acceptodds
Under review as a conference paper at ICLR 2027

ProtoAdapter: Data-Centric Vertical Adaptation for Pretrained Generative Recommenders

Abstract

Generative recommenders must adapt pretrained models to new verticals, where item-token alignment alone does not teach how users transition between items and target-domain interactions are often scarce. Existing continual pretraining and supervised fine-tuning primarily learn from individual histories, leaving two complementary signals implicit: recurring behavioral patterns shared across the domain and history-specific evidence that identifies a user's current position within such a pattern. We propose ProtoAdapter, an offline data construction framework for vertical adaptation that makes both signals explicit. ProtoAdapter builds a bank of representative behavioral prototypes from the target-domain corpus, retrieves compatible prototypes for each user history, and selects supporting events from that history. A teacher model then organizes them into a structured trace containing the vertical pattern, observed evidence, and transition bridge, with the target action appended as the supervised answer. Trace-augmented and direct-answer examples are jointly used for continual pretraining and supervised fine-tuning, while deployment retains direct item-token generation without requiring traces. Experiments on several recommendation datasets with Qwen3 and OneRec show consistent improvements across ranking metrics. The gains become more pronounced in low-data settings: on KuaiRand, using only 10% of the training data yields a relative improvement of up to 29.7%. Further analyses verify the effects of trace construction and demonstrate that the proposed two-scale supervision improves data efficiency by connecting reusable domain transitions with sequence-specific evidence. Training code: https://anonymous.4open.science/r/ProtoAdapter/.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Recommender Systems, Generative Recommendation, Domain Adaptation, Data-Centric Learning, Rationale Supervision

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_9h3qOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ProtoAdapter, a data-centric framework for adapting pretrained generative recommenders to new domains. It constructs a bank of shared behavioral prototypes, retrieves relevant patterns, and selects history-specific evidence to generate structured training traces using a teacher model. Trace-augmented and direct-answer supervision are jointly employed during CPT and SFT, while inference retains direct item-token generation. Experiments on three datasets using Qwen3 and OneRec demonstrate consistent ranking improvements, particularly under limited training data.

Critical strengths

S1. Well-motivated and practical framework. Combining domain-level behavioral prototypes with history-specific evidence provides a reasonable approach to sparse-domain adaptation. Offline trace construction also avoids additional inference-time overhead.

S2. Consistent empirical improvements. The method improves two different backbones across three datasets, with particularly encouraging gains in low-data settings. The shuffled and cross-domain trace ablations further support the importance of coherent supervision.

Critical weaknesses

1.Insufficient isolation of the core contributions. The ablations do not separately establish the benefits of prototype retrieval and local evidence selection. Since the training target also guides trace construction, the gains may partly arise from target-aware rationale distillation rather than the proposed two-scale mechanism.

  1. Incomplete efficiency and validation analysis. Offline prototype construction and teacher-based trace generation introduce additional costs that are not quantified. The low-data protocol and statistical evidence also require clarification to better establish the claimed data efficiency.
Questions for authors
  1. It would be helpful to include a teacher-generated rationale baseline without prototypes, along with prototype-only and evidence-only ablations, to better understand the contribution of each component.

  2. Are prototype banks rebuilt using only the corresponding 10%/50% training subsets? Reporting offline construction costs and comparing against direct training under similar computational budgets would also help assess the practical efficiency of the method.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 9h3q

Thank you for recognizing the complementary motivation and the direct-decoding setting. We respond in two parts.

  1. Response to W1/Q1 (component contributions and fair budgets). Shared Table S1 shows that prototype-only and evidence-only each improve over direct training, while the full method gives the largest gain on all three datasets. Shared Table S2 shows that generic summaries, matched-length filler, and Direct-extended remain below the full method, so the benefit comes from organized evidence even after accounting for additional text and training tokens.
  2. Response to W2/Q2 (independently rebuilt low-data assets and cost). The low-data experiment checks whether each fraction uses only that fraction of information. For each fraction, we rebuild the fragment pool, prototype bank, selector statistics, retrieval index, and teacher cache using only that fraction of target-domain training histories. The test set and evaluator remain fixed. Three paired data subsamples are shared by direct and ProtoAdapter, with training seeds 17, 43, and 91. Results are in Table 1 below.

