acceptodds
Under review as a conference paper at ICLR 2027

TDE: Target-conditioned Decision Evidence for Generative Recommendation

Abstract

Generative recommendation (GR) autoregressively generates semantic token sequences for the next item, but its generation objective specifies what target to generate without explicitly supervising which historical behaviors support that target. We introduce Target-conditioned Decision Evidence (TDE), a new supervision framework that converts target-conditioned decision explanations into representation learning signals. Given an observed training target, TDE uses a reference model to identify which historical behaviors contribute to generating that specific target. We adapt Transformer attribution to autoregressive encoder-decoder recommendation by defining target scores over generated target tokens and tracing target-to-history contributions through decoder-to-encoder cross-attention. The resulting positive decision evidence is normalized into a soft distribution over historical behaviors, from which two independently perturbed evidence-grounded views supervise a trainable recommender through a contrastive objective. At inference time, TDE introduces no target input: the trained recommender performs standard autoregressive generation from the history alone. Experiments show that TDE consistently improves GRs. Controlled comparisons further show that its gains cannot be explained by heuristic augmentation, generic contrastive learning, target-supervised sequence pairing, or uniform token weighting. Additional analyses reveal that more target-relevant attribution provides more effective supervision, that evidence-guided augmentation and dual-view grounding play complementary roles, and that evidence transfer depends on alignment between the attribution pathway and the backbone’s decision objective. TDE establishes an explanation-to-supervision paradigm for GRs: learning not merely from user history, but from the historical evidence that supports a specific target decision.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Generative Recommendation, Target-conditioned Decision Evidence, Decision Evidence Supervision, Transformer Attribution, Evidence-Grounded Representation Learning

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_9YsyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TDE to turn attribution from a frozen reference model into a representation-learning signal. For an observed training target, it defines a score over the generated target tokens and traces contributions from the target to the history through feature gradients and decoder--encoder cross-attention. Positive attributions become weights over historical tokens. Two independently perturbed versions of these weights produce views for contrastive training. At inference, the recommender generates from history alone. On four Amazon datasets, the paper reports gains for two generative backbones and compares TDE with heuristic augmentation, generic contrastive learning, same-target pairing, and uniform weighting. It also studies attribution choices, evidence views, and transfer across architectures.

Critical strengths
  1. The target-conditioned supervision is well motivated. The generation loss tells the model which item to predict, but not which past interactions matter for that prediction; different targets may draw on different parts of the same history. TDE turns that distinction into a training signal. I see this as the paper's strongest motivation.

  2. The paper gives a concrete path from attribution to supervision. It defines a score over the generated target tokens, traces contributions through decoder--encoder cross-attention, and uses two evidence-grounded views to train representations rather than stopping at a post-hoc explanation. Comparisons with heuristic augmentation, generic contrastive learning, same-target pairing, and uniform weights make the value of this design more credible.

Critical weaknesses
  1. The paper does not independently test attribution faithfulness. Target-dependent weights and better recommendation scores do not show that high-weight interactions actually support the target prediction. Since the reference and student start from the same backbone, the auxiliary loss may also reinforce mistaken reference-model preferences. This leaves the central “decision evidence” interpretation unconvincing.

  2. The advance over prior work is unclear. Explanation-guided contrastive learning for sequential recommendation has been studied, for example in EC4SRec (CIKM 2022; arXiv:2209.01347). Adapting attribution to generated target tokens is a concrete contribution, but the paper does not establish what this adaptation adds over earlier explanation-guided methods. This limits my assessment of its novelty.

Questions for authors
  1. Implementation clarification: Did the experiments use the sum normalization in Eq. (8) or the Softmax in Algorithm 1? If Softmax was used, zero-attribution tokens still receive positive weight. How should the claim of retaining only positive evidence be understood?

  2. Attribution validity: Beyond downstream ranking metrics, how do you establish that high-weight history items support the reference model's target prediction? Could you compare target-score changes after perturbing high- and low-attribution items, and test whether the gains survive mismatching weights with training targets?

  3. Negative evidence: What is the intuition for discarding negative attribution? Is the signed ablation in Fig. 3d comparable in normalization and view construction? Could negative attribution carry useful opposing evidence?

  4. Novelty: Relative to explanation-guided contrastive sequential recommendation such as EC4SRec (CIKM 2022; arXiv:2209.01347), what capability or gain specifically comes from target-token attribution and the two evidence-grounded views?

  5. Scope: TDE hurts BERT4Rec on Baby and Sports in Table 4. Under which architectures or training conditions should it help? Although inference is unchanged, does the training cost in Table 10 limit practical use?

4 comments
AuthorsOpenReview ↗

Q1/Q3: Normalization and Negative; W1/Q2: Faithfulness

Q1/Q3: Attribution Normalization and Negative Evidence

We investigate attribution normalization and evidence polarity on Amazon Tools.

MethodAttribution WeightR@5R@10N@5N@10
ActionPiece (AP)—0.02680.04440.01710.0229
(1) TDE-SoftmaxReLU (Algo.1)Softmax(ReLU(α))0.03240.04990.02160.0272
(2) TDE-SumReLU (Eq.8)ReLU(αᵢ) / Σⱼ ReLU(αⱼ)0.03140.04930.02060.0264
(3) TDE-SoftmaxSignedSoftmax(α)0.03200.04770.02180.0269
(4) TDE-AbsL1abs(αᵢ) / Σⱼ abs(αⱼ)0.03110.04810.02040.0259
(5) TDE-SignedL1αᵢ / Σⱼ abs(αⱼ)0.02910.04900.01930.0257
  1. Normalization robustness.
  • Both SoftmaxReLU and SumReLU are effective.
  • We confirm that the original TDE uses SoftmaxReLU as specified in Algo. 1 (temperature=1.0). We will fix Eq. 8.
  1. Negative evidence.
  • Direct signed pooling is less effective. Both SoftmaxReLU and SumReLU outperform AbsL1 and SignedL1.
  • Negative attribution might still contain useful information. We will fix the claim of retaining only positive-support tokens.

