acceptodds
Under review as a conference paper at ICLR 2027

A Generative Recommender Also Needs a Discrete Language for Memory

Abstract

Generative recommendation provides a discrete language for items, typically through Semantic IDs (SIDs), but leaves user memory implicit in the interaction sequence. Even when the complete history is available, the model must repeatedly infer from raw events the predictive state that matters for future behavior. We argue that history and memory should not be conflated: interaction tokens record what happened, whereas memory should explicitly represent what those interactions imply about the future. This motivates an explicit predictive memory that exposes the future-relevant meaning of the interaction history. We instantiate this principle with DeYi○Φ. DeYi distills earlier interactions into a user-specific predictive state trained to preserve future-relevant information while the original interaction sequence remains available to the generator. Φ then maps these private continuous states into a shared discrete vocabulary according to their induced future-item predictions rather than latent-space proximity, giving recurring predictive states across users reusable symbolic identities. During generation, the resulting memory token is prepended to the full interaction sequence and also defines a context-dependent predictive route before the target item's native SID. Remarkably, adding only a single predictive memory token can outperform generation from the full history alone, showing that explicit memory provides information complementary to the event sequence itself. These results suggest a simple principle for generative recommendation: history provides evidence, while memory provides predictive abstraction. Code is available at https://anonymous.4open.science/r/DeYiPhi-1537.

Primary area:
generative models
Keywords:
Generative Recommendation, Representation Learning, Sequence Modeling

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

The paper identifies an important limitation in generative recommendation: while items are represented as discrete tokens, user memory remains implicit. To address this gap, the paper proposes a novel model, DEYI, and its variant, DEYI-phi, which incorporate user history into future-predictive states through either user-specific or user-shared discrete identities. Experimental results demonstrate that the proposed models improve recommendation performance over baseline models using full interaction history, highlighting the potential benefits of explicitly modeling user memory in generative recommendation.

Critical strengths
  1. The paper is well motivated and addresses an important yet underexplored problem in generative recommendation. Specifically, user memory remains implicit when modeling SID-based interaction sequences, potentially limiting the information available for predicting the target item's SID.
  2. The paper is well written and easy to understand. The description of the memory components is particularly helpful in clarifying the types of information captured by the memory and how this information contributes to target item prediction.
  3. The proposed models are evaluated on large-scale datasets, and the results demonstrate improved performance over baseline models using full interaction history, providing empirical evidence for the effectiveness of the proposed memory component.
  4. Section 4.3, which analyzes the information encoded in the memory, provides useful insights into the internal workings of the proposed approach and helps readers better understand the role of memory in the model.
Critical weaknesses
  1. Despite the promising empirical results, the proposed methodology remains insufficiently convincing, and the underlying mechanism responsible for the observed improvements requires further investigation. The proposed approach can be interpreted as combining a summary of users' long-term interaction histories with recent interactions, thereby providing additional context through the memorization of long-term user preferences. The results in Table 3 appear to support this interpretation. For products with shorter interaction histories, the full-history baseline and the DEYI variants achieve similar performance, whereas DEYI-phi performs noticeably better, potentially because quantizing user histories introduces user-user collaborative signals through the memory mechanism. For short-video recommendation, the full-history variant substantially outperforms DEYI, possibly because DEYI loses access to long-term historical context. DEYI-phi performs better, potentially because quantized user histories capture collaborative signals similar to those exploited by user-user or item-based collaborative filtering. However, the current experiments do not sufficiently investigate these possible explanations or provide deeper insights into the underlying mechanism. Consequently, it remains unclear whether the observed improvements stem primarily from explicitly modeling user memory, incorporating long-term historical information, or capturing collaborative signals, which weakens the methodological justification.
  2. The use of memory to improve (generative) recommendation has been extensively studied in prior work [a, b, c, d]. However, the paper lacks an in-depth discussion of these related studies and does not sufficiently clarify how the proposed approach differs from or improves upon existing methods. A more thorough comparison with prior work is necessary to establish the novelty of the proposed method and better substantiate its effectiveness.

[a] Sequential Recommendation with User Memory Networks. WSDM 2018.

[b] TokenRec: Learning to Tokenize ID for LLM-based Generative Recommendation. TKDE. https://arxiv.org/pdf/2406.10450.