Table 1: Low-data results with independently rebuilt banks

Training fractionOneRecProtoAdapter, rebuilt bankAbsolute gainRelative gain
10%2.12 ± 0.052.75 ± 0.06+0.63 [0.53, 0.74]+29.7%
50%3.09 ± 0.043.55 ± 0.05+0.46 [0.38, 0.55]+14.9%
100%3.64 ± 0.033.98 ± 0.04+0.34 [0.28, 0.41]+9.3%

The larger relative benefit under limited data remains after rebuilding all construction assets. As an additional diagnostic, using a full-data bank with 10% student-training data gives 2.91 NDCG@5, compared with 2.75 for the strict 10% bank. The former measures access to a larger training bank, while the latter measures performance under a complete 10% construction budget. Shared Table S3 gives the complete offline account: teacher generation takes 5.6, 19.7, and 14.9 A100-hours on All_Beauty, KuaiRec, and KuaiRand, and adaptation takes 9.4, 31.2, and 24.6 A100-hours. At deployment, OneRec and OneRec+ProtoAdapter have p50 latencies of 118ms and 119ms on the pooled 6,144-request held-out evaluation (2,048 per dataset), and peak allocated memory of 19.8GB and 19.9GB.

Thank you for the detailed rebuttal and additional experimental results. I appreciate the authors' efforts to address my concerns. The prototype-only and evidence-only ablations, together with the generic-summary and matched-budget controls, provide stronger evidence for the effectiveness of the proposed framework. The clarification of the low-data protocol, statistical results, and offline construction costs also improves the overall experimental evaluation.

Nevertheless, some limitations remain. While the new ablations demonstrate that both evidence sources contribute to performance, their measured interaction is relatively small, leaving the extent of their synergistic benefit less clear. Moreover, although the matched-budget controls improve comparison fairness, the additional offline construction costs are not fully considered in an end-to-end cost-effectiveness evaluation.

Overall, the rebuttal has satisfactorily addressed most of my original concerns and strengthened the empirical evidence. However, I believe my original rating still appropriately reflects the paper's contributions and remaining limitations. Therefore, I will maintain my rating of 3 (Weak Acceptance).

Thank you again for your careful review, and if you have any new questions in the future, you are welcome to continue the discussion with us!

Reviewer_TS3sOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ProtoAdapter, an offline data construction framework for adapting pretrained generative recommenders to new verticals (e.g., beauty products, short videos). The core observation is that individual user histories are often too short to reveal complete behavioral transitions, but the same transitions recur across many users in a domain. ProtoAdapter addresses this by (1) building a bank of "behavioral prototypes" from the target corpus that capture reusable transition patterns, (2) retrieving compatible prototypes for each user and selecting supporting events from that user's history, and (3) having a teacher model synthesize structured traces with three slots (vertical pattern, observed evidence, transition bridge) with the target item appended as the supervised answer. These trace-augmented examples are mixed with standard direct-answer examples during both continual pretraining and supervised fine-tuning. At deployment, traces are discarded and the model just does direct item-token generation. Experiments on All Beauty, KuaiRec, and KuaiRand with Qwen3-1.7B and OneRec-1.7B show consistent improvements, with the largest relative gains (up to 29.7%) in low-data settings. Ablations show that shuffled, cross-domain, and raw-pattern traces all underperform the full construction.

Critical strengths

The most consequential strength is the two-scale evidence design and the way it addresses a real gap in vertical adaptation. Existing rationale distillation methods generate explanations for individual examples without connecting them to reusable domain-level patterns. ProtoAdapter's insight that a prototype bank can supply transition context that sparse histories cannot reveal on their own, while selected history events anchor that context to the specific user is clean and well-motivated. The ablations in Table 3 provide real evidence that this structure matters: cross-domain traces and shuffled traces both degrade performance substantially, which means the gain is not just from having more text in the training signal. The data efficiency results in Figure 3(b) and Table 5 are the most compelling validation of this idea: at 10% training data, augmented OneRec nearly matches the baseline at 50% data, which is a practical result that could matter for cold-start verticals.

Critical weaknesses

The biggest weakness is the unexamined role of label leakage through the evidence selection step. The saliency(i_t, y_u) term in Eq. 4 scores history events by their relevance to the target item. This means the offline construction effectively tells the model which events predict the label, and the teacher then wraps this into a narrative. It is plausible that much of the measured improvement comes from this privileged selection rather than from the prototype-based behavioral structure that the paper emphasizes. Without an ablation that removes the saliency term (or replaces it with a proxy that does not use y_u), the paper cannot distinguish between these two explanations. This is not a fatal flaw, since label-informed data augmentation is a legitimate technique, but the current framing overstates the role of domain-level behavioral prototypes relative to what might simply be label-guided attention over history events.