W1/Q2 Part I: Attribution Faithfulness

We conduct two complementary intervention experiments on Amazon Tools, using the same frozen reference model:

  • (1) Deletion, which removes selected historical items, and
  • (2) Sufficiency, which retains only selected items.

We compare four strategies: High (highest attribution), Low (lowest positive attribution), Random, and Recent (most recent interactions). All interventions are performed at the item level, followed by re-tokenization.

ΔlogP denotes target log-probability drops relative to the full history.

High-attribution items are more influential

Table: Deletion faithfulness: higher values indicate greater prediction degradation.

Deleted historyDelete-HighDelete-LowDelete-RandomDelete-Recent
ΔlogP @10% deletion0.08720.03500.02150.0699
ΔlogP @20% deletion0.10060.04100.02730.0835
ΔlogP @30% deletion0.13910.06490.04140.1100
Rank worsening @20% deletion254.8100.688.8169.8
  • Removing high-attribution items consistently causes the largest target log-probability degradation across all deletion ratios. At 20% deletion, High produces 3.68× larger logP degradation than Random, and 2.87× greater exact catalog-rank worsening.
  • Removing high-attribution items also exceeds Recent, indicating that attribution captures more than recency effects.

High-attribution items preserve predictive information

Table: Evidence sufficiency: lower is better.

Retained historyRetain-HighRetain-LowRetain-RandomRetain-Recent
ΔlogP 20% retaining0.08710.30420.18520.2528
ΔlogP 40% retaining-0.01880.24140.10050.1903
ΔlogP 60% retaining-0.04470.18580.06490.1459
ΔlogP 80% retaining-0.04270.11600.02520.1175
  • Retaining High consistently yields the smallest logP loss. With only 20% of historical items retained, High reduces the loss by 53.0% relative to Random.
  • Retaining the top 40%–80% even slightly improves target logP over the full history, suggesting that removing less relevant interactions may reduce interference.
AuthorsOpenReview ↗

W1/Q2 Part II: Target–Evidence Mismatch; W2/Q4: Novelty and Comparison with EC4SRec

W1/Q2 Part II: Target–Evidence Mismatch

We conduct a WrongTarget ablation, following the target-substitution setting in Fig. 1 of our paper. Specifically, we keep the same historical sequence and original near-future training target, but replace the attribution evidence with that extracted for a temporally later target from the same user's interaction sequence. WrongTarget evidence is constructed exclusively from training interactions. The recommendation objective remains unchanged, while the attribution evidence is conditioned on a different target. The dual-view construction and contrastive learning objective remain unchanged.

  1. Results of target–evidence mismatch on Amazon Office.
MethodR@5R@10N@5N@10
ActionPiece0.05250.08350.03430.0446
TDE-WrongTarget0.05280.08690.03430.0452
TDE0.05770.09000.03760.0479
  1. Results of target–evidence mismatch on Amazon Tools.
MethodR@5R@10N@5N@10
ActionPiece0.02680.04440.01710.0229
TDE-WrongTarget0.03050.04900.02020.0262
TDE0.03240.04990.02160.0272
  1. Observations.
  • TDE consistently outperforms TDE-WrongTarget. Mismatching the target evidence reduces the gains.
  • Together with Fig. 1 in our paper, these results provide complementary evidence. Fig. 1 demonstrates that attribution changes with the prediction target, while this WrongTarget ablation demonstrates that correctly aligning this attribution with the training target is important for effective representation supervision.

W2/Q4: Novelty and Comparison with EC4SRec

We implement EC4SRec-style learning under the same ActionPiece backbone and training budget. Following EC4SRec, this baseline uses importance-guided cropping, masking, and reordering to construct augmented histories, together with its contrastive objective EC4SRec(SSL). Results on Amazon Office and Tools are as follows.

  1. Comparing TDE with EC4SRec-style learning on Office.
MethodR@5R@10N@5N@10
ActionPiece0.05250.08350.03430.0446
AP+CL4SRec0.05220.08620.03270.0437
AP+EC4SRec(SSL)0.04950.08540.03320.0447
TDE0.05770.09000.03760.0479
Improvement (%)+16.57+5.39+13.25+7.16
  1. Comparing TDE with EC4SRec-style learning on Tools.
MethodR@5R@10N@5N@10
ActionPiece0.02680.04440.01710.0229
AP+CL4SRec0.02880.04630.01970.0253
AP+EC4SRec(SSL)0.03020.04740.02030.0258
TDE0.03240.04990.02160.0272
Improvement (%)+7.28+5.27+6.40+5.43
  1. Observations.
  • TDE consistently outperforms AP+EC4SRec across all eight metrics on both datasets.
  • TDE is effective beyond explanation-guided discrete sequence augmentation.
  1. Novelty of TDE comparing with EC4SRec.
  • TDE explicitly attributes the target token generation in autoregressive generative recommendation to historical interactions.
  • TDE extracts fine-grained, token-level, target-conditioned decision evidence by tracing the decoder-to-encoder cross-attention pathway.
  • TDE transforms decision explanations into representation-level supervision by constructing two evidence-weighted contrastive views, beyond directly modifying historical sequences through explanation-guided augmentation.
AuthorsOpenReview ↗

Q5: architectures; Training Cost

Q5 Part I: Under which architectures should TDE help?

