{"id":"b128bf2b-9703-4199-a3b2-5ec511a8cf49","arxiv_id":"2412.08911","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A recommender model conditioned on a multi-dimensional goal vector of future click and purchase rewards outperforms multi-objective baselines on two e-commerce datasets with lower computational cost.","lead":"The paper trains a recommender model to predict the next item a user will choose, conditioning on a vector of desired future rewards for multiple objectives such as clicks and purchases. It reports that this goal-conditioned approach outperforms standard multi-task and offline RL baselines on two e-commerce datasets while using fewer parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Denoising claim is the load-bearing explanation for MOGCSL's gains, but it is only tested on a synthetic dataset whose noise is generated from the goals, not on real data.","rationale":"The reader's weakest_assumption identifies the denoising premise, and I agree. The central empirical comparisons are large and consistent, and no obvious flaw in the MOGCSL training/inference setup would invalidate the reported HR numbers. However, the paper presents denoising as a core contribution and as the mechanism behind the gains. The only direct evidence is synthetic and circular: the noise process in Eq. (20) enforces the very goal-noise correlation that Section 3.4 assumes. The real datasets have sparse purchase goals (mean about 0.2), so the existence of a meaningful 'high-goal, low-noise' regime is not established. Thus the load-bearing assumption is not merely a missing ablation; it is the explanatory bridge between vector-goal conditioning and the observed improvements. A real-data noise-proxy correlation check would settle it. If the check fails, the paper remains conditionally acceptable as an empirical method, but the denoising and mechanism claims must be removed or revised. Since the reader already conditioned on this and no independent flaw changes the verdict, I keep the verdict unchanged.","tokens_in":16893,"tokens_out":10309,"duration_ms":120889,"concrete_test":"On RetailRocket and Challenge15, construct a real-data noise proxy: for each interaction, mark the action as noisy if it is inconsistent with the user's long-term category distribution in the same session (or if the session ends immediately after it despite a click). Compute the rank correlation between the multi-objective goal g_t and the clean-action indicator, and evaluate MOGCSL-S separately on test actions stratified by goal quantile. If the correlation is near zero, or if high-goal conditioning does not improve predictive accuracy more on low-noise actions than on high-noise actions, then the Section 3.4 denoising premise is unsupported on real data and the paper must weaken the denoising claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 frames MOGCSL's central advantage as discounting noisy interactions, and Appendix B.2 is offered as empirical validation. But Eq. (20) defines the noisy label as uniformly random exactly when any goal dimension is ≤ -1, i.e., the very correlation the section assumes ('interactions achieving higher goals are generally less noisy') is hard-coded into the data generator. The experiment then shows that an XGBoost classifier given g can use that programmed correlation. It does not test the proposed transformer/GCSL training, and it cannot falsify the denoising claim. On RetailRocket and Challenge15 no noise labels exist and no proxy analysis links higher cumulative rewards to lower noise; the paper itself notes the mean purchase goal is only about 0.2, so 'high' purchase goals are extremely sparse. Consequently the mechanism offered for the large purchase-HR improvements (48.67 to 65.43 and 39.06 to 56.82) is an unverified premise. If the premise fails, MOGCSL could still be a useful goal-conditioned predictor, but the paper's headline contribution of 'first to reveal... mitigate harmful effects of noisy instances' and the design justification for multi-dimensional goals would not be supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MOGCSL, a goal-conditioned supervised learning method for multi-objective sequential recommendation. The method relabels each interaction with a vector of cumulative future rewards (Eq. 1), conditions a transformer-based next-action predictor on that vector, and trains with a standard cross-entropy loss (Eq. 5). For inference, goals are chosen either by scaling the training-set mean reward or by a CVAE-based algorithm that models the distribution of achievable goals (Algorithm 2, Eqs. 6-7). Experiments on RetailRocket and Challenge15 report large gains on purchase-oriented HR/NDCG over multi-task, Pareto, and offline-RL baselines, while using a 9.1M-parameter model with the shortest training time in Table 2. A theorem about the distribution of achievable goals is stated in Section 