[c] VARIUM: Variational Autoencoder for Multi-Interest Representation with Inter-User Memory. WSDM 2025.

[d] Self-Evolving Memory for Generative Recommendation. https://arxiv.org/pdf/2609.15598.

  1. The paper lacks substantial comparisons with other state-of-the-art generative recommendation models. Such comparisons are important for determining whether the proposed user tokens provide tangible benefits over existing methods. Without these comparisons, it is difficult to assess the relative effectiveness of the proposed approach and establish whether its improvements are attributable to the proposed user memory mechanism rather than other factors.
  2. The evaluation relies exclusively on item SIDs from the existing OpenOneRec Rec-IF and the T5 generative recommender. Evaluating the proposed method with different tokenization strategies and generative recommendation architectures would provide more convincing evidence of its generalizability and effectiveness. Relying on a single tokenizer and generative recommender may introduce methodological bias and makes it difficult to determine whether the observed improvements are robust across different modeling choices.
Questions for authors

Please see the review.

2 comments

We thank the reviewer for recognizing the motivation, controlled analyses, and results on large-scale recommendation tasks. We clarify the mechanism, relationship to prior work, and generalization below.

W1. What Drives the Gains: Long-Term History or Cross-User Collaboration?

DeYiΦ provides an explicit, reusable predictive interpretation of history. All variants in Tables 2–3 retain the complete interaction sequence. Dummy and Pooling offer little benefit, while continuous DeYi alone does not reproduce the gains from shared discrete memory. Thus, the benefit comes from how historical information is represented and used, rather than access to additional interactions.

Cross-user sharing is central to Φ. The key question is what different users should share. Φ groups memories by future-item predictions rather than behavioral or latent-space proximity.

To test this distinction, we fix the DeYi states, 256 memory codes, and TIGER generation interface, changing only the quantization criterion.

Table R1. Predictive vs. Geometric Quantization on Product (%).

MethodP@10R@10N@10P@32R@32N@32
Euclidean VQ10.581.951.8019.483.922.64
Cosine VQ10.731.951.7719.283.872.58
DeYiΦ11.412.081.9320.514.092.79

Predictive quantization leads on all six metrics. Section 4.4 further shows that users assigned the same memory token exhibit more similar held-out future behavior than users assigned different tokens, across both domains.

Together, these results support prediction-aligned sharing as the mechanism that turns private user states into reusable memory identities, beyond generic history compression.

W2. Distinction from Existing Memory Methods

Our contribution is a shared discrete memory language whose identities are defined by future predictions and exposed to an autoregressive item-SID generator.

The four cited works address complementary aspects of recommendation memory:

Table R2. Comparison with Related Memory Methods.

WorkMain FocusDistinction from DeYiΦ
User Memory Networks (2018)Historical behavior memoryPer-user behavior modeling vs. shared predictive identities
TokenRecCollaborative user/item tokenizationCF-derived tokens vs. future-predictive memory tokens
VARIUMInter-user interest prototypesVAE memory retrieval vs. explicit SID-generation conditioning
Self-Evolving Memory (LION)Sparse memory for continual adaptationEvolving key-value memory vs. a shared predictive vocabulary

DeYi learns what earlier interactions predict; Φ assigns predictively equivalent states a common identity. The generator then uses this identity alongside the unchanged native Item SID sequence.

The distinction is therefore in the semantics and sharing criterion of memory, rather than memory or discretization alone. Item SIDs encode what happened; memory tokens encode what those events predict. We will make these relationships more explicit in the revised related-work section.

W3 / W4. Generalization Beyond TIGER and a Single SID Interface

To evaluate a second generative architecture, we extend DeYiΦ to OneRec on Short Video, reporting both Semantic ID (SID) and Product ID (PID) recommendation metrics.

Table R3. OneRec on Short Video, Seed 2026 (%).

Level / MethodP@10R@10N@10P@32R@32N@32
SID · FULL11.041.571.9814.792.112.25
SID · DeYiΦ12.221.872.2814.532.232.47
PID · FULL9.871.381.7512.971.821.97
PID · DeYiΦ11.191.692.0613.232.002.23

Compared with OneRec FULL, DeYiΦ improves all PID metrics and five of six SID metrics; SID Pass@32 remains higher for FULL. These results show that shared predictive memory can also benefit a different generative recommender and item-token interface.