In fact, we have explicitly investigated this issue in Section 5.5 (RQ5), Appendix E, and Table 18.

Empirical evidence across architectures

Table 18 (Appendix E) summarizes the transferability of TDE across homogeneous and heterogeneous architectures.

Table: Average improvements are computed over four datasets and four evaluation metrics, as reported in Table 18.

Evidence TransferArchitectureDecision-Pathway AlignmentObjective AlignmentAvg. Improvement
ActionPiece → ActionPieceEncoder–DecoderStrongStrong+10.7%
GenPAS → GenPASEncoder–DecoderStrongStrong+11.2%
SASRec → SASRecDecoder-onlyStrongStrong+31.2%
BERT4Rec → BERT4RecEncoder-onlyWeak / EntangledWeak / Potential conflict-2.9%
ActionPiece → SASRecEncoder–Decoder → Decoder-onlyStrongStrong+34.3%

TDE benefits both encoder–decoder (ActionPiece, GenPAS) and decoder-only (SASRec) architectures, and even heterogeneous ActionPiece → SASRec.

When should TDE help? A twofold alignment principle

As discussed in Section 5.5 and Appendix E, the effective transfer depends on two complementary conditions.

  • Decision-evidence alignment: The reference model should expose a meaningful target-to-history decision pathway, allowing attribution to identify historical evidence supporting the target. ActionPiece provides decoder–encoder cross-attention, while SASRec provides causal self-attention at the prediction position.
  • Representation-objective alignment: Evidence-guided contrastive supervision should be compatible with the trainable backbone's recommendation objective. ActionPiece and SASRec both optimize future-target prediction, making their representation objectives naturally aligned with TDE's target-conditioned evidence supervision.

Why does BERT4Rec behave differently? Appendix E provides explanations

  • BERT4Rec's bidirectional self-attention produces globally contextualized historical representations, making individual historical contributions less directly attributable (attribution entanglement).
  • Its Cloze-style masked-item reconstruction objective differs from the future-target prediction objective underlying TDE, potentially creating conflicting representational preferences.
  • Consistent with this weaker alignment, TDE improves BERT4Rec on Office and Tools for NDCG metrics but degrades performance on Baby and Sports (Table 4). These observations are consistent with our proposed alignment hypothesis.

Q5 Part II: Training Cost and Practicality

We report the complete training time (seconds), including backbone pretraining (Stage I) and continuation training (Stage II).

Note, Shared Stage I Pretraining:

  • Stage I is performed only once for each backbone and dataset, rather than repeated separately for the baseline and TDE. Specifically, we select the pretrained checkpoint with the best validation performance and use this same checkpoint to initialize both AP-continued and TDE-AP for their respective Stage II continuation training.
  • The same procedure applies to GenPAS-continued and TDE-GenPAS.
DatasetMethodStage I (s)Stage II (s)Total (s)Change
OfficeAP-continued3,4555,0308,485—
OfficeTDE-AP3,4557,67611,131+31.18%
OfficeGenPAS-continued4,39012,60016,990—
OfficeTDE-GenPAS4,3907,16011,550-32.02%
ToolsAP-continued12,46212,46224,924—
ToolsTDE-AP12,46219,83832,300+29.59%
ToolsGenPAS-continued12,48017,43029,910—
ToolsTDE-GenPAS12,4807,14019,620-34.40%
  • ActionPiece bacbone: Compared with AP-continued, TDE-AP increases total training time on the two datasests.
  • GenPAS bacbone: Compared with GenPAS-continued, TDE-GenPAS reduces the total training time on the two datasests. Its Stage II continuation time is reduced.
  • Inference efficiency: TDE uses the frozen reference model and evidence-guided contrastive learning exclusively during training. At inference, these components are removed, and the trained backbone performs standard recommendation without additional computational overhead.
AuthorsOpenReview ↗

Rebuttal follow-up.

Thank you again for your valuable time and comments on our paper. As the discussion period ends soon, we would like to know whether our responses and revised PDF address your concerns. If you have any further questions or suggestions, we would be very happy to discuss them. Many thanks!

Reviewer_Ef48OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TDE, a training framework for generative recommendation. A frozen reference model provides target-conditioned attribution over historical tokens. Two perturbed versions of the attribution weights are used to pool the trainable model's representations for contrastive learning. The additional branch is removed at inference. Experiments use full-ranking evaluation on four Amazon datasets and examine several backbones, attribution methods, and contrastive-learning controls.

Critical strengths
  1. This paper introduces target-conditioned decision evidence as a supervision signal for generative recommendation. Using target-specific historical attribution to guide representation learning offers an interesting perspective on improving recommendation models.

  2. The proposed framework adapts Transformer attribution to autoregressive encoder-decoder models and incorporates it into contrastive learning. The additional components are used only during training, so inference cost is unchanged.

  3. Experiments on four datasets show consistent improvements on ActionPiece and GenPAS. Comparisons with heuristic augmentation, generic and target-supervised contrastive learning, and uniform weighting help identify the contribution of the proposed evidence supervision.

Critical weaknesses
  1. The proposed method directly uses the reference model's attributions as supervision. However, these attributions may contain noise or bias and do not necessarily reflect the actual contribution of historical behaviors to the prediction target.