The second weakness is the lack of specificity around the teacher model and the cost of the construction pipeline. Building the prototype bank, retrieving prototypes, selecting evidence, and generating traces for every training example is a substantial offline investment. The paper presents this as a practical data construction framework, but gives no information about wall-clock time, teacher model identity, API cost, or failure rate of the quality filter. For a method positioned around data efficiency, knowing the total cost of producing that data (including teacher inference) is essential for assessing whether the efficiency gain is real or just shifts the expense.

Questions for authors

First, what happens to the results if you remove the saliency term from Eq. 4 and select evidence using only recency, continuity, and domain fit? This would clarify how much of the gain comes from label-informed evidence selection versus the prototype construction itself.

Second, which specific model serves as the teacher G_phi, and what fraction of generated traces are rejected by the quality filter? Relatedly, how sensitive are the results to teacher capability? Have you tried a weaker teacher?

Third, the cross-domain ablation in Table 3 actually falls below the direct baseline on both datasets. Can you comment on why cross-domain traces hurt rather than just being neutral? This seems to suggest the model is sensitive to domain mismatch in ways that go beyond simply not helping.

3 comments
AuthorsOpenReview ↗

Response to Reviewer TS3s

Thank you for focusing on target saliency, teacher choice, and the full construction cost. We respond in three parts.

  1. Response to W1/Q1 (benefit after removing target guidance). Target-free results are in shared Table S2. This version rebuilds selection, retrieval, and traces using only observed histories, without target saliency or teacher access to the label. KuaiRec improves from 9.19 to 9.58 and KuaiRand from 3.64 to 3.87. Full target-aware construction reaches 9.77 and 3.98. The difference combines target-guided evidence selection with the teacher's label-based consistency checking.
  2. Response to W2/Q2 (teacher strength and cost). Table 1 below compares Qwen3-7B and Qwen3-32B using the same prompts, filters, common-request protocol, and training budgets, with a fresh common accepted set for the two-teacher comparison.

Table 1: Teacher-strength comparison

TeacherAll_Beauty NDCG@5KuaiRec NDCG@5KuaiRand NDCG@5Acceptance, All_Beauty / KuaiRec / third
Qwen3-7B-Instruct10.589.613.8691.8% / 90.9% / 90.4%
Qwen3-32B-Instruct10.769.773.9893.1% / 93.3% / 92.9%

The smaller teacher also improves over direct training, showing that the construction works with a smaller teacher. The stronger teacher further improves acceptance and final quality through better trace organization. Shared Table S3 gives the complete offline account: teacher generation takes 5.6, 19.7, and 14.9 A100-hours on All_Beauty, KuaiRec, and KuaiRand, and adaptation takes 9.4, 31.2, and 24.6 A100-hours. Bank construction, retrieval, and teacher generation happen before adaptation; all are offline costs. Serving remains direct decoding, with 118/119ms p50 for OneRec and ProtoAdapter on the pooled 6,144-request held-out evaluation (2,048 per dataset). 3. Response to Q3 (why cross-domain traces hurt). Table 2 defines low agreement as a mean evidence–prototype maximum-cosine score below 0.45, computed over all traces in the corresponding arm. The loss column is the relative increase in direct-answer loss on fixed validation for cross-domain versus domain-matched traces.

Table 2: Cross-domain error analysis

DatasetLow agreement, cross-domainLow agreement, domain-matchedCross-domain NDCG@5Direct NDCG@5Direct-answer loss change vs. domain-matched, validation
KuaiRec31.8%8.6%9.119.19+6.4%
KuaiRand34.5%9.2%3.543.64+5.8%

Cross-domain traces contain substantially more weakly aligned transitions, alongside higher validation direct-answer loss and lower NDCG. These results are consistent with one interpretation: weakly aligned cross-domain connections can interfere with direct-answer learning. We now describe domain compatibility as a construction requirement and include the full error analysis in Appendix F.

Reviewer_TS3sOpenReview ↗

Reponse to authors' comments

Thank you for the new experiments.

On W1/Q1. The target-free ablation was the experiment I most wanted to see. KuaiRec 9.19→9.58 and KuaiRand 3.64→3.87 without any label access accounts for roughly two-thirds of the full gain. That supports the prototype contribution claim. Two remaining issues. First, "target-free" removes the saliency term and the teacher's label access simultaneously. The remaining third of the gain could come from either mechanism, and the current ablation cannot distinguish them. Second, in Table 4, evidence-only (9.51/3.83) outperforms prototype-only (9.43/3.78) on both datasets, and evidence-only still benefits from target saliency. The data suggests the two scales are closer to equal partners than the prototype-first narrative implies. I'd like the framing adjusted accordingly.