Unlike comparing absolute scores across architectures, the within-OneRec comparison directly evaluates the contribution of predictive memory in that system.

Together with the original TIGER results on Product and Short Video, these findings support the applicability of shared predictive memory beyond a single generative backbone. Our three-seed Product evaluation further shows a Pass@32 gain in each generator initialization.

Thank you for providing a detailed response. However, the main concerns remain insufficiently addressed

  1. The underlying mechanism responsible for the performance improvements is still not convincingly established. The additional experiments demonstrate that predictive quantization performs better than conventional geometric quantization, but they do not adequately disentangle the respective contributions of long-term historical information, explicit user memory modeling, and cross-user collaborative signals. In particular, the observation that users sharing the same memory token exhibit similar future behavior does not necessarily demonstrate that the memory mechanism facilitates beneficial cross-user information transfer. More controlled ablation experiments are needed to isolate these factors.
  2. The discussion of related memory-based recommendation methods remains largely conceptual. Although the authors provide a comparison table describing the differences between the proposed method and existing approaches, they do not sufficiently establish the technical novelty or empirical advantages of the proposed method over closely related methods, particularly those involving collaborative tokenization and inter-user memory.
  3. The concern regarding comparisons with state-of-the-art generative recommendation models remains largely unresolved. The additional OneRec experiments only compare FULL and DeYi within the same architecture, rather than evaluating the proposed approach against representative state-of-the-art generative recommendation baselines. Consequently, it remains difficult to assess the relative effectiveness and competitiveness of the proposed method.
  4. Although the OneRec experiments with SID and PID provide some evidence of generalizability, the evaluation is still limited in terms of generative architectures, tokenization strategies, and recommendation scenarios. The current results are insufficient to establish that the proposed memory mechanism is robust across different modeling choices.

Thus, I would like to maintain my original score as it reflects my overall assessment of this paper.

Reviewer_u263OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that SID-based generative recommenders give items a discrete vocabulary but leave the user's state implicit in the raw interaction sequence. Its proposed fix is an explicit memory token, built around two questions: what the memory should represent, and when two users' memories should count as the same. The answer to both is prediction. A memory should keep what matters for predicting future behavior. Two memories should share a token when they lead to similar predictions about future items, even if their vectors are far apart. The method has two parts. DeYi splits off the older part of a user's history and encodes it into a unit vector with self-attention plus learned cross-attention queries. Each vector defines a softmax distribution over the item catalog and is trained to predict the user's next items. Φ then turns these continuous, per-user states into a shared codebook. It assigns each state to the codeword with the smallest KL divergence between their predicted item distributions. The paper derives a closed form for this KL that makes it computable without materializing the full catalog. For large catalogs, it uses a scalable version based on ANN retrieval for the high-probability items, importance sampling for the tail, medoid codewords and a distilled router. The resulting token is prepended to the user's SID sequence in TIGER, and the model is also trained to decode it before the target item's SID.

Critical strengths

The paper’s main strength is its central idea: assigning the same token to memories that predict similar future-item distributions, measured by KL divergence, rather than simply clustering nearby vectors. This better aligns quantization with the goal of recommendation. The method is also designed to scale, with a closed-form solution and practical techniques for handling a catalog of 12.9M items.

The intervention is lightweight, requiring only an extra token without changing the backbone or item SIDs. Under this setup, the method consistently outperforms plain TIGER across both domains, all cutoffs, and all three metrics, while the continuous-memory variant does not help. If these results hold, the method could offer meaningful gains at low adoption cost, and the controlled design strengthens the evidence.

Critical weaknesses

1: The method changes two things at once. It adds the memory token to the input, and it changes the decoder target from SID(y) to [g*, SID(y)]. With K=1, g* is simply the input token, so the first decoding step is nominally a copy. During beam search, however, the 32 beams can branch across different memory tokens, and that could act as a test-time ensemble over routes. The paper has no input-only or prefix-only ablation, and it does not say whether the Uniform VQ and Geometric VQ baselines in Table 3 use the same prefix target. It also omits the simple baselines that would show whether Φ's machinery matters:

  • DeYi used alone as a retriever;
  • a score-level ensemble of DeYi and TIGER;
  • k-means on the expected future-item embedding μ_z.