  2. The method relies on a fixed reference model without systematically studying how reference quality affects the results. The reference model itself may be unreliable, and its errors and biases may be passed to the trainable model.

  3. Both contrastive views are constructed by randomly dropping weights from the same evidence distribution. When evidence is concentrated on a few historical behaviors, the views may be nearly identical or lose the key supporting evidence, which may weaken the contrastive signal.

Questions for authors
  1. Does removing highly weighted historical items affect the target ranking more than removing low-weight items?

  2. How do the gains change when using reference models at different training stages? Could the authors also compare instances on which the reference model makes correct and incorrect predictions?

9 comments
AuthorsOpenReview ↗

W1: Target–Evidence Mismatch; W1/Q1: Attribution Faithfulness

W1: Target–Evidence Mismatch

We conduct a WrongTarget ablation, following the target-substitution setting in Fig. 1 of our paper. Specifically, we keep the same historical sequence and original near-future training target, but replace the attribution evidence with that extracted for a temporally later target from the same user's interaction sequence. WrongTarget evidence is constructed exclusively from training interactions. The recommendation objective remains unchanged, while the attribution evidence is conditioned on a different target. The dual-view construction and contrastive learning objective remain unchanged.

  1. Results of target–evidence mismatch on Amazon Office.
MethodR@5R@10N@5N@10
ActionPiece0.05250.08350.03430.0446
TDE-WrongTarget0.05280.08690.03430.0452
TDE0.05770.09000.03760.0479
  1. Results of target–evidence mismatch on Amazon Tools.
MethodR@5R@10N@5N@10
ActionPiece0.02680.04440.01710.0229
TDE-WrongTarget0.03050.04900.02020.0262
TDE0.03240.04990.02160.0272
  1. Observations.
  • TDE consistently outperforms TDE-WrongTarget. Mismatching the target evidence reduces the gains.
  • Together with Fig. 1 in our paper, these results provide complementary evidence. Fig. 1 demonstrates that attribution changes with the prediction target, while this WrongTarget ablation demonstrates that correctly aligning this attribution with the training target is important for effective representation supervision.

W1/Q1: Attribution Faithfulness

We conduct two complementary intervention experiments on Amazon Tools, using the same frozen reference model:

  • (1) Deletion, which removes selected historical items, and
  • (2) Sufficiency, which retains only selected items.

We compare four strategies: High (highest attribution), Low (lowest positive attribution), Random, and Recent (most recent interactions). All interventions are performed at the item level, followed by re-tokenization.

ΔlogP denotes target log-probability drops relative to the full history.

High-attribution items are more influential

Table: Deletion faithfulness: higher values indicate greater prediction degradation.

Deleted historyDelete-HighDelete-LowDelete-RandomDelete-Recent
ΔlogP @10% deletion0.08720.03500.02150.0699
ΔlogP @20% deletion0.10060.04100.02730.0835
ΔlogP @30% deletion0.13910.06490.04140.1100
Rank worsening @20% deletion254.8100.688.8169.8
  • Removing high-attribution items consistently causes the largest target log-probability degradation across all deletion ratios. At 20% deletion, High produces 3.68× larger logP degradation than Random, and 2.87× greater exact catalog-rank worsening.
  • Removing high-attribution items also exceeds Recent, indicating that attribution captures more than recency effects.

High-attribution items preserve predictive information

Table: Evidence sufficiency: lower is better.

Retained historyRetain-HighRetain-LowRetain-RandomRetain-Recent
ΔlogP 20% retaining0.08710.30420.18520.2528
ΔlogP 40% retaining-0.01880.24140.10050.1903
ΔlogP 60% retaining-0.04470.18580.06490.1459
ΔlogP 80% retaining-0.04270.11600.02520.1175
  • Retaining High consistently yields the smallest logP loss. With only 20% of historical items retained, High reduces the loss by 53.0% relative to Random.
  • Retaining the top 40%–80% even slightly improves target logP over the full history, suggesting that removing less relevant interactions may reduce interference.
AuthorsOpenReview ↗

W2/Q2: Reference Quality and Training Stages

W2/Q2: Reference Model Quality and Training Stages

We use four reference checkpoints at different training stages. The maximum training budget is 280 epochs (following ActionPiece setting) with early stopping and validation-based checkpoint selection. We reserve the epoch 280 checkpoint beyond the validation optimum on the Office dataset for this ablation. AP (continued) denotes the ActionPiece-continued (seed 2026); see the response to the reviewer aYw3's W2/Q2 Part II: Multi-Seed Reliability of the ActionPiece Backbone.

  1. Results of reference model quality on Amazon Office.

Table: Reference model performance/quality (selection on validation set)

Reference checkpointR@5R@10N@5N@10
Early (epoch 40)0.03360.05690.02100.0285
Middle (epoch 60)0.04280.07340.02620.0360
Best (epoch 160)0.05340.08480.03470.0448
Last (epoch 280)0.04590.07240.03010.0385

Table: TDE Student performance (test set)

MethodR@5R@10N@5N@10
AP (pretrained)0.05250.08350.03430.0446
AP (continued)0.05120.08690.03210.0435
TDE (w/ Early Ref.)0.05260.08640.03510.0459
TDE (w/ Middle Ref.)0.05630.08850.03680.0473
TDE (w/ Best Ref.)0.05770.09000.03760.0479
TDE (w/ Last Ref.)0.05140.08360.03400.0443
  1. Results of reference model quality on Amazon Tools.

