acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Conditional Transition Memory for Generative Recommendation

Abstract

Generative recommendation (GR) with Semantic-ID (SID) represents each item as a short sequence of semantic tokens. These tokens introduce dual-role dependencies: fine-grained token transitions capture reusable structures shared across instances, whereas next-item selection demands dynamic contextual inference over user history. Standard dense Transformers entangle both roles within uniform computations, lacking an explicit retrieval pathway for recurrent structural patterns. Our empirical investigation reveals that existing GR models fail to fully exploit such reusable transitions, hindering their representational capacity. To address this, we propose Hierarchical Conditional Transition Memory (HCTM), which decouples generation into retrieval-augmented transition lookup and context-dependent preference inference. HCTM integrates an encoder-side Item-Internal Token Memory with a decoder-side Prefix-Conditioned Transition Memory, fusing retrieved memories into hidden states via context-conditioned residual gating. Extensive experiments across three benchmark datasets demonstrate that HCTM consistently outperforms competitive baselines. In-depth analyses further confirm that the performance gains stem directly from selective transition retrieval, while the memory substantially accelerates training convergence. The anonymized code repository is available at https://anonymous.4open.science/r/HCTM-C3B5/.

Primary area:
generative models
Keywords:
Generative Recommendation, Sequential Recommendation, Memory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_BRcQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper considers generative recommendation. Motivated by the commonly used memory modules in recent deep learning literature (especially sequence modeling), the authors proposed a new model architecture called Hierarchical Conditional Transition Memory (HCTM), which was designed to enhance the reusable structure to achieve performance gains. They first carefully examined the limitations of existing methods including the dense transformers only and pattern lookups only to show they are not sufficient, then they designed two types of memory banks including the item-internal one and the prefix-conditioned one. Experiments show performance gains over existing methods like TIGER and OneRec on some public datasets.

Critical strengths
  1. The motivation is clear and easy to follow -- memory modules have been proven effective in sequence modeling in recent years, and it is natural to ask if SID for generative recommendation can also benefit from this type of model architectures.

  2. The authors present some systematic study (Section 4) to show the necessity of using memory banks to avoid redundancy from structural reuse, and they investigate each building block of existing architectural designs including dense transformers alone and pattern lookup alone to showcase that each individual part are insufficient.

  3. Code is provided. Experiments show performance gains over existing methods.

Critical weaknesses
  1. Memory modules are not new in literature -- they have been introduced in some LLMs from frontier labs such as Meta [1] and DeepSeek [2].

  2. Certain notations or figures might be confusing. For example in Figure 5, different items should in general have different codes and deduplication suffixes, and thus it might be better to have some subscripts or superscripts to indicate the codes/suffixes are for item j (1<= j <= i). Also there seems no reference to Figure 5, so might be difficult for readers to find definitions for the superscripts written in the "Prefix-Conditioned Transition Memory" box.

  3. Experiments are on some public datasets and might not be applicable to industry settings yet.

References

[1] Memory Layers at Scale

[2] Conditional memory via scalable lookup: A new axis of sparsity for large language models.

Questions for authors
  1. Since the newly proposed memory banks could achieve structural reuse, could this also bring efficiency gains as we could avoid the redundancy?

  2. There is a DeepSeek-Engram baseline added to the memory augmented baselines. However there seems no too many details disclosed to show how it is implemented in this experiment? Also there are some other memory augmented baseline choices such as memory layers. Wondering if the authors can provide some insights on this.

  3. Does the word "Hierarchical" refer to the RQ-VAE in SID? If I understand correctly there is no "Hierarchical" structure in the newly introduced memory banks?

  4. Are there any results on industrial-scale datasets?

2 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

We sincerely thank Reviewer BRcQ for the constructive comments and questions regarding prior memory architectures, implementation details, efficiency, and terminology. Following your suggestions, we have incorporated Memory Layers directly into our main benchmark comparisons, added full baseline formulations, and clarified our architectural definitions.

