.0875 \pm 0.0004150.1163±0.0004710.1163 \pm 0.000471
InstrumentsPATH + PMI0.0881±0.0003690.0881 \pm 0.0003690.1178±0.0007290.1178 \pm 0.000729
YelpNTP0.0228±0.0006500.0228 \pm 0.0006500.0429±0.0009900.0429 \pm 0.000990
YelpPATH (no calibration)0.0286±0.0006920.0286 \pm 0.0006920.0514±0.0010330.0514 \pm 0.001033
YelpPATH + Conditional Prior0.0292±0.0006380.0292 \pm 0.0006380.0524±0.0010240.0524 \pm 0.001024
YelpPATH + PMI0.0293±0.0006310.0293 \pm 0.0006310.0527±0.0010800.0527 \pm 0.001080

The results demonstrate consistent improvements from PATH over NTP across all three datasets. On average, PATH improves NDCG@10 by 37.5% and Hit@10 by 25.8% on Beauty, 5.9% and 5.3% on Instruments, and 25.4% and 19.8% on Yelp, respectively. The relatively small standard deviations further indicate that the improvements are stable across different random initializations.

Regarding inference calibration, both calibration variants generally improve over uncalibrated PATH in the aggregate results, although the conditional prior slightly reduces Hit@10 on Instruments. However, we acknowledge that the additional advantage of PMI over the conditional prior is relatively small. The mean differences between PMI and the conditional prior are 0.0002 in NDCG@10 and 0.0005 in Hit@10 on Beauty, 0.0006 and 0.0015 on Instruments, and 0.0001 and 0.0003 on Yelp, respectively.

Therefore, we do not interpret these results as establishing a robust superiority of PMI over the conditional prior. Rather, they suggest that both calibration strategies provide comparable supplementary ranking improvements, while PATH's main contribution comes from joint-prefix alignment during training.

In the revised manuscript, we have included the multi-seed results across all three datasets and moderated the discussion of the incremental advantage of PMI accordingly.

AuthorsOpenReview ↗

Response to Question 3

Thank you for raising the computational-efficiency question. We provide additional details on training-step time, peak GPU memory, transition-table storage, and the number of conditional branch evaluations.

(1) Training time and peak GPU memory.

We profiled NTP and PATH on a single RTX 3090 using FP32 precision. Each training step includes the forward pass, backward pass, and optimizer update. The results are summarized in Table 3.

Table 3: Training-step time and peak GPU memory.

Batch sizeMethodAdditional branchesStep time (ms)Peak memory (GiB)
128NTP034.341.553
128PATH446.822.394
32NTP019.970.471
32PATH441.690.859
32Exact two-marginal197494.728.444

At batch size 128, PATH increases step time by 36.3% and peak allocated memory by 54.1% relative to NTP. At batch size 32, the corresponding increases are 108.8% and 82.4%, respectively. The exact two-marginal control is substantially more expensive because it requires evaluating all 197 supported first-token branches. At batch size 32, its training-step time and peak memory are approximately 11.87 and 9.83 times those of sampled PATH, respectively. We note that the PATH profiling above uses Kmc=4K_{\mathrm{mc}}=4, whereas the main experiments use Kmc=2K_{\mathrm{mc}}=2. Thus, the reported profiling results use a larger conditional-branch sampling budget than the main experimental configuration.

(2) Transition-table storage.

Let PP denote the number of valid two-token SID prefixes. The dense transition matrix requires O(P2)O(P^2) storage, with 4P24P^2 bytes under FP32 precision. Table 4 reports the actual single-copy storage requirements.

Table 4: Dense transition-table storage.

Dataset / upper boundPrefixes PPFP32 storage
Beauty7,931239.95 MiB
Instruments6,702171.34 MiB
Yelp15,879961.85 MiB
Theoretical upper bound65,53616 GiB

With a fixed codebook size of 256 at each of the first two SID levels, the maximum number of joint prefixes is $ P_{\max}=256^2=65,536.$ Consequently, a dense transition matrix contains at most Pmax⁡2=232P_{\max}^2=2^{32} entries. Its theoretical storage requirements under different numerical precisions are shown in Table 5.

Table 5: Theoretical maximum transition-table storage under different numerical precisions.

Data typeBytes per entryMaximum storage
FP64832 GiB
FP32416 GiB
BF16 / FP1628 GiB
INT814 GiB

The current implementation uses FP32. The lower-precision values are storage estimates rather than experimentally validated alternatives. The reported dataset-specific storage corresponds to a single matrix copy. In our data-parallel implementation, one CPU copy and one GPU copy per rank are maintained, so the total storage requirement also depends on the number of training ranks.

(3) Sampled conditional branch evaluations.

The main computational advantage of PATH comes from the chain-rule decomposition of joint KL. The first-token marginal KL is computed exactly using the standard forward pass, while the conditional term is estimated by sampling KmcK_{\mathrm{mc}} branches from the history-conditioned first-token distribution.

Let MM denote the number of supported first-token branches and DD the cost of one conditional branch evaluation. Exact enumeration requires O(MD)O(MD) additional conditional decoding work, whereas PATH requires only O(KmcD)O(K_{\mathrm{mc}}D). For Beauty, M=197M=197. Therefore, using four sampled branches reduces the number of conditional evaluations from 197 to 4, a reduction of approximately 98%. The main experimental setting uses only two samples.

Overall, PATH introduces additional training computation and transition-table storage, but the conditional decoding overhead is controlled by a small sampling budget rather than exhaustive branch enumeration. We have clarified these computational costs and the distinction between the profiling and main-experiment configurations in the revised manuscript.

We again thank you for these valuable suggestions, which have helped us strengthen the experimental analysis and clarify the contributions and computational characteristics of PATH.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Manuscript Revision Update

Dear Reviewers,

Thank you for your constructive and valuable comments. We have uploaded a revised version of the manuscript, incorporating the additional experiments and analyses discussed in our responses.

In response to your comments, we conducted further experiments to investigate the identified concerns more thoroughly and made corresponding revisions to the manuscript. The major revisions are concentrated in the Experimental Section and the Appendix, with all modified contents highlighted in blue for ease of identification. In particular, we have focused on addressing concerns regarding experimental fairness, computational overhead, and the benefits of modeling joint dependencies. These revisions are supported by additional experimental evidence and more detailed analyses.

We sincerely appreciate the time and effort you have devoted to reviewing our work. We hope that the revisions and additional evidence adequately address your concerns. In light of these clarifications and improvements, we would be grateful if you could reconsider your overall assessment of the manuscript.

Thank you again for your valuable feedback and consideration.

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.