On W2/Q2. The cost breakdown and the Qwen3-7B comparison are what was missing. So seems to be resolved.

On Q3. The low-agreement analysis is convincing on both datasets (31.8% vs 8.6% on KuaiRec, 34.5% vs 9.2% on KuaiRand, with clear loss increases). Misaligned traces actively interfere rather than just adding noise. Addressed. Updated assessment. The rebuttal fully resolves the cost and cross-domain concerns and makes real progress on label leakage. The core claim that prototype construction contributes beyond label information is now supported. I raise my score to 3: Weak accept (from 2), contingent on the camera-ready (1) presenting the two evidence scales as roughly co-equal rather than leading with prototypes, and (2) moving the cost table and cross-domain error analysis into the main paper.

Thank you again for your careful review. We will revise the paper according to your suggestions, and if you have any new questions in the future, you are welcome to continue the discussion with us!

Reviewer_hNg3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ProtoAdapter builds offline training data for adapting generative recommenders to a new domain. It selects recurring transition prototypes from target-domain histories, retrieves relevant prototypes and past events for each training example, and asks a teacher to write a three-part trace before the labeled next item. Joint trace and direct-answer training occurs during CPT and SFT; inference still directly generates item tokens.

Critical strengths
  1. Combining domain-wide patterns with examples' own observed events is a plausible way to teach a model about sparse target-domain histories. Training includes direct-answer examples, so deployment retains the original item-token interface (Sections 3–4).

  2. Both backbones improve on all three datasets and three metrics (Table 2). The 10% KuaiRand result rises by 0.63 NDCG@5 percentage points, a larger absolute gain than at full data (Table 5). Shuffled, cross-domain, and raw-pattern traces underperform the full construction (Table 3), suggesting that trace content matters.

Critical weaknesses
  1. The offline event selector explicitly uses the target item, and the teacher receives the label for consistency checks (Section 4). This is legitimate training-time supervision, not test leakage. However, the ablations do not compare against an equally sized teacher trace made from the same history and target without the prototype bank, or a compute/token-matched direct-answer control. Shuffling or replacing traces also changes their coherence and quality, so Table 3 cannot show that prototype retrieval, rather than additional target-aware teacher information, drives the gain.

  2. Section 5.1 states that experiments use multiple runs and the improvements are significant, but the paper provides neither run counts nor variation, confidence intervals, or the test used. The teacher identity and key construction/training settings are also insufficiently specified in the manuscript. These omissions matter when full-data NDCG@5 improvements are 0.28–0.70 percentage points (Table 2); they prevent assessment of robustness and fair cost comparisons.

Questions for authors
  1. With the same teacher, labels, training-token budget, and CPT/SFT schedule, what happens if traces use the history and target but no retrieved prototypes? How much does each of prototype retrieval and target-guided event selection contribute?

  2. How many independent seeds were used for each result, what statistical test produced the significance claim, and what are the per-run values or uncertainty intervals? Which teacher and exact prompt/selection settings were used?

3 comments
AuthorsOpenReview ↗

Response to Reviewer hNg3

Thank you for identifying why the original ablations did not isolate prototype retrieval. Shuffled, cross-domain, and raw-pattern traces changed several factors at once, so we separate the two evidence scales and the target-aware teacher contribution as follows.

  1. Response to W1/Q1 (fair attribution and target-free construction). The four-cell results are in shared Table S1, and target-free plus matched-budget controls are in Table S2. Direct, prototype-only, evidence-only, and full construction reach 9.19, 9.43, 9.51, and 9.77 on KuaiRec, and 3.64, 3.78, 3.83, and 3.98 on KuaiRand. The three teacher-trace variants share the teacher, target visibility, common accepted request pool, trace lengths, direct:trace ratio, supervised tokens, and optimizer updates. The submitted Direct row is the original reference; Direct-extended reaches 9.24 and 3.67 under the exact same token and update budget. Target-free construction removes saliency and withholds the label from the teacher, reaching 9.58 and 3.87, while full target-aware construction reaches 9.77 and 3.98. The difference includes target-guided evidence selection and the teacher's label-based consistency checking.
  2. Response to W2/Q2 (seed-level results and statistics). See shared Table S4, located immediately after Table S3 and before the individual responses. It lists the three seed-level values for Full and target-free construction and the full-minus-direct paired CIs. Tables S1 and S2 contain the corresponding mean results and Direct-extended control. The SD describes training variation, while the CI is a request-level paired bootstrap over seed-averaged predictions. Appendix B provides every variant's seed-level values, exact teacher/prompt settings, selection formulas, and training configuration.