Relation to prior memory modules and baseline details (W1, Q2)

To clarify HCTM's positioning relative to generic memory modules and address the implementation details (Q2), we highlight our recommendation-specific structural specialization, provide the exact formulations added to the appendix for both memory baselines, and report their controlled comparison on TIGER:

  • Recommendation-specific specialization: Unlike Memory Layers (trainable hidden-state KV lookup) and DeepSeek-Engram (local flat nn-gram reuse), HCTM aligns memory with hierarchical SID structure—indexing intra-item token compositions and inter-item successor prefixes as explicit predictive evidence for next-item generation.
  • DeepSeek-Engram baseline: Implemented as hash-based nn-gram retrieval (n∈{2,3,4}n \in \{2,3,4\}) over flat Semantic-Token sequences in the TIGER encoder, injected via gated residual after Layer 0 (added to Appendix D.3). For each order nn and hash head qq, hn,q(zt−n+1:t)h_{n,q}(z_{t-n+1:t}) addresses a separate table returning en,q,t\mathbf{e}_{n,q,t}:

mt=Avg⁡(e(n,q,t))\mathbf{m}_t = \operatorname{Avg}(\mathbf{e}^{(n,q,t)})

st=⟨RMSNorm⁡(ht),RMSNorm⁡(Wkmt)⟩ds_t = \frac{\langle\operatorname{RMSNorm}(\mathbf{h}_t),\operatorname{RMSNorm}(W_k\mathbf{m}_t)\rangle}{\sqrt{d}}

αt=σ(sign⁡(st)∣st∣)\alpha_t = \sigma(\operatorname{sign}(s_t)\sqrt{|s_t|})

ht←ht+αtmt\mathbf{h}_t \leftarrow \mathbf{h}_t+\alpha_t\mathbf{m}_t

  • Memory Layers baseline (parameter controlled comparison): To evaluate continuous parametric memory under a strictly parameter budget, we replace the Layer-0 FFN sub-layer in both TIGER stacks with a sparse key-value memory of 1,0241,024 trainable 128128-d keys and values (matching the exact parameter count of the original dff=1024d_{\text{ff}}=1024 FFN layer; initialized with σk=128−1/2,σv=0.02\sigma_k=128^{-1/2}, \sigma_v=0.02). Given this compact, parameter-matched memory pool (1,0241,024 slots), we compute exact top-4 dot-product retrieval directly rather than using product-key approximation (added to Appendix D.4):

It=TopK⁡j(ht⊤kj,4)\mathcal{I}_t=\operatorname{TopK}_j(\mathbf{h}_t^\top\mathbf{k}_j,4)

α(j)=softmax⁡(ht⊤k(j)),j∈I\alpha^{(j)} = \operatorname{softmax}(\mathbf{h}_t^\top\mathbf{k}^{(j)}), \quad j \in \mathcal{I}

m(t)=∑j∈Iα(j)v(j)\mathbf{m}(t) = \sum_{j\in\mathcal{I}} \alpha^{(j)}\mathbf{v}^{(j)}

The retrieved output replaces the FFN sub-layer within the standard TIGER Transformer block, preserving normalization, dropout, and residual connections without external index construction.

DatasetTIGERDeepSeek-EngramMemory LayersHCTM
Beauty0.03550.03010.03820.0406
Sports0.01970.01640.02050.0220
Toys0.03270.03200.03340.0404

Memory Layers improves upon TIGER across all benchmarks by expanding latent FFN capacity via continuous hidden-state queries, whereas DeepSeek-Engram underperforms TIGER—likely because assuming flat, language-like token adjacency fails to respect the discrete compositional syntax of SID sequences. In contrast, HCTM achieves the highest NDCG@10 across all datasets by explicitly aligning memory retrieval with SID's dual-scale hierarchy—capturing intra-item token compositions and predictive inter-item transitions.

AuthorsOpenReview ↗

Rebuttal (2/2)

Figure 5 notation (W2)

We sincerely apologize for the notation ambiguity in the initial submission and thank the reviewer for catching this. In the revised manuscript, we have updated Figure 5 and its caption to explicitly define the successor-candidate superscripts, which also distinguishes SID code positions from the deduplication suffix.


Industrial applicability (W3, Q4)

While our current evaluation focuses on standard public benchmarks, HCTM possesses two structural properties that offer promising potential for large-scale industrial deployment:

  1. Bounded memory footprint: Indexing quantized Semantic-ID prefixes with a fixed budget (K=8K=8) scales with the hierarchical codebook space rather than the full item catalog, keeping storage overhead manageable as item catalogs grow.
  2. Off-accelerator lookup potential: Because the transition store is non-parametric, it could be hosted in distributed host memory (RAM/NVMe) and queried asynchronously alongside standard serving pipelines, avoiding GPU HBM bottlenecks.