Since the memory is a deterministic function of the same history, it cannot add information. The paper's claim is therefore about why the token helps: predictive abstraction and prediction-aligned sharing. This weakness weighs most in my assessment because, without these controls, the same Table 2 numbers fit equally well with less interesting explanations: beam diversification, distillation from a second model trained on a multi-item future window, or a compressed summary that helps TIGER cope with 512-event histories.

2: For Short Video at K=1, Table 2 gives gains of +7.4% Pass@32, +8.1% Recall@32 and +11.5% NDCG@32. Fig. 5 reports +21.6%, +23.9% and +28.0% for what appears to be the same configuration, against the same FULL values. Appendix A's discussion is built on the larger figure numbers. On Product, the Recall and NDCG percentages in Fig. 4 also do not match Table 2. On top of this, no seeds, standard deviations or confidence intervals are reported anywhere. Several of the paper's conclusions rest on differences in the third or fourth decimal place, for example the claim that prediction gives "the strongest continuous abstraction" (0.1881 vs. 0.1877).

Questions for authors
  1. Which numbers are correct? For Short Video at K=1, Table 2 implies gains of +7–11% at @32. Fig. 5 reports +21.6%, +23.9% and +28.0% for what appears to be the same configuration and the same FULL baseline. On Product, the Recall and NDCG percentages in Fig. 4 also differ from Table 2. Please explain which runs each number comes from, and which numbers the paper's claims rest on.
  2. Is the gain from the input token or from the prefix token? Please report three variants: the memory token in the input only, the token as a decoding prefix only, and both. Do Uniform VQ and Geometric VQ in Table 3 use the same [g*, SID(y)] target? At inference with K=1, how often do the 32 beams contain a memory token different from the input token? I am trying to tell whether part of the gain comes from beam search branching over memory tokens.
  3. Simple baselines. Please report: (a) DeYi alone as a retriever, scoring items by zᵀe(i); (b) a score-level ensemble of DeYi and TIGER; (c) k-means on μ_z in place of Φ. Also, does the FULL baseline include TIGER's hashed user-ID token? These baselines would show whether the gain comes from Φ's prediction-aligned sharing or from adding a second model's predictions.
  4. Train/test consistency and model selection. Were the DeYi memories for TIGER's training users produced out-of-fold? If not, the training-time tokens come from a model that was fit on those users' targets. Table 1 lists only train and test splits, so how were G, τ, K and the TIGER hyperparameters selected?
  5. Online deployment. Has the method been deployed or tested online? If so, please report A/B results, even as anonymized relative changes, together with the serving cost: (a) the DeYi encoder and router at request time; (b) the extra decoding step; (c) how often the codebook and user memories are refreshed as histories and the catalog change.
2 comments

We thank the reviewer for highlighting predictive equivalence and efficient integration. We address the concerns about mechanism, numerical consistency, training, and deployment.

W1 / Q2. Input Token, Decoder Prefix, and Beam Branching

DeYiΦ uses shared predictive memory in both encoder conditioning and decoder-prefix prediction. We evaluate input-only, prefix-only, and combined variants.

Table R1. Component Analysis on Product, Seed 2026 (%).

MethodPass@32Recall@32NDCG@32
FULL18.793.742.57
Input-only19.923.992.70
Prefix-only17.543.472.43
Both (DeYiΦ)20.514.152.81

FULL and Both are from Table 2; single-component results use supplemental 32-beam training settings.

The combined model achieves the highest reported scores. The component results show distinct roles for encoder conditioning and decoder-prefix prediction.

At K=1K=1, Eq. (13) reduces to g∗(r,y)=gr1g^{*}(r,y)=g_{r1}. Our audit covers 18,744 users and 599,808 valid beams: each user has one route, with zero route mismatches. Thus, the main model does not rely on beam diversification across memory routes. Target-dependent route selection applies to K>1K>1.

W1 / Q3. Predictive Memory vs. Simple Baselines

DeYiΦ turns future-predictive states into reusable discrete identities, rather than simply adding a second predictor. We evaluate DeYi retrieval, score fusion with TIGER, and K-means on expected future-item embeddings.

Table R2. Simple Baselines on Product (%).