3.3, and a denoising mechanism is claimed in Section 3.4 and Appendix B.2.","tokens_in":17104,"tokens_out":7064,"duration_ms":73428,"significance":"The empirical core of the paper is valuable: the comparison uses the same transformer/self-attention base module across eight multi-objective baselines plus SQN and MOPRL, reports means and standard deviations over five seeds, and the purchase-metric gains are consistent on two public datasets. If these results hold, MOGCSL is an attractive simple alternative to multi-task towers and weighted-loss optimization for recommendation, and the complexity comparison in Table 2 supports the scalability claim. The paper also gives a candid discussion of when simple statistical goal selection is competitive with the CVAE-based method. However, the theoretical analysis in Appendix A and the denoising validation in Appendix B.2 are not at the same standard as the experimental core, and these items directly support the paper's novelty claims. The real-data experiments do not depend on the synthetic construction, so the central empirical result remains plausible, but the mechanism story needs to be substantially strengthened or reframed.","major_comments":[{"comment":"The synthetic denoising experiment is circular with respect to the claim in Section 3.4. Equation (20) defines the noisy label ε(a) as a deterministic function of the goal vector: whenever any goal dimension gi ≤ −1, ε(a) is uniformly random, and otherwise ε(a)=a. The goal vector is then given as an input feature to XGBoost-mg, and at inference every goal dimension is set to 1. The experiment therefore only shows that a supervised model that receives g can learn the programmed threshold; it cannot confirm that MOGCSL's conditioning on cumulative rewards discounts noise in real recommender data. In addition, the experiment uses an XGBoost classifier rather than the transformer/GCSL training described in Section 3.2, so it does not test the proposed method itself. Because the abstract and Section 3.4 present denoising as a central benefit, this evidence needs to be replaced with a test that does not bake in the assumed correlation, or the claim should be substantially softened.","section":"Appendix B.2, Eq. (20)"},{"comment":"The proof of Theorem 1 contains an unjustified marginalization. Equation (16) writes P(rn|sn) as a sum over an, gn of 1(R(sn,an)=rn)π(an|sn,gn)P(sn,gn)/P(sn), but the correct conditional would involve P(an,gn|sn) or P(gn|sn); there is no reason for P(sn,gn)/P(sn) to appear in that form. Equation (17) also conditions P(sm+1|sm,xn−1) on xn−1, which contains rewards after time m, yet the right-hand side uses only rm; conditioning on future rewards changes the transition distribution, so the displayed factorization does not follow. Since Theorem 1 is used to motivate the CVAE goal-selection algorithm in Section 3.3, the proof must be corrected or the theoretical claim should be removed or weakened.","section":"Appendix A, Eqs. (15)-(17)"},{"comment":"The proposed CVAE-based goal-selection algorithm does not consistently outperform the simple statistical strategy in the primary experiments. In Table 4, MOGCSL-C is worse than MOGCSL-S on Challenge15 purchase HR@10 (55.13 vs 56.82) and only marginally different on RetailRocket. The paper acknowledges this and attributes it to low average goals, but the goal-selection algorithm is listed as a key contribution and is the only component that depends on Theorem 1. The Tenrec result in Appendix B.8 is helpful, but the current main-dataset results do not establish the practical value of the CVAE-based method; either the contribution should be scaled back or the analysis should be integrated more directly with the main experimental claims.","section":"Section 4.4, Table 4"},{"comment":"The denoising explanation is not verified on either real dataset. The paper never measures noise on RetailRocket or Challenge15, and Section 4.4 reports that the mean purchase goal is only about 0.2, so 'high' purchase goals are extremely sparse. Without a real-data proxy linking higher cumulative rewards to lower label noise, the large purchase-metric gains (e.g., HR@10 from 48.67 for Share-PE to 65.43 for MOGCSL on RetailRocket) are not evidence for the noise-discounting mechanism; they may reflect other properties of goal-conditioned training, such as additional input features or differences in the training objective. The mechanism should be presented as a hypothesis, or supported by a real-data analysis that gives direct evidence for the assumed correlation.","section":"Section 3.4 and Section 4.4"}],"minor_comments":[{"comment":"Equation (3) appears to show only the goal embedding being fed into the self-attention block, whereas the text says the state embedding is fed in; please clarify the notation and the exact input to Atten(·).","section":"Section 3.2, Eq. (3)"},{"comment":"The variational