We acknowledge that real-world production deployment involves additional system-level trade-offs (e.g., distributed cache consistency and tail latency), and we leave large-scale industrial validation to future work.


Efficiency of structural reuse (Q1)

Yes, while structural retrieval introduces a modest per-step overhead, it yields substantial overall training efficiency by significantly accelerating optimization convergence:

  • Minimal parameter and per-step overhead: Adding both memory pathways increases trainable parameters by only 4.8% (13.76M vs. 13.13M for TIGER) and reduces per-step training token throughput by 7.8%.
  • Substantial overall speedup (48.2% faster total training): Because explicit memory supplies ready-to-use structural priors that accelerate optimization fitting, HCTM converges 2.1× faster—reaching early stopping in 1,050 optimizer steps compared to 2,200 for TIGER—which yields an 48% reduction in total end-to-end training time.

Meaning of "hierarchical" (Q3)

We use “hierarchical” in a structural-granularity sense: HCTM aligns memory across the two compositional levels inherent to Semantic-ID sequences:

  1. Lower level (intra-item composition): Item memory models how coarse-to-fine Semantic-ID codes combine within individual items to enrich historical representations on the encoder side.
  2. Higher level (inter-item transition): Prefix memory models sequential transitions across consecutive items by retrieving candidate successor prefixes on the decoder side.

We have clarified this distinction explicitly in the revised manuscript to prevent ambiguity.

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

HCTM augments SID-based generative recommendation with item-internal token memory and prefix-conditioned transition memory. Context-conditioned gates fuse retrieved representations into a Transformer backbone. Experiments show improved ranking accuracy and faster convergence, with modest architectural overhead.

Critical strengths
  1. The motivation is supported by informative diagnostics of repeated SID transitions.
  2. The framework integrates structural lookup with contextual modeling through simple gated residuals.
  3. Ablations, frequency-stratified analyses, and efficiency measurements help characterize its benefits and costs.
Critical weaknesses
  1. Comparisons lack simple transition-aware alternatives, making the contribution beyond statistical priors unclear.
  2. Generalization claims exceed the evidence: unseen transitions slightly degrade, while memory-miss analyses use training loss.
  3. Most gains come from decoder transition memory; the necessity of the full hierarchical design is only weakly supported.
  4. Experiments are conducted on Amazon review datasets alone. Lack of evaluation on other domain datasets (e.g., micro-videos, food)
Questions for authors
  1. How does HCTM compare with TIGER interpolated with transition probabilities or a transition-based logit bias?
  2. Does context-dependent candidate weighting improve over the source-specific mixture used here?
  3. How robust is HCTM to temporal distribution shifts, unseen transitions, and different candidate budgets?
  4. Can the authors report free-running ranking results separately for seen and unseen transitions?
3 comments
AuthorsOpenReview ↗

Rebuttal (1/3)

We sincerely thank Reviewer YHzW for the insightful and thoughtful feedback, particularly for highlighting the distinction between structural transition evidence and contextual inference. Your suggestions regarding output-level transition controls, budget sensitivity, and cross-domain evaluation motivated additional experiments, which we have included in the revised manuscript.

Transition-aware TIGER controls (W1, Q1)

To isolate the benefit of HCTM's latent gated fusion over output-level heuristic adjustments, we evaluate both suggested controls on TIGER (incorporating empirical transition priors via static logit bias and probability interpolation, respectively):

DatasetTIGERTIGER + transition logit biasTIGER + probability interpolationHCTM
Beauty0.03550.03860.03940.0406
Sports0.01970.02060.02130.0220
Toys0.03270.03350.03470.0404

While static output-level adjustments yield dataset-dependent gains—performing competitively on Beauty and Sports but poorly on Toys (+0.0020 vs. +0.0077 for HCTM)—HCTM consistently achieves the best results across all datasets. This gap confirms that static score fusion cannot dynamically filter noisy transitions, whereas HCTM's context-conditioned latent gating adapts structural priors to each user's interaction state.

AuthorsOpenReview ↗

Rebuttal (2/3)

Controlled learning and held-out transitions (W2)

We thank the reviewer for catching this overclaim and have revised the text to clarify that Appendix G demonstrates faster training convergence rather than unseen-transition generalization.