Table: Reference model performance/quality (selection on validation set)

Reference checkpointR@5R@10N@5N@10
Early (epoch 70)0.02850.04410.01790.0230
Middle (epoch 140)0.02860.04540.01910.0233
Best = Last (epoch 280)0.03160.05010.02090.0269

Table: TDE Student performance (test set)

MethodR@5R@10N@5N@10
AP (pretrained)0.02680.04440.01710.0229
AP (continued)0.02950.04800.02010.0261
TDE (w/ Early Ref.)0.02880.04680.01920.0249
TDE (w/ Middle Ref.)0.02920.04740.01910.0250
TDE (w/ Best=Last Ref.)0.03240.04990.02160.0272
  1. Observations.
  • TDE generally benefits from higher-quality reference models.
  • TDE can benefit from imperfect reference models, suggesting some tolerance to imperfect reference supervision.
AuthorsOpenReview ↗

Q2: Reference-Correct vs. Reference-Incorrect

Q2: Reference-Correct vs. Reference-Incorrect Instances

We explicitly stratify test instances into reference-correct and reference-incorrect groups based on the ground-truth target's ranking under the frozen reference model. Specifically, we define a reference prediction as Correct@p% if the ground-truth target ranks within the top (p%) of the full item catalog, and Incorrect@p% otherwise, where p∈{1,5,10}.p \in \{1,5,10\}.

AP-Cont. denotes the ActionPiece-continued (seed 2026); see the response to the reviewer aYw3's W2/Q2 Part II: Multi-Seed Reliability of the ActionPiece Backbone.

  1. Results of reference-correct and reference-incorrect instances on Office (2,421 items; 4,905 test instances).
ThresholdReference Group#InstancesAP-Cont. R@5TDE R@5AP-Cont. R@10TDE R@10AP-Cont. N@5TDE N@5AP-Cont. N@10TDE N@10
Top 1% (≤25)Correct5240.32060.25190.45040.36830.22290.17690.26460.2142
Top 1% (≤25)Incorrect4,3810.02350.02940.04790.05770.01430.01780.02210.0269
Top 5% (≤121)Correct1,3780.16760.15240.26270.24460.11220.09970.14270.1291
Top 5% (≤121)Incorrect3,5270.01130.01450.02380.03090.00710.00950.01110.0147
Top 10% (≤242)Correct2,0120.12670.11530.20480.19230.08420.07610.10910.1007
Top 10% (≤242)Incorrect2,8930.00550.01000.01180.02040.00350.00610.00560.0095
  1. Results of reference-correct and reference-incorrect instances on Tools (10,218 items; 16,638 test instances).
ThresholdReference Group#InstancesAP-Cont. R@5TDE R@5AP-Cont. R@10TDE R@10AP-Cont. N@5TDE N@5AP-Cont. N@10TDE N@10
Top 1% (≤103)Correct8880.19710.15650.26010.22970.14360.12180.16420.1456
Top 1% (≤103)Incorrect15,7500.01930.02370.03290.03960.01200.01470.01640.0198
Top 5% (≤511)Correct2,0600.13010.10100.17770.15190.09020.07420.10570.0908
Top 5% (≤511)Incorrect14,5780.01450.02090.02630.03530.00900.01280.01280.0175
Top 10% (≤1022)Correct3,0680.10270.08050.14240.12190.07030.05720.08320.0706
Top 10% (≤1022)Incorrect13,5700.01210.01960.02300.03350.00750.01210.01100.0166
  1. Observations.
  • Across all three correctness thresholds on both datasets, TDE consistently outperforms AP-Cont. on reference-incorrect instances.
  • TDE can still learn useful representations from instances where the reference model fails to rank the ground-truth target highly.
  • Improvements are concentrated on reference-hard instances.
Reviewer_aYw3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TDE uses a frozen pretrained recommender to extract target-conditioned attribution over historical tokens and turns positive attribution into two stochastic, evidence-weighted representations for contrastive training. The target is used only during training; inference retains the original backbone. Experiments on four Amazon domains show improvements with ActionPiece and GenPAS, with controls separating evidence weighting from augmentation, generic contrastive learning, and target-based positive pairing. Additional experiments examine attribution components, reference freezing, and transfer across architectures.

Critical strengths
  1. The contribution is a useful way to turn decision attribution into training supervision for generative recommendation. It complements the existing generation objective without adding inference cost. Improvements across all four domains with both ActionPiece and the stronger GenPAS framework (Tables 2–3) make the approach practically relevant beyond a single backbone.

  2. The controlled experiments provide unusually helpful evidence about the source of the gains. Comparisons with AP+CL4SRec, AP+TSCL, uniform weighting, and partial versus dual grounding (Fig. 3 and Figs. 5–6) address plausible alternative explanations. The attribution and frozen-reference ablations further support the design, and the mixed BERT4Rec results are reported candidly rather than suggesting universal transfer.

Critical weaknesses
  1. The central evidence distribution is specified inconsistently: Eq. (8) normalizes ReLU scores by their sum, whereas Algorithm 1 applies Softmax to those scores. These operations differ in concentration and in whether nonpositive evidence receives zero weight, directly affecting view construction and the uniform-weight control. Code and checkpoints are promised upon acceptance (p. 10), but no repository is provided for checking the implementation. Resolving this discrepancy and documenting zero-mass handling would address a concrete reproducibility concern.

  2. Tables 2–5 and the key controlled comparisons report point estimates without the number of runs or variability. This particularly limits confidence in the small GenPAS gains on Office (1.40–2.95% in Table 3) and the incremental differences between evidence variants. Matched-seed repetitions of GenPAS/TDE and the main CL controls on Office and Tools would establish how reliably these differences hold.