objectives are written with a plus sign before the KL term (L = E[log P] + D_KL); the standard ELBO uses a minus sign, or the quantity is a negative ELBO. Please check the sign and state explicitly whether these expressions are to be maximized or minimized.","section":"Section 3.3, Eqs. (6)-(7)"},{"comment":"There are several typos and grammatical slips: 'Intialization' in Algorithm 1, 'We also introduces' in the abstract, 'am MOMDP' in Appendix A, and 'mutli-objective' in Section 3.1. A proofreading pass is needed.","section":"Algorithm 1 and Abstract"},{"comment":"The units in Table 2 should be defined: 'Ks' presumably means kiloseconds, but it is not explained; also '3.0Ks' and similar entries would be clearer as '3.0×10^3 s' or '50 min'.","section":"Table 2"},{"comment":"The text states that MOGCSL achieves the best performance on click HR on RetailRocket, but on Challenge15 the click metrics are not consistently best; the wording should be adjusted to avoid overgeneralizing the click results.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The strongest part of the paper is the controlled real-data evaluation, which supports the practical usefulness of goal-conditioned multi-objective prediction. The weakest parts are the mismatch between the bold denoising and theoretical claims and the evidence provided: the synthetic denoising experiment is circular, the proof of Theorem 1 has gaps, and the CVAE goal-choosing algorithm is not consistently better than the simple baseline. A revision that either supplies real-data evidence for the denoising mechanism or explicitly reframes it as a hypothesis, and that repairs or de-emphasizes the theorem, would make the paper publishable. No concerns about attribution or overlap were evident."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the empirical result is real and worth engaging with; the denoising and theory claims need to be trimmed to what the evidence supports.\n\nWhat's actually new: representing the GCSL goal as a vector of per-objective cumulative rewards, trained with a transformer encoder and plain cross-entropy. On RetailRocket and Challenge15 it beats scalar-goal GCSL (MOPRL) and the multi-objective baselines on purchase HR/NDCG by a wide margin, with a 9.1M-parameter model and the shortest training time. The comparison is well controlled: same base encoder, five seeds, sensible baselines including Shared-Bottom/MMOE with DWA, PE, and FAMO. I buy the main claim.\n\nThe honest surprises in the paper are also useful: the CVAE-based goal chooser does not beat a simple mean-scaled goal on the two main datasets, and the authors say so. They show high-goal training data is sparse and follow up on Tenrec where the advanced chooser does win. That is a fair, non-hyped treatment.\n\nSoft spots. The denoising story is the weakest part. Section 3.4 assumes higher goals mean less noisy actions, and the synthetic experiment in Appendix B.2 builds that correlation in by defining noise as a function of the goal (Eq. 20). So it cannot validate the mechanism on real data. The paper itself reports mean purchase goal around 0.2, which makes high-goal examples scarce. The gains on purchase metrics may come from something else—maybe the vector conditioning itself, or the goal input acting as a useful feature. The paper doesn't need the denoising story to stand; the empirical comparison carries the contribution. But the abstract's 'first to reveal' and the filtering claim go beyond what is shown. Theorem 1 is a Markov-property restatement with a proof that has an unjustified marginal over (sn, gn) in Eq. (16); I'd downgrade it to a lemma or drop it.\n\nAlso no code is released, which makes the strong results harder to check, and the paper's own tables show MOGCSL's click NDCG is not always best—fine, but worth acknowledging.\n\nWho it's for: anyone working on multi-objective learning or sequential recommendation. It deserves a serious referee: the empirical design is solid, the claims are mostly commensurate with the evidence, and the weaknesses are fixable by revision. I'd accept it conditionally.","headline":"A genuinely strong empirical comparison for vector-goal GCSL in multi-objective recommendation, but the denoising story is circular and should be cut or reframed.","tokens_in":17668,"tokens_out":2250,"would_cite":true,"duration_ms":21757,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Conditioning a recommender on a vector of future rewards outperforms weighted multi-objective baselines.","keywords":["multi-objective recommendation","goal-conditioned supervised learning","sequential recommendation","cumulative reward relabeling","denoising","variational autoencoder goal selection","transformer recommender","offline learning"],"falsifier":"On a dataset where high-goal interactions are generated by noisy or atypical behavior (for example, accidental clicks that lead to long