Robustness to temporal shifts, unseen transitions, and successor-prefix budgets (Q3, Q4)

Temporal distribution shift

We agree with the reviewer that temporal distribution shift is a critical challenge in real-world recommendation. While our evaluation follows the standard chronological leave-one-out protocol for fair comparison with prior work, HCTM is specifically designed to accommodate temporal evolution via two properties:

  1. Zero-retraining memory refresh: Decoupling transitions from static parametric weights into a non-parametric retrieval store allows incremental sliding-window updates without retraining the backbone.

  2. Semantic stability: Indexing quantized SID prefixes preserves category-level transitions even as atomic items churn over time.

Seen and unseen transitions

To verify this, we partition the held-out test set into seen transitions (Target hit, where the ground-truth prefix appears in retrieved training candidates) and two unseen regimes—where the source prefix retrieves non-target candidates (Target-miss, nonempty lookup) or has no recorded successors (Empty decoder lookup):

DatasetTest groupn (%)TIGER NDCG@10HCTM NDCG@10
BeautyTarget hit2,032 (9.09%)0.23230.2873
BeautyTarget-miss, nonempty lookup20,274 (90.66%)0.01770.0169
BeautyEmpty decoder lookup57 (0.25%)0.02380.0055
SportsTarget hit1,968 (5.53%)0.14250.2115
SportsTarget-miss, nonempty lookup33,551 (94.25%)0.01260.0112
SportsEmpty decoder lookup79 (0.22%)0.01690.0049
ToysTarget hit2,241 (11.54%)0.18320.2618
ToysTarget-miss, nonempty lookup17,092 (88.05%)0.01290.0118
ToysEmpty decoder lookup79 (0.41%)0.01280.0035

These results confirm a highly favorable precision–coverage trade-off: HCTM boosts Target hit NDCG@10 by 23.7%–48.4% (+0.0550 to +0.0786), while incurring a minimal −0.0008-0.0008 to −0.0014-0.0014 difference on the ~90% Target-miss, nonempty lookup cases. This slight margin confirms that context-conditioned gating effectively down-weights misleading historical candidates when they conflict with the user's current context, ensuring net performance gains across the entire test set.

Successor-prefix budget

HCTM is robust to moderate budget variations and achieves an optimal precision–recall balance at K=8K=8. To evaluate sensitivity to the candidate budget, we fix the trained Toys model and vary the number of retained successor prefixes per source in the training transition index:

Successor prefixes per sourceNDCG@5NDCG@10Recall@5Recall@10MRR@5MRR@10
10.02880.03370.04040.05540.02500.0270
20.03190.03720.04570.06240.02740.0295
40.03310.03940.04780.06740.02820.0308
80.03400.04040.04980.06980.02880.0315
160.03350.03990.04880.06870.02840.0309
320.03340.03950.04850.06840.02740.0301

Ranking metrics rise steadily from K=1K=1 to a peak at K=8K=8 (0.0404 NDCG@10) and remain stable across K∈[4,16]K \in [4, 16], whereas overly large budgets (K=32K=32) introduce low-frequency candidate noise that slightly dilutes performance.

AuthorsOpenReview ↗

Rebuttal (3/3)

Relative contribution of the two branches (W3)

To address this concern, we extend the component ablation to all three datasets. The results confirm that both memory components contribute significantly and are indispensable to HCTM's full performance:

DatasetTIGERItem-onlyPrefix-onlyFull HCTMItem-only vs. TIGER pHCTM vs. Prefix-only p
Beauty0.0355±7.226e-40.0374±7.599e-40.0392±7.375e-40.0406±7.673e-40.00370.0187
Sports0.0197±3.704e-40.0203±3.895e-40.0214±3.780e-40.0220±3.933e-40.03720.0394
Toys0.0327±5.074e-40.0335±5.335e-40.0386±9.826e-40.0404±1.022e-30.04130.0219
  • Empirical evidence: While Prefix-only accounts for the primary performance gain, Item-only still significantly outperforms TIGER across all three datasets (p<0.05p < 0.05). Moreover, adding Item memory on top of Prefix-only yields a consistent, statistically significant NDCG@10 improvement of 3.6%, 2.8%, and 4.7% on Beauty, Sports, and Toys, respectively (p=0.0187,0.0394,0.0219p = 0.0187, 0.0394, 0.0219).
  • Conceptual design: The two branches model complementary structural scales—Item memory captures intra-item SID token dependencies, whereas Prefix memory captures inter-item transitions. Combining both aligns retrieval with the full hierarchy of SID generation.