Questions for authors
  1. Which normalization was used for the reported results: Eq. (8) or Algorithm 1? If Softmax was used, was attribution rescaled or temperature-adjusted? How are examples with no positive evidence, or views whose positive evidence is entirely dropped, handled? This clarification determines the reproducible definition of TDE.

  2. Were the ActionPiece/GenPAS baselines and RQ2 controls continued from the same pretrained checkpoint with a matched update budget, optimizer/scheduler initialization, and validation rule? Please specify the pretraining and continuation schedules. This would strengthen the attribution of gains to the added supervision.

5 comments
AuthorsOpenReview ↗

W1/Q1 Normalization and Negative; Availability and Reliability

W1/Q1 Part I: Attribution Normalization and Negative Evidence

We investigate attribution normalization and evidence polarity on Amazon Tools.

MethodAttribution WeightR@5R@10N@5N@10
ActionPiece (AP)—0.02680.04440.01710.0229
(1) TDE-SoftmaxReLU (Algo.1)Softmax(ReLU(α))0.03240.04990.02160.0272
(2) TDE-SumReLU (Eq.8)ReLU(αᵢ) / Σⱼ ReLU(αⱼ)0.03140.04930.02060.0264
(3) TDE-SoftmaxSignedSoftmax(α)0.03200.04770.02180.0269
(4) TDE-AbsL1abs(αᵢ) / Σⱼ abs(αⱼ)0.03110.04810.02040.0259
(5) TDE-SignedL1αᵢ / Σⱼ abs(αⱼ)0.02910.04900.01930.0257
  1. Normalization robustness.
  • Both SoftmaxReLU and SumReLU are effective.
  • We confirm that the original TDE uses SoftmaxReLU as specified in Algo. 1 (temperature=1.0). We will fix Eq. 8.
  1. Negative evidence.
  • Direct signed pooling is less effective. Both SoftmaxReLU and SumReLU outperform AbsL1 and SignedL1.
  • Negative attribution might still contain useful information. We will fix the claim of retaining only positive-support tokens.

W1/Q1 Part II: Evidence Availability and Dual-view Reliability

Zero-evidence and dropout fallback handling

We conduct validation using all 16,638 real test instances from Amazon Tools, with attribution scores extracted from the frozen ActionPiece reference model. We also performed independent dual-view dropout simulations and synthetic edge-case unit tests.

Table: Boundary-case frequencies and handling.

Boundary caseObserved frequencyHandling
(1) No positive raw attribution5.76%Softmax yields uniform weights
(2) SumReLU: zero positive mass5.76%Valid-token uniform fallback
(3) SoftmaxReLU: zero weight mass0%Not possible for nonempty valid histories
(4) At least one view fully dropped (p=0.2p=0.2)0%Restore original normalized weights
(5) Both views fully dropped (p=0.2p=0.2)0%Same fallback per empty view
(6) At least one view fully dropped (p=0.3p=0.3)0.16%Restore original normalized weights
(7) Both views fully dropped (p=0.3p=0.3)0%Same fallback per empty view
  • No positive attribution. We found that 5.76% of real instances contain no strictly positive raw attribution. Under the implemented SoftmaxReLU normalization, all valid tokens then receive uniform weights, so the resulting distribution remains well-defined. For SumReLU, the implementation uses valid-token uniform fallback.
  • Complete dropout. For each real instance, we simulate one pair of independently dropped views at p=0.2p=0.2 and p=0.3p=0.3. No complete-dropout events are observed at p=0.2p=0.2, while 0.16% of view pairs contained at least one completely dropped view at p=0.3p=0.3. When a view loses all valid tokens, the implementation restores the original normalized weights, reapplies the validity mask, and renormalizes.

Dual-view reliability under evidence dropout

We conduct a Monte Carlo analysis on 2,000 Amazon Tools test instances, using attribution weights extracted from the frozen ActionPiece reference model. Each instance undergoes 100 independent dual-view dropout simulations with p=0.2p=0.2. Instances are divided into three quantile groups based on evidence concentration.

Table: Dual-view reliability across evidence-concentration groups.

GroupTop-1 retained in both viewsAll-valid-token dropoutDual-view cosineRandom-pair cosine
High63.82%0.0180%0.79440.5909
Medium63.70%0.0075%0.79520.5798
Low64.30%0.0060%0.79610.5812
  • The probability of retaining the highest-weight token in both views closely matches the theoretical value (1−p)2(1-p)^2=64.0%. The probability of retaining it in exactly one view is 2p(1−p)2p(1-p)=32.0%, indicating stochastic differences in key-token inclusion.
  • The probability that at least one view drops all NuN_u valid tokens is 2pNu−p2Nu2p^{N_u}-p^{2N_u}. Empirically, this event occurs in only 0.0060%–0.0180%. When it occurs, our implementation restores the original normalized attribution weights and renormalizes them over valid positions, ensuring finite, nonnegative, unit-sum view weights.
  • The mean cosine similarity between independently dropped and renormalized weight vectors is approximately 0.79, compared with 0.58 for randomly paired vectors. The two views are neither identical nor unrelated.
AuthorsOpenReview ↗

W2/Q2 Multi-Seed

W2/Q2 Part I: Multi-Seed Reliability of the GenPAS Backbone

We conduct three matched-seed runs (2024, 2025, and 2026), comparing GenPAS-continued(the pretrained GenPAS backbone with continued training under the same protocol as TDE) against TDE-GenPAS. Both methods are initialized from the same GenPAS checkpoint, selected based on the best validation performance during Stage I pretraining.