Thank you for the additional experiments. The component ablations, matched-budget controls, and target-free results address my main concerns. The seed-level results also strengthen the evidence. I am updating my recommendation to positive.

Thank you very much for your suggestions. If you have any further questions, we would be glad to continue the discussion.

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

The paper studies how to effectively adapt pretrained generative recommenders to new verticals. The authors propose ProtoAdapter, which constructs a prototype bank to capture recurring behavioral patterns across the domain and selects sequence-specific evidence to identify relevant interactions within the user's history. These two sources of information are then organized into structured traces using a teacher LLM, which are incorporated into continual pretraining and supervised fine-tuning alongside direct recommendation supervision. Experimental results demonstrate improvements in overall recommendation quality, with particularly notable gains in low-data settings.

Critical strengths
  1. The paper presents an interesting approach to capturing recurring behavioral patterns across the target domain, which may otherwise remain implicit during conventional recommendation training. By using a teacher LLM to summarize these patterns in natural language and connect them with sequence-specific evidence, the proposed framework provides structured supervision that can help the model better leverage domain-level behavioral knowledge.

  2. The experimental results show promising improvements when training data is limited. For example, on KuaiRand, ProtoAdapter achieves a 29.7% relative improvement in NDCG@5 with only 10% of the training data, compared to 9.3% with the full dataset. This suggests that the proposed approach may be especially beneficial for adapting generative recommenders to data-scarce verticals.

Critical weaknesses

Several important components of ProtoAdapter are described with their concrete implementations not sufficiently specified.

  1. Section 4.1 (Eq. (4)) introduces a weighted combination of recency, continuity, domain compatibility, and target-related saliency to select relevant historical events. However, the paper does not provide precise computational definitions for these scoring functions or explain whether their outputs are normalized. Also, from what I’ve checked, the values of wrw_r, wcw_c, wdw_d, and wsw_s, as well as the criteria used to determine them, are not reported. Since evidence selection directly affects prototype retrieval and subsequent trace generation, these missing details concern a central component of the proposed method. Furthermore, Section 4.2 (Eq. (5)) introduces α\alpha to balance history similarity and evidence agreement, while Section 4.3 (Eq. (6)) introduces β\beta to balance direct and trace supervision. However, their actual values and selection procedures are not specified. It is also unclear how the representation function q(⋅)q(\cdot) and the similarity and matching functions in Eq. (5) are implemented.

  2. Section 4.2 states that "Traces that copy target identifiers or introduce unsupported evidence are rejected." However, the paper does not explain how these violations are detected, particularly when unsupported historical events are expressed in natural language rather than through explicit identifiers. It is unclear whether the verification relies on deterministic rules, an additional LLM-based verifier, or another mechanism. Moreover, the frequency of rejected traces is not reported. Since the proposed approach relies heavily on the teacher-generated traces, the absence of these details raises concerns about the reliability and reproducibility of the constructed supervision.

Questions for authors
  1. In evidence selection, how are the four scoring functions implemented, and how are wrw_r, wcw_c, wdw_d, and wsw_s determined? Could the authors provide their actual values and selection criteria?

  2. How are α\alpha and β\beta selected? Are they tuned on a validation set, and if so, what search ranges are used? Could the authors also clarify the implementations of q(⋅)q(\cdot), sim, and match?

  3. Given that both stages of CPT and SFT use direct and trace supervision, how do they differ in terms of training objectives, data configurations, and optimization settings?

  4. In trace verification, how are copied target identifiers and unsupported historical evidence detected and filtered? Is an additional verification model used? What proportion of generated traces is rejected for each reason?

  5. Why is ProtoAdapter applied only to Qwen3 and OneRec, rather than other generative recommendation models such as TIGER or HSTU? Could the authors clarify whether the method is compatible with these architectures and, if possible, provide additional experiments demonstrating its effectiveness beyond the two evaluated backbones?

4 comments
AuthorsOpenReview ↗

Response to Reviewer zcdN

Thank you for listing the implementation details needed to reproduce the method. We have expanded the construction, filtering, and training description as follows.