Other domains (W4)

To evaluate cross-domain generalizability beyond e-commerce, we add experiments on two non-shopping benchmarks—MicroLens-100K (micro-video) and MovieLens-1M (movie)—and report test NDCG and Recall below:

MicroLens-100K

MethodNDCG@5NDCG@10Recall@5Recall@10
TIGER0.02270.02810.03390.0485
HCTM0.02510.02920.03580.0503

MovieLens-1M

MethodNDCG@5NDCG@10Recall@5Recall@10
TIGER0.06910.08890.10640.1680
HCTM0.07830.09900.11750.1819

HCTM consistently outperforms TIGER on both MicroLens-100K (+4.1% NDCG@10) and MovieLens-1M (+11.3% NDCG@10), confirming its generalizability across distinct recommendation domains.


Context-dependent candidate weighting (Q2)

We evaluate a context-dependent candidate weighting variant against HCTM's source-specific mixture on test NDCG@10:

DatasetHCTMContext-dependent candidate weighting
Beauty0.04060.0410
Sports0.02200.0216
Toys0.04040.0407

We thank the reviewer for suggesting this intuitive and interesting variant. As shown above, context-dependent candidate weighting yields comparable performance to HCTM (+0.0004 on Beauty, -0.0004 on Sports), indicating that our source-specific mixture weights already capture the primary prior distribution while leaving fine-grained disambiguation to the context-conditioned residual gate. Therefore, we retain the simpler source-specific mixture in the main model for a better efficiency–performance trade-off, and have included this comparison in Appendix.

Reviewer_iQz5OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

HCTM adds two memory pathways to TIGER: within-item token lookup in the encoder and successor-prefix lookup in the decoder. Context-dependent gates mix the retrieved vectors into hidden states. The paper reports ranking results on three Amazon datasets, alongside ablations, transition diagnostics, and convergence and efficiency measurements.

Critical strengths

Ranking gains hold across all three datasets. Switching decoder memory off on the same checkpoint also gives useful evidence that the model uses the retrieved signals, rather than merely benefiting from a different training run.

Critical weaknesses
  1. The convergence experiment doesn't establish capacity offloading or better generalization. Appendix G measures NLL on training subsequences. In its memory-miss group, the correct target isn't among the retrieved candidates, but other candidates and encoder memory can still contribute. Faster fitting on this group doesn't show that the backbone has gained capacity for novel contexts, as Section 8.2 claims. 2.The need for the full hierarchy isn't well established beyond a simpler Prefix-only model. The full model adds about 2.5% NDCG@10 over Prefix-only on Toys, with no repeated-seed ablation or uncertainty estimate. The main table's significance markers are against TIGER, not Prefix-only. There's no architectural ablation on Beauty or Sports to show that the two-pathway conclusion carries over.
Questions for authors
  1. Why is Table 2 limited to Toys? Appendix H gives cross-dataset diagnostics, not model ablations. Do the component and gating findings hold on Beauty or Sports?
  2. Does the capacity-offloading claim hold on held-out memory-miss examples, after separating a missing target from a genuinely empty lookup?
  3. Would context-dependent gating still win against a learned constant gate, rather than only the fixed all-ones control?
  4. There's too much blank space in the appendix, especially on pages 15, 16 and 19. Figure 5's style is inconsistant with the other figures, and it's hard to read without zooming in a lot. Could you tighten the layout and make the figure clearer and visually consistent?
2 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

We sincerely thank Reviewer iQz5 for the thorough and constructive review, as well as for recognizing our ranking improvements. Your questions regarding held-out generalization, multi-dataset statistical significance, and learned constant gating helped us clarify these aspects with additional analyses. We have incorporated new results into the revised manuscript and address each point below.

Convergence and capacity offloading (W1, Q2)

We agree that Appendix G demonstrates faster optimization rather than held-out generalization: it evaluates teacher-forced NLL on 1,953 training subsequences, where "memory-miss" (1,088 cases) denotes target-miss rather than an empty lookup (so other candidates and encoder memory still contribute).