sessions) while low-goal interactions are deliberate, train MOGCSL and compare with a uniform-weight baseline; if goal-conditioned training does not improve or worsens prediction, the denoising premise fails. More directly, measure a label-noise proxy, such as repeat-action or re-click consistency, as a function of cumulative goal in RetailRocket and Challenge15; the claimed mechanism requires higher-goal steps to be measurably cleaner.","tokens_in":16639,"feed_emoji":"🛒","tokens_out":4338,"duration_ms":45557,"temperature":0.7,"pith_summary":"The paper proposes MOGCSL, a training scheme for multi-objective sequential recommendation that conditions next-item prediction on a vector of cumulative future rewards, one dimension per objective. It argues that this turns multi-objective learning into plain supervised learning, removing the need for separate task towers, loss-weight tuning, or Pareto optimization. The method also implicitly discounts noisy interactions: because it learns to imitate actions that led to high long-term goals, it concentrates on what the authors argue are cleaner demonstrations of user interest. On two e-commerce datasets, purchase hit rate at rank 10 improves markedly over previous multi-objective baselines, with fewer parameters and shorter training time. A second contribution is a variational-autoencoder goal-selection procedure for inference, which the paper shows is most useful when training data contains many high-goal sessions.","feed_headline":"Vector goals lift recommendation hits from 48.7 to 65.4","feed_subtitle":"Conditioning next-item prediction on a vector of future click and purchase rewards beats multi-task towers and weighted losses, using…","key_machinery":"The central object is the goal vector $g_t = \\sum_{t'=t}^{|\\tau|} r_{t'}$, the cumulative multi-objective reward from timestep $t$ to the end of the session, used both as a training relabeling and as a conditioning input. The model concatenates a transformer-encoded interaction history with a timestep embedding and a learned embedding of the goal vector, passes the result through a self-attention block, and outputs action logits over the item set; training is standard cross-entropy on the observed next action given state and goal. A conditional VAE learns the distribution of achievable goals for a new state and selects a high, achievable goal as the inference input. The mechanism that carries the argument is the goal's role as a filter: actions followed by low cumulative rewards are implicitly discounted because the model is trained to imitate only actions consistent with the input goal, and high goals at inference activate the clean-behavior patterns.","core_discovery":"The central claim is that multi-objective sequential recommendation can be solved by relabeling each interaction with a multi-dimensional goal, defined as the vector of cumulative rewards from that step to the end of the session, and training a transformer to predict the observed next action conditioned on that goal. Because the model is trained to imitate actions that actually led to high goals, it preferentially learns clean behavior patterns; at inference, feeding a high goal steers it toward high-reward actions. The paper reports purchase HR@10 rising from 48.67 with the best weighted-loss baseline to 65.43 on RetailRocket and from 39.06 to 56.82 on Challenge15, using a 9.1M-parameter model and the shortest training time in the comparison. It also proves that the distribution of goals actually achieved during inference is determined by the initial state, the input goal, and the policy, which motivates the proposed CVAE-based goal chooser.","pith_inferences":["An implication not drawn in the paper is that the denoising mechanism depends on a monotone relationship between cumulative reward and data quality; if real logs contain high-goal noise (for example, bot-driven sessions), the method could amplify rather than discount noise.","The CVAE goal-chooser's weaker performance on low-average-goal datasets suggests a testable extension: augment training data with synthetic high-goal trajectories or use goal relabeling so the model sees enough high-goal examples to make advanced goal selection worthwhile.","Theorem 1 implies the goal chooser must be retrained whenever the policy changes; otherwise the selected goals may not be achievable under the new policy, a practical constraint the paper does not discuss.","The comparison between MOGCSL-S and MOGCSL-C indicates that simple statistics on training goals can be near-optimal for sparse-goal data, so production systems with tight latency budgets can safely start with the simpler strategy."],"forward_implications":["MOGCSL removes the need for separate towers, dynamic loss weights, or Pareto-optimization loops in multi-objective recommendation; a single cross-entropy loss on goal-relabeled data suffices.","Because the goal is a