  1. Response to W1/Q1 (evidence scores and weights). For historical event i_t, the four inputs are:
  • Recency: normalized temporal position (t+1)/|h_u|.
  • Continuity: mean cosine similarity to the adjacent observed events, using one neighbor at an endpoint.
  • Domain fit: cosine similarity between the event's item-text embedding and the target-domain item-text centroid.
  • Target saliency: cosine similarity between the event embedding and the training target's item-text embedding. Each score is min–max normalized using statistics from the corresponding training domain. The default weights are (0.20, 0.25, 0.20, 0.35), selected on validation from weights in [0,0.5] with sum one. The target-free variant removes the last term and renormalizes the remaining weights to approximately (0.3077, 0.3846, 0.3077).
  1. Response to W1/Q2 (retrieval functions and parameter choices). The representation q is the mean-pooled, L2-normalized item-text embedding from a frozen Qwen3-Embedding-0.6B encoder. sim is cosine similarity. match(e,p) averages, over selected events, the maximum cosine similarity to the prototype's supporting events. Retrieval uses score = alpha × sim(q(history),q(prototype)) + (1−alpha) × match(evidence,prototype), with alpha=0.6 selected from {0.3, 0.5, 0.6, 0.7, 0.9}. The loss uses beta=1.0, selected from {0.25, 0.5, 1.0, 2.0}. The reference bank has K=32, M=4, and a 300-character trace cap.
  2. Additional response to Q1 (selector-weight sensitivity). The five target-aware rows change only the four weights (w_r,w_c,w_d,w_s) and rebuild the selected evidence and traces. The target-free row is an additional reference that also removes saliency and withholds the training label from the teacher. | Weights | KuaiRec NDCG@5 | KuaiRand NDCG@5 | |---|---:|---:| | Target-free, renormalized (0.31,0.38,0.31,0) | 9.58 | 3.87 | | Balanced (0.25,0.25,0.25,0.25) | 9.69 | 3.93 | | Default (0.20,0.25,0.20,0.35) | 9.77 | 3.98 | | Higher saliency (0.15,0.15,0.15,0.55) | 9.81 | 4.00 | | Recency-dominant (0.55,0.15,0.15,0.15) | 9.61 | 3.89 | | Continuity-dominant (0.15,0.55,0.15,0.15) | 9.65 | 3.91 |

All six settings improve over direct training. The default is close to the best tested setting, and the target-free row again shows a substantial gain without saliency.

AuthorsOpenReview ↗

Continued: Response to Reviewer zcdN

  1. Response to W2/Q4 (teacher and deterministic filtering). The offline teacher is Qwen3-32B-Instruct, with temperature 0.2, top-p 0.9, and at most 256 generated tokens. The prompt template hash is 7c91…a42e. The teacher receives the observed history, prototype descriptions, selected evidence, and the training label for consistency checking. The requested trace excludes target identifiers and future events. Filtering uses three deterministic checks: exact target item/SID string matching detects copying; event pointers and normalized item-metadata matching detect unsupported evidence; slot parsing checks the required sections. Rejected traces become direct-answer examples. | Dataset | Generated traces | Accepted | Rejected: target copy | Rejected: unsupported evidence | Rejected: malformed slots | Direct-answer fallback | |---|---:|---:|---:|---:|---:|---:| | All_Beauty | 36,420 | 33,905 (93.1%) | 1.2% | 4.4% | 1.3% | 2,515 | | KuaiRec | 128,000 | 119,424 (93.3%) | 1.4% | 4.0% | 1.3% | 8,576 | | KuaiRand | 96,000 | 89,184 (92.9%) | 1.6% | 4.1% | 1.4% | 6,816 |

The controlled comparisons use the intersection of requests accepted by full, prototype-only, evidence-only, generic-summary, and target-free traces. This is why the common trace pools are smaller than the raw full-method accepted counts.

  1. Response to Q3 (CPT and SFT). CPT uses two passes with learning rate 1e−5; SFT uses one pass with learning rate 5e−6. Both use AdamW, weight decay 0.01, 3% warmup, and the same validation-based checkpoint rule. A token-batched sampler caps each optimizer update at 4,096 supervised tokens, so the number of records varies with the direct/trace composition and the update counts follow directly from the total budgets. Direct and trace losses are normalized separately by their supervised token counts, and the target answer is supervised once in each corresponding branch. Every attribution variant uses its corresponding data construction in both stages. The submitted direct schedule uses seven direct-only passes, giving 18.65M/13.99M tokens on KuaiRec/KuaiRand. Direct-extended repeats direct examples until it reaches the same total supervised tokens and optimizer updates as the trace variants. The stage comparison also retains the submitted SFT-only result of 9.50/3.76: moving to full CPT+SFT at 9.77/3.98 adds another 0.27/0.22 points.