Both methods follow the same two-stage training protocol. In each stage (pretraining and continuation), training is capped at 320,000 optimization steps, with validation performed every 400 steps and early stopping patience set to 10 validation evaluations. During continuation, both methods use matched random seeds and identical training budgets.

Table: Three individual seeds and mean ± standard deviation on Office.

SeedMetricGenPAS-continuedTDE-GenPAS
2024Recall@50.08620.0938
2024Recall@100.12990.1360
2024NDCG@50.05840.0653
2024NDCG@100.07230.0788
2025Recall@50.08770.0885
2025Recall@100.12800.1293
2025NDCG@50.05930.0609
2025NDCG@100.07270.0733
2026Recall@50.08730.0930
2026Recall@100.12330.1284
2026NDCG@50.06080.0614
2026NDCG@100.07390.0746
Mean ± StdRecall@50.0871 ± 0.00080.0918 ± 0.0029
Mean ± StdRecall@100.1271 ± 0.00340.1312 ± 0.0042
Mean ± StdNDCG@50.0595 ± 0.00120.0625 ± 0.0024
Mean ± StdNDCG@100.0730 ± 0.00080.0756 ± 0.0029
  • TDE-GenPAS achieves higher mean performance across all four metrics compared with GenPAS-continued.

W2/Q2 Part II: Multi-Seed Reliability of the ActionPiece Backbone

We conduct three matched-seed runs (2024, 2025, 2026), comparing ActionPiece-continued (the pretrained ActionPiece backbone with continued training like TDE, max-epochs 280+280) against TDE-AP.

  1. Results of three individual seeds and mean ± standard deviation on Office.
SeedMetricAP-continuedTDE-AP
2024R@50.05120.0577
2024R@100.08010.0900
2024N@50.03430.0376
2024N@100.04360.0479
2025R@50.04950.0557
2025R@100.08560.0909
2025N@50.03360.0357
2025N@100.04520.0475
2026R@50.05120.0565
2026R@100.08690.0917
2026N@50.03210.0366
2026N@100.04350.0469
Mean ± StdR@50.0506 ± 0.00100.0566 ± 0.0010
Mean ± StdR@100.0842 ± 0.00360.0909 ± 0.0009
Mean ± StdN@50.0333 ± 0.00110.0366 ± 0.0010
Mean ± StdN@100.0441 ± 0.00100.0474 ± 0.0005
  1. Results of three individual seeds and mean ± standard deviation on Tools.
SeedMetricAP-continuedTDE-AP
2024R@50.03070.0324
2024R@100.04890.0499
2024N@50.02080.0216
2024N@100.02670.0272
2025R@50.03170.0319
2025R@100.04960.0503
2025N@50.02080.0214
2025N@100.02660.0270
2026R@50.02950.0316
2026R@100.04800.0495
2026N@50.02010.0211
2026N@100.02610.0271
Mean ± StdR@50.0306 ± 0.00110.0320 ± 0.0004
Mean ± StdR@100.0488 ± 0.00080.0499 ± 0.0004
Mean ± StdN@50.0206 ± 0.00040.0214 ± 0.0003
Mean ± StdN@100.0265 ± 0.00030.0271 ± 0.0001
  1. Observations.
  • TDE-AP consistently outperforms AP-continued across all three seeds and all four metrics on both datasets.
  • TDE-AP exhibits lower standard deviations than AP-continued on most metrics.
AuthorsOpenReview ↗

W2/Q2: Matched-Seed CL Controls and Pretraining/Continuation Protocol

W2/Q2 Part III: Matched-Seed CL Controls and Pretraining/Continuation Protocol

Matched-seed CL controls

We repeat AP+CL4SRec, TDE-Uniform, and TDE using three matched random seeds (2024, 2025, 2026). All methods start from the same best pretrained ActionPiece checkpoint and follow the same continuation training protocol.

  1. Results of matched-seed CL controls on Amazon Office.
SeedMetricAP+CL4SRecTDE-UniformTDE
2024R@50.05320.05240.0577
2024R@100.08710.08810.0900
2024N@50.03510.03460.0376
2024N@100.04600.04610.0479
2025R@50.04910.04970.0557
2025R@100.08300.08400.0909
2025N@50.03130.03240.0357
2025N@100.04220.04330.0475
2026R@50.04670.05100.0565
2026R@100.08260.08440.0917
2026N@50.02860.03400.0366
2026N@100.04010.04470.0469
Mean ± StdR@50.0497 ± 0.00330.0510 ± 0.00130.0566 ± 0.0010
Mean ± StdR@100.0842 ± 0.00250.0855 ± 0.00220.0909 ± 0.0009
Mean ± StdN@50.0317 ± 0.00330.0337 ± 0.00110.0366 ± 0.0010
Mean ± StdN@100.0428 ± 0.00300.0447 ± 0.00140.0474 ± 0.0005
  1. Results of matched-seed CL controls on Amazon Tools.
SeedMetricAP+CL4SRecTDE-UniformTDE
2024R@50.02970.03100.0324
2024R@100.04870.04740.0499
2024N@50.01940.02050.0216
2024N@100.02550.02550.0272
2025R@50.03050.03020.0319
2025R@100.04780.04580.0503
2025N@50.01970.02080.0214
2025N@100.02530.02580.0270
2026R@50.03050.02980.0316
2026R@100.04820.04640.0495
2026N@50.01990.02010.0211
2026N@100.02560.02550.0271
Mean ± StdR@50.0302 ± 0.00050.0303 ± 0.00060.0320 ± 0.0004
Mean ± StdR@100.0482 ± 0.00050.0465 ± 0.00080.0499 ± 0.0004
Mean ± StdN@50.0197 ± 0.00030.0205 ± 0.00040.0214 ± 0.0003
Mean ± StdN@100.0255 ± 0.00020.0256 ± 0.00020.0271 ± 0.0001
  1. Observations.
  • TDE outperforms both AP+CL4SRec and TDE-Uniform on all four metrics across all three seeds.