vector, objectives are disentangled without scalarization weights, and changing priorities at inference reduces to choosing a different goal vector rather than retraining weights.","On datasets with high average cumulative goals, the learned CVAE goal-chooser outperforms simple statistical goal selection, implying that richer goal distributions unlock the benefit of principled inference goals.","The denoising analysis implies that MOGCSL should be most beneficial when log data contains many low-reward interactions from distracted or mismatched users, and least beneficial when high- and low-goal interactions are equally informative.","The same vector-goal conditioning could apply to any sequential decision problem with multiple reward dimensions, provided the offline data contains trajectories with varying long-term outcomes."],"supporting_citations":[{"why":"Supplies the goal-conditioned supervised learning paradigm that MOGCSL extends from scalar goals to multi-dimensional vectors.","marker":"[8]"},{"why":"Provides the PRL baseline and the statistical goal-selection rule that MOGCSL-S adopts for inference.","marker":"[11]"},{"why":"Provides the SQN offline-RL baseline and the dataset preprocessing conventions used in the experiments.","marker":"[28]"},{"why":"Provides the Shared-Bottom and MMOE multi-task architectures that serve as primary baselines.","marker":"[4]"},{"why":"Provides the FAMO baseline for balanced multi-objective optimization and a comparison point for efficiency.","marker":"[13]"},{"why":"Documents the common practice of setting inference goals as a multiple of training-set statistics, which the paper compares against.","marker":"[10]"},{"why":"Supplies the Tenrec dataset with high average goals used to validate the advanced CVAE goal-chooser.","marker":"[31]"}],"fun_headline_variants":["Vector goals power multi-objective recs to 65% hit rate","Goal-conditioned learning beats multi-task for recs","MOGCSL lifts purchase hits from 48.7 to 65.4","Vector goal relabeling cleans noisy recommender data","Multi-objective recs via goal-conditioned imitation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that interactions followed by higher cumulative multi-objective rewards are systematically less noisy and more representative of true user interest; the paper encodes this in a synthetic experiment but does not verify the correlation on the real e-commerce datasets.","fun_headline_variants_meta":{"raw":{"variants":["Vector goals power multi-objective recs to 65% hit rate","Goal-conditioned learning beats multi-task for recs","MOGCSL lifts purchase hits from 48.7 to 65.4","Vector goal relabeling cleans noisy recommender data","Multi-objective recs via goal-conditioned imitation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000534,"raw_usage":{"total_tokens":2600,"prompt_tokens":1010,"completion_tokens":1590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":1503}},"tokens_in":626,"tokens_out":1590,"duration_ms":10951,"temperature":1.0,"reasoning_tokens":1503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:26:24.141908+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a dataset where high-goal interactions are generated by noisy or atypical behavior (for example, accidental clicks that lead to long sessions) while low-goal interactions are deliberate, train MOGCSL and compare with a uniform-weight baseline; if goal-conditioned training does not improve or worsens prediction, the denoising premise fails. More directly, measure a label-noise proxy, such as repeat-action or re-click consistency, as a function of cumulative goal in RetailRocket and Challenge15; the claimed mechanism requires higher-goal steps to be measurably cleaner.","supporting_citations":[{"cited_title":"Rethinking reinforcement learning for recommendation: A prompt per- spective","cited_arxiv_id":null,"evidence_quote":"Provides the PRL baseline and the statistical goal-selection rule that MOGCSL-S adopts for inference."},{"cited_title":"Self-supervised reinforcement learning for recommender systems","cited_arxiv_id":null,"evidence_quote":"Provides the SQN offline-RL baseline and the dataset preprocessing conventions used in the experiments."},{"cited_title":"Modeling task relationships in multi-task learning with multi-gate mixture-of-experts","cited_arxiv_id":null,"evidence_quote":"Provides the Shared-Bottom and MMOE multi-task architectures that serve as primary baselines."},{"cited_title":"Famo: Fast adaptive multitask optimization","cited_arxiv_id":null,"evidence_quote":"Provides the FAMO baseline for balanced multi-objective optimization and a comparison point for efficiency."},{"cited_title":"Tenrec: A large-scale multipurpose benchmark dataset for rec- ommender systems","cited_arxiv_id":null,"evidence_quote":"Supplies the Tenrec dataset with high average goals used to validate the advanced CVAE goal-chooser."}],"review_version":1}