  2. Response to Q5 (additional backbone experiments and integration). We added TIGER and HSTU using the same direct/trace construction, CPT+SFT schedule, and evaluator. Their semantic item IDs are mapped into paired direct/trace records, with the corresponding item-token output heads adapted. The table below aligns Pass@1, HR@5, and NDCG@5 with the main paper. | Dataset | Backbone | Variant | Pass@1 | HR@5 | NDCG@5 | |---|---|---|---:|---:|---:| | All_Beauty | TIGER | Submitted direct baseline | 3.20 | 8.73 | 5.92 | | All_Beauty | TIGER | + ProtoAdapter | 3.58 | 9.41 | 6.26 | | All_Beauty | HSTU | Submitted direct baseline | 3.08 | 8.46 | 5.72 | | All_Beauty | HSTU | + ProtoAdapter | 3.45 | 9.19 | 6.12 | | KuaiRec | TIGER | Submitted direct baseline | 5.53 | 10.84 | 8.59 | | KuaiRec | TIGER | + ProtoAdapter | 5.69 | 11.26 | 8.98 | | KuaiRec | HSTU | Submitted direct baseline | 5.38 | 10.55 | 8.38 | | KuaiRec | HSTU | + ProtoAdapter | 5.55 | 10.98 | 8.87 | | KuaiRand | TIGER | Submitted direct baseline | 1.30 | 3.96 | 2.86 | | KuaiRand | TIGER | + ProtoAdapter | 1.46 | 4.37 | 3.07 | | KuaiRand | HSTU | Submitted direct baseline | 1.21 | 3.81 | 2.72 | | KuaiRand | HSTU | + ProtoAdapter | 1.36 | 4.22 | 2.97 |

Both additional backbones improve all three metrics over their respective baselines. The relative NDCG@5 gains are roughly 5%–9%, with variation across domains and backbones. Their absolute NDCG@5 remains below Qwen3+ProtoAdapter, while the same construction works across the different item-token interfaces. The revised paper now includes this table in the backbone-coverage experiment rather than limiting the response to an interface discussion.

I sincerely thank the authors for the careful and detailed response, as well as the additional experiments.

I still have one concern remaining. Regarding the response to Q1, the target saliency weight wsw_s appears to be the dominant factor in overall performance, which seems quite natural, whereas the other weights (wrw_r, wcw_c, wdw_d) seem to contribute relatively little. On both KuaiRec and KuaiRand, performance increases almost monotonically as wsw_s increases, but the component-level effect of each component weights is not fully explained. Since evidence selection is presented as one of the main contributions of the paper, I believe this is an important issue to clarify. In its current form, the configuration of the selector appears somewhat arbitrary, and the paper would be strengthened by a more principled justification of these design choices.

But overall, the authors have provided extensive additional experiments, which I really appreciate. Although the original submission lacked several important details and explanations, I believe the paper has improved substantially throughout the rebuttal process, with stronger evidence supporting the proposed method. Therefore, I am raising my score to 3.

Thank you very much for your careful review. Owing to the limited time available, we could not incorporate the relevant content during the rebuttal period; however, we consider your suggestions to be highly valuable. We will add the corresponding justification to the appendix of the paper later.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Shared supplementary results

We thank the reviewers for their careful reading and constructive suggestions. The shared tables below provide the evidence used across the four responses.

Common construction, training, and statistical protocol

All variants use the same OneRec-1.7B initialization, item-token mapping, training requests, test requests, and direct-decoding evaluator. Each constructed variant supplies its corresponding data in both CPT and SFT. The four attribution variants are:

  • Direct: history→answer training without a trace.
  • Prototype-only: retrieval uses history similarity only (alpha=1); the teacher writes the vertical pattern and transition bridge without selected event pointers or an observed-evidence slot.
  • Evidence-only: prototype retrieval is disabled; the teacher writes observed evidence and a transition bridge from selected current-history events.
  • Full: history/evidence retrieval with alpha=0.6, followed by the complete vertical-pattern, observed-evidence, and transition-bridge trace.

For controlled comparisons, the common accepted intersection contains full, prototype-only, evidence-only, generic-summary, and target-free traces. Every variant must pass filtering on a request for that request to enter the shared trace pool. This creates pools of 32,036 / 112,640 / 84,480 requests on All_Beauty / KuaiRec / KuaiRand. A rejection in any variant makes that request a direct-answer example in every variant, so request identities and the direct:trace ratio stay aligned. Raw full-method generation and rejection counts are reported separately.

The mean supervised trace length is 64 tokens under the 300-character cap, and the mean direct-answer supervision is 20.82 tokens. Trace variants use two CPT passes and one SFT pass. With these common pools, total supervised training tokens are 8.43M / 29.62M / 22.22M, with 2,058 / 7,232 / 5,424 optimizer updates, respectively. The final update in each run contains the shorter remainder batch. Training seeds 17, 43, and 91 are shared across variants. We report mean ± sample SD over training runs. Paired CIs use 10,000 request-level bootstrap resamples of the seed-averaged per-request difference. Low-data experiments additionally pair the data subsample across direct and ProtoAdapter within each seed. All NDCG@5 values below are percentages.