MethodPass@32
DeYi Retrieval1.07
DeYi + TIGER Score Fusion13.33
K-means on Expected Future Embedding19.72

Retrieval and score fusion do not reproduce the benefit of predictive memory. K-means also trails predictive quantization. Along with the Dummy, Pooling, and VQ ablations in Table 3, these results support prediction-aligned sharing beyond generic summarization.

Memory tokens are derived from observed history; their value is to expose future-relevant structure in a form shared across users.

In our supplemental controls, Euclidean, Cosine, and Predictive VQ use the same decoder target, [g*, SID(y)]. Predictive VQ leads on all six metrics. FULL does not use a hashed user-ID token or learned user-ID embedding.

W2 / Q1. Consistency Between Table 2 and Figures 4–5

The percentages use different reference models. Table 2 reports absolute scores, comparing DeYiΦ primarily with FULL. Figures 4–5 annotate gains over continuous DeYi at the same KK; the dashed line denotes FULL. This explains the apparent differences for both domains. The main performance claims rest on Table 2.

We will clarify the captions: Percentage annotations denote relative gains over continuous DeYi at the same K; the dashed line denotes FULL.

We also evaluate three TIGER seeds (2026, 2027, 2028) on Product, keeping the DeYi teacher, Φ vocabulary, and test users fixed.

Table R3. Three-seed Product Results, Mean ± SD (%).

MethodP@10R@10N@10P@32R@32N@32
FULL10.26±0.161.93±0.011.76±0.0218.78±0.253.80±0.062.56±0.04
DeYiΦ11.33±0.122.07±0.021.90±0.0220.49±0.174.14±0.052.78±0.01

DeYiΦ improves Pass@32 in all three seeds and the mean of all six metrics. The gain persists across generator initializations, supporting shared predictive memory alongside complete interaction history.

Q4. Train/Test Consistency and Model Selection

Training and test users are disjoint; predictive-memory models and the vocabulary are frozen at evaluation time. Product contains 75,987 training users and 18,744 test users. Of the training users, 72,328 fit DeYi and 3,659 serve for checkpoint validation. Φ constructs its vocabulary from training-user states.

Some training-user memories are generated in-sample. Test users do not enter DeYi fitting or Φ vocabulary construction; their tokens come from frozen models.

The main configuration uses G=256G=256, τ=0.15\tau=0.15, and K=1K=1, with consistent TIGER settings for FULL and DeYiΦ. The capacity study examines different KK values. We will clarify the data splits and model-selection protocol in the revision.

Q5. Online Deployment and Serving Cost

Our work focuses on offline recommendation. Online A/B testing and deployment-performance evaluation are future work.

Thank you for your reply. I will remain my score.

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

This paper proposes DEYI ◦Φ, a framework for adding discrete predictive memory to generative recommendation. DeYi learns user-specific memory states using future-item supervision over frozen item embeddings, while Φ quantizes these states based on the KL divergence between their induced item distributions. The resulting tokens are prepended to the full interaction history and predicted before the target item SID. Experiments with TIGER on Product and Short Video report improvements over full-history generation.

Critical strengths

The paper is clearly organized, and the distinction between interaction history and predictive abstraction is well motivated. The ablations comparing continuous memory with different quantization strategies provide useful evidence that the discrete representation contributes to the reported gains.

Critical weaknesses

The novelty is not sufficiently established. The proposed method combines a next-item-supervised history encoder, KL-based quantization, and discrete prefix conditioning. While this is a reasonable design, the paper does not convincingly distinguish it from existing approaches to user-interest modeling, history compression, and discrete representations. The empirical evaluation is also narrow: all recommendation experiments use TIGER on two tasks from a single benchmark, without evaluation on another generative-recommender backbone or reporting uncertainty across random seeds.

More importantly, the central claim of predictive equivalence is supported by a surrogate distribution defined through frozen item embeddings, rather than the downstream generator’s actual next-item distribution. The paper does not establish that KL similarity in this surrogate space reliably reflects changes in recommendation behavior. Furthermore, in the default K = 1 setting, Eq. (13) reduces to (g^*(r,y)=g_{r1}), which is independent of the target item (y). Thus, the main experiments do not validate the target-conditioned routing interpretation. These limitations leave the central methodological claim insufficiently substantiated.