Pretraining and continuation schedules

  1. As illustrated in Fig. 2 and described in Sec. 5 (Experimental Setup), our experiments follow a unified two-stage training protocol (details described in Algo. 1):
  • Stage I — Backbone pretraining: We pretrain ActionPiece for up to 280 epochs, selecting the checkpoint with the highest validation NDCG@10 on the validation set. The selected pretrained checkpoints correspond to epoch 280 on Tools and epoch 160 on Office (see Table 10 in the paper).
  • Stage II — Matched continuation: All compared methods, including AP-continued, AP+CL4SRec, TDE-Uniform, and TDE, initialize from the same best pretrained checkpoint learned in Stage I. Each method is allowed up to 280 additional continuation epochs, subject to the same early-stopping criterion (patience=20).
  1. Optimizer/scheduler initialization: Continuation inherits the optimizer state (including momentum) and learning-rate scheduler state from pretraining. The same initialization policy applies to all methods.

  2. Comparisons use identical pretrained starting points (learned in Stage I), matched maximum continuation budgets (max-epoch 280), consistent optimizer/scheduler state handling, and identical validation/early-stopping rules (patience=20).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal and revised PDF

Dear (Senior) Area Chair and reviewers,

Thank you for the helpful reviews and constructive discussions. Your comments clarified the paper and strengthened the analysis.

We have substantially strengthened TDE through new faithfulness tests, controls, multi-seed experiments, and clarifications, now incorporated into the revised manuscript (changes in blue).

Strengths recognized by reviewers

Reviewer aYw3 described TDE as a "useful way to turn decision attribution into training supervision" and praised the controlled comparisons.

Reviewer Ef48 recognized its "interesting perspective" and lack of inference overhead.

Reviewer 9Ysy called the target-conditioned motivation the paper's "strongest motivation." Our core idea is turning decision explanations into training-only supervision.

Major concerns addressed

1. Attribution faithfulness (Ef48, 9Ysy). We added item-level deletion and sufficiency interventions against Low, Random, and Recent controls. On Tools, deleting the top 20% high-attribution items drops target log-probability by 0.1006, compared with 0.0273 for random deletion (3.68×). Keeping only the top 20% yields a drop of 0.0871, versus 0.1852 for random retention. These directly validate attribution faithfulness for the reference decision.

2. Target-specific supervision and reference quality (Ef48, 9Ysy). Our WrongTarget control changes only the target used to extract evidence while holding the recommendation objective fixed. TDE outperforms WrongTarget in R@5: 0.0577 vs. 0.0528 (Office) and 0.0324 vs. 0.0305 (Tools). Early/Middle/Best/Last and correctness-stratified evaluations reveal that TDE can still improve reference-hard instances even from weaker references.

3. Normalization and reproducibility (aYw3, 9Ysy). We verified that the reported method uses Softmax⁡(ReLU⁡(α))\operatorname{Softmax}(\operatorname{ReLU}(\alpha)) with temperature 1.0, corrected the equation, and clarified nonpositive attribution handling. Ablations additionally test SumReLU and signed alternatives. Among 16,638 Tools instances, 5.76% have no positive raw attribution; these receive uniform valid-token weights. We documented dropout fallbacks and matched continuation/optimizer/validation protocols.

4. Multi-seed reliability and strong controls (aYw3). We added three matched seeds (2024–2026) for ActionPiece (Office/Tools), GenPAS (Office), and CL controls. TDE also exceeds AP+CL4SRec and TDE-Uniform in every metric and seed on both datasets.

5. Novelty over EC4SRec (9Ysy). We added a matched-backbone EC4SRec-style SSL adaptation and expanded Related Work. Unlike explanation-guided sequence augmentation, TDE traces autoregressive target-token generation through decoder–encoder cross-attention and constructs evidence-weighted representation views. Its R@5 exceeds the adapted baseline: 0.0577 vs. 0.0495 (Office), 0.0324 vs. 0.0302 (Tools); it wins on all eight reported metrics.

6. View stability, scope, and efficiency (Ef48, 9Ysy). Across evidence-concentration groups, top-evidence retention in both views is 63.7%–64.3% (theoretical 64.0%), with view cosine similarity near 0.79 versus 0.58 for random pairs. We clarified architectural alignment (including BERT4Rec limitations) and full training costs. No additional inference cost is introduced.

Reviewer Ef48's follow-up and score reconsideration

We especially appreciate Ef48's explicit acknowledgment that our new experiments addressed some original questions. The reviewer listed three remaining issues. We have replied to all three: (i) reconstructed group-weighted results and explained discrepancies from independent stochastic five-view inference; (ii) acknowledged weak-reference limits and added reference-hard analyses; (iii) addressed near-uniform Softmax scores with matched-mask TDE-versus-Uniform encoder-gradient diagnostics over 14 checkpoints, including zero-difference null controls.

Overall, the rebuttal now provides independent faithfulness evidence, target-mismatch tests, reference-quality boundaries, matched-seed reliability, and a direct prior-work comparison. We believe the central concerns are addressed, strengthening TDE's explanation-to-supervision contribution.

Many thanks!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.