Table S1. Two-scale attribution with shared trace budgets

The three teacher-trace variants share accepted requests, trace-length distribution, direct:trace ratio, supervised tokens, and optimizer updates. The submitted Direct row is retained as the original reference. Direct-extended in the next table provides the separate strict token-and-update-matched direct control.

VariantAll_Beauty NDCG@5KuaiRec NDCG@5KuaiRand NDCG@5
Direct OneRec, submitted reference10.12 ± 0.089.19 ± 0.053.64 ± 0.03
Prototype-only10.39 ± 0.099.43 ± 0.063.78 ± 0.04
Evidence-only10.48 ± 0.089.51 ± 0.063.83 ± 0.04
Full ProtoAdapter10.76 ± 0.099.77 ± 0.063.98 ± 0.04

Both components improve over the submitted direct reference, and their combination gives the largest gain. The descriptive interaction full − prototype-only − evidence-only + direct is positive in all three domains, indicating a small additional benefit when the two evidence scales are combined. The Direct-extended comparison below separately checks the complete method under the exact same token and update budget.

AuthorsOpenReview ↗

Continued: Shared supplementary results

Table S2. Target-free construction and equal-token controls

Target-free construction removes target saliency, renormalizes the other three selector weights, and generates traces from observed histories without providing the label to the teacher. It is included in the common accepted intersection above, so its requests, trace count, direct:trace ratio, and training budget remain aligned with the other trace variants. Generic summaries use the same teacher, full history, and training label for consistency, without retrieved prototypes or selected evidence. Random filler contains domain-neutral text with no target identifiers or future events. Direct-extended repeats direct examples to the same total supervised-token and update budget as the trace variants.

VariantKuaiRec NDCG@5KuaiRand NDCG@5Mean trace tokensSupervised tokens, KuaiRec / third (M)
Direct, submitted schedule9.19 ± 0.053.64 ± 0.03018.65 / 13.99
Direct-extended9.243.67029.62 / 22.22
Generic teacher summary9.373.736429.62 / 22.22
Random/filler, matched length9.223.656429.62 / 22.22
Shuffled traces, matched budget9.243.606429.62 / 22.22
Raw patterns, matched budget8.933.586429.62 / 22.22
Target-free full construction9.58 ± 0.063.87 ± 0.046429.62 / 22.22
Full ProtoAdapter9.77 ± 0.063.98 ± 0.046429.62 / 22.22

Target-free construction remains above the direct baseline, generic summaries, and matched-length controls. Restoring full target-aware construction improves further. This difference includes both target-guided evidence selection and the teacher's use of the training label for consistency checking.

Table S3. Offline cost and direct-decoding inference

Offline construction is performed once per dataset. Teacher output tokens below refer to generated trace text before filtering; at 64 mean tokens per generated trace, they are consistent with the raw generation counts in the filtering table.

DatasetBank construction (A100-hours)Retrieval + selection (CPU-hours)Teacher output tokens (M)Teacher generation (A100-hours)Adaptation (A100-hours)
All_Beauty0.71.82.35.69.4
KuaiRec2.38.68.219.731.2
KuaiRand1.86.46.114.924.6

At deployment, ProtoAdapter directly decodes item tokens from the original history. The serving process loads only the adapted model and tokenizer; bank construction, selection, retrieval, and teacher generation remain offline. The table reports the pooled held-out evaluation over 2,048 requests per dataset, with 6,144 requests in total. OneRec and OneRec+ProtoAdapter have essentially the same serving latency and memory use.

Table S4. Seed-level values and paired statistics

Each cell lists seeds 17 / 43 / 91. The corresponding means and sample SDs are used in Tables S1 and S2.

VariantAll_BeautyKuaiRecKuaiRand
Full ProtoAdapter10.66 / 10.78 / 10.849.71 / 9.76 / 9.833.94 / 3.99 / 4.01
Target-free full10.46 / 10.55 / 10.649.51 / 9.59 / 9.633.83 / 3.87 / 3.91

For full minus the submitted Direct reference, the request-level paired CIs over seed-averaged predictions are +0.58 [0.48, 0.69] on KuaiRec and +0.34 [0.28, 0.41] on KuaiRand. The SD describes training variation, while the CI describes paired request variation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.