To directly evaluate held-out generalization (Q2), we partition the free-running test set into three mutually exclusive groups:

DatasetTest groupn (%)TIGER NDCG@10HCTM NDCG@10
BeautyTarget hit2,032 (9.09%)0.23230.2873
BeautyTarget-miss, nonempty lookup20,274 (90.66%)0.01770.0169
BeautyEmpty decoder lookup57 (0.25%)0.02380.0055
SportsTarget hit1,968 (5.53%)0.14250.2115
SportsTarget-miss, nonempty lookup33,551 (94.25%)0.01260.0112
SportsEmpty decoder lookup79 (0.22%)0.01690.0049
ToysTarget hit2,241 (11.54%)0.18320.2618
ToysTarget-miss, nonempty lookup17,092 (88.05%)0.01290.0118
ToysEmpty decoder lookup79 (0.41%)0.01280.0035

Key Findings:

  • Targeted gains on supported transitions (Target hit, ~5.5%–11.5%): HCTM improves NDCG@10 by 23.7%–48.4% over TIGER, confirming that gains stem directly from valid historical transitions.

  • Controlled trade-off under distracting candidates (Target-miss, nonempty lookup, ~88%–94%): When retrieved historical successors do not contain the ground-truth item, HCTM exhibits a marginal NDCG@10 difference of −0.0008-0.0008 to −0.0014-0.0014 compared to TIGER. This demonstrates that context-conditioned gating successfully prevents non-target candidates from overwhelming the backbone's contextual reasoning, trading a negligible perturbation on ambiguous transitions for a substantial +0.0550 to +0.0786 NDCG@10 gain on Target hit cases—driving strong overall improvements.

  • Negligible tail risk & zero-cost fallback (Empty decoder lookup, 0.22%–0.41%): The slight drop on this rare tail (57–79 instances) stems from representation shift caused by null-vector substitution. Since empty lookups are deterministically known at inference, bypassing the memory residual directly restores TIGER-level performance at zero computational overhead.

AuthorsOpenReview ↗

Rebuttal (2/2)

Component contributions and full-model necessity (W2, Q1)

To address this concern, we extend the component ablation to all three datasets. The results confirm that both memory components contribute significantly and are indispensable to HCTM's full performance:

DatasetTIGERItem-onlyPrefix-onlyFull HCTMItem-only vs. TIGER pHCTM vs. Prefix-only p
Beauty0.0355±7.226e-40.0374±7.599e-40.0392±7.375e-40.0406±7.673e-40.00370.0187
Sports0.0197±3.704e-40.0203±3.895e-40.0214±3.780e-40.0220±3.933e-40.03720.0394
Toys0.0327±5.074e-40.0335±5.335e-40.0386±9.826e-40.0404±1.022e-30.04130.0219
  • Empirical evidence: While Prefix-only accounts for the primary performance gain, Item-only still significantly outperforms TIGER across all three datasets (p<0.05p < 0.05). Moreover, adding Item memory on top of Prefix-only yields a consistent, statistically significant NDCG@10 improvement of 3.6%, 2.8%, and 4.7% on Beauty, Sports, and Toys, respectively (p=0.0187,0.0394,0.0219p = 0.0187, 0.0394, 0.0219).
  • Conceptual design: The two branches model complementary structural scales—Item memory captures intra-item SID token dependencies, whereas Prefix memory captures inter-item transitions. Combining both aligns retrieval with the full hierarchy of SID generation.

Gating ablation (Q1)

To evaluate whether the gating mechanism generalizes beyond Toys, we compare context-dependent gating against the fixed unit-gate control across all datasets:

DatasetFixed unit gateContext-dependent gate
Beauty0.03950.0406
Sports0.02130.0220
Toys0.03890.0404

Context-dependent gating consistently outperforms the fixed unit-gate control across all datasets.


Learned constant gate (Q3)

Following the reviewer's suggestion, we compare our context-dependent gate against a learned context-independent (constant) gate on test NDCG@10:

DatasetLearned constant gateContext-dependent gate
Beauty0.03980.0406
Sports0.02150.0220
Toys0.03950.0404

Context-dependent gating consistently outperforms the learned constant gate across all datasets, confirming the benefit of context-specific memory modulation.


Figure and appendix presentation (Q4)

We thank the reviewer for the helpful presentation feedback. In the revised manuscript, we have made the following concrete updates:

  • Figure 5: Added an explicit legend distinguishing quantized semantic codes (ci,1:3c_{i,1:3}) from the non-quantized deduplication suffix (dd), highlighted the causal boundary between observed history and partially generated targets, and clarified in the caption that superscripts index successor candidates while subscripts denote SID positions.
  • Appendix: Reorganized Sections C and D to cleanly separate baseline formulations (Appendices D.3–D.4) from HCTM's causal input and training specifications (Appendices D.5–D.6), and tightened table/equation spacing to eliminate excessive whitespace.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.