Questions for authors

Can you demonstrate that the KL distortion in Eq. (7) correlates with changes in TIGER’s actual next-item distribution, rather than merely reflecting similarity in the frozen embedding space? Do the gains persist across multiple random seeds and a second generative-recommender backbone?

2 comments

We thank the reviewer for recognizing the distinction between interaction history and predictive abstraction. We address the concerns about novelty, predictive equivalence, routing, and generalization below.

W1. Novelty: What makes discrete predictive memory distinct?

The key question is when different user histories should share a memory symbol. Our answer is predictive equivalence: histories with similar future-item predictions may share a symbol even when their events or latent vectors differ.

DeYi learns future-predictive states from earlier interactions. Φ groups these states by KL distortion between their induced item distributions. The generator uses the resulting memory tokens alongside the unchanged full SID history. Thus, unlike methods that replace history with a compressed summary, we preserve the original events; unlike geometric quantization, we define shared identity by predictive behavior, not vector distance.

To isolate this choice, we hold DeYi states, the 256-code vocabulary size, and the TIGER generation interface fixed.

Table R1. Product: Predictive vs. geometric quantization (%).

MethodP@10R@10N@10P@32R@32N@32
Euclidean VQ10.581.951.8019.483.922.64
Cosine VQ10.731.951.7719.283.872.58
DeYiΦ11.412.081.9320.514.092.79

Predictive quantization leads on all six metrics. The sharing criterion itself matters, beyond making memory discrete. Item SIDs encode what happened; memory tokens encode what those events predict.

W2 / Q1. Does teacher-side KL reflect downstream generation?

Eq. (7) defines equivalence between memory states, not equality between generator outputs: dpred(z,c)=DKL(pz∥pc)d_{\mathrm{pred}}(z,c)=D_{\mathrm{KL}}(p_z\parallel p_c).

Here, pzp_z and pcp_c are distributions induced in the frozen item-embedding space. Φ uses this distortion to assign shared memory identities; TIGER separately learns to generate native SIDs from the full history and these identities. Teacher-side KL therefore does not guarantee matching generator-side KL.

We test two observable consequences instead. Section 4.4 of the paper examines whether users sharing a memory symbol show more similar held-out future behavior. Table R1 tests whether prediction-aligned sharing improves downstream recommendation under a matched generation interface. Both support the utility of predictive memory identities; neither is presented as proof of generator-KL preservation.

W3. Is routing target-dependent when K = 1?

No. At K=1K=1, Eq. (13) reduces exactly to g∗(r,y)=gr1g^{*}(r,y)=g_{r1}.

Each user has one fixed memory route. Our Product audit covers 18,744 users and 599,808 valid beams: every user has one route, with zero route mismatches. The K = 1 results thus test conditioning on a shared predictive memory, not nontrivial target-dependent route selection or beam diversification over routes. Target-dependent selection arises only when K > 1.

W4 / Q2. Are results robust to seeds and a second backbone?

We reran the Product comparison with seeds 2026, 2027, and 2028.

Table R2. Product: Three-seed mean ± SD (%).

MethodP@10R@10N@10P@32R@32N@32
FULL10.26±0.161.93±0.011.76±0.0218.78±0.253.80±0.062.56±0.04
DeYiΦ11.33±0.122.07±0.021.90±0.0220.49±0.174.14±0.052.78±0.01

DeYiΦ improves mean performance on every metric, with a Pass@32 gain in each seed. The submitted Table 2 also reports improvements on both Product and Short Video.

For a second generative backbone, we evaluate OneRec on Short Video, reporting both Semantic ID (SID) and item ID (PID) metrics.

Table R3. OneRec, Short Video (%).

Level / MethodP@10R@10N@10P@32R@32N@32
SID · FULL11.041.571.9814.792.112.25
SID · DeYiΦ12.221.872.2814.532.232.47
PID · FULL9.871.381.7512.971.821.97
PID · DeYiΦ11.191.692.0613.232.002.23

DeYiΦ improves all PID metrics and most SID metrics; SID Pass@32 remains slightly higher for FULL. These results show that predictive memory can also help a different generative backbone, while its effect varies by metric. The OneRec result is from seed 2026.

Reviewer_zEsEOpenReview ↗

response

Thanks for your responses, which have addressed some concerns. I have increased my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.