{"id":"0286a20a-121f-4729-8b46-67a3e014bfad","arxiv_id":"2504.16454","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"UniGRF trains one generative model to output both next-item predictions and click probabilities, and reports consistent gains over separate retrieval and ranking models on MovieLens and Amazon-Books.","lead":"A single generative Transformer can be trained to do both recommendation retrieval and ranking, treating each as a sequence generation task. UniGRF adds a ranking-driven enhancer and an adaptive loss weighter, and reports consistent gains over two-stage baselines on MovieLens-1M, MovieLens-20M, and Amazon-Books.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never specifies how the ranking head f_phi is applied to unseen candidate items at inference, so the reported AUC and the enhancer's hard-negative scores rest on an undefined scoring path.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the paper does not explain how f_phi, trained on contextualized history-position latents, is applied to arbitrary candidate items when computing ranking scores. This is not a minor implementation detail. The central claim is that a single generative model performs both retrieval and ranking while sharing information without extra cost. If candidates are scored by f_phi on raw item embeddings, then Eq. (3) violates the training distribution of Eq. (2) and the ranking scores are not user-contextualized. If candidates are scored by appending them to the sequence and running the Transformer, then the framework requires an extra forward pass per candidate, which changes the efficiency claim and should have been stated. The reported AUC in Table 3 and the enhancer module both depend on this unresolved choice. I do not see this as evidence of misconduct; the authors may well have a working protocol that they simply omitted. But the omission is load-bearing because it determines whether the empirical results support the paper's central claim. Secondary issues such as the unreported alpha, T, lambda_a, lambda_b, the lack of code, and the p-value caption without seeds are real but would not by themselves change the conditionality of the verdict. The paper presents broad, consistent experimental improvements and a plausible framework, so a conditional acceptance posture remains appropriate; clarifying the ranking inference protocol is the minimal requirement.","tokens_in":20285,"tokens_out":4063,"duration_ms":43746,"concrete_test":"Request the exact candidate-scoring protocol used for Table 3 and reproduce it. On MovieLens-1M, train exactly as in Eq. (2), then compute AUC under two protocols: (a) score each candidate by sigmoid(f_phi(e_item)) using raw item embeddings, as Eq. (3) literally states; (b) append each candidate to the user's sequence, run Transformer(e_u + candidate), and apply f_phi to the candidate's contextualized output. Also rebuild the hard-to-detect negative set under protocol (b) and rerun the enhancer ablation. If protocol (a) reproduces the reported AUC, Eqs. (2) and (3) are inconsistent. If only protocol (b) reproduces it, the 'no additional computational cost' claim must account for one extra forward pass per candidate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (2) trains f_phi on e_b'_k, the contextualized latents produced by Transformer(e_u) for items already in the user's history. The ranking stage is then supposed to score candidates from the candidate set (Sec. 3), and Table 3 reports AUC, but no inference-time procedure is given for computing an e_b' representation for candidate items that are not in the history. Eq. (3) computes score_ranking = sigmoid(f_phi(eS)) with eS described as 'the embedding of the negative sample set S,' which reads as raw item embeddings. If that is the implementation, f_phi is evaluated on inputs from a different distribution than Eq. (2) trained on; if eS instead means contextualized latents, the paper omits the additional forward passes and their cost, contradicting the 'minimal computational overhead' claim in Sec. 4.2. Either way, both the Table 3 ranking results and the enhancer's hard-negative construction depend on an unspecified candidate-scoring protocol. Because no code or data splits are released, nothing in the manuscript currently verifies that the same model can actually rank arbitrary candidates without extra cost or that the enhancer's scores are computed on the right representation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniGRF, a framework that unifies retrieval and ranking in a single autoregressive generative recommendation model. It treats next-item prediction (retrieval) and click prediction (ranking) as two output positions of one Transformer, trains with sampled softmax and binary cross-entropy losses, and adds a ranking-driven enhancer that generates hard negatives and potential-favorite pseudo-positives, plus a gradient-guided adaptive weighter. Experiments on MovieLens-1M, MovieLens-20M, and Amazon-Books instantiate the framework with HSTU and Llama and report consistent gains over baselines in retrieval metrics and AUC.","tokens_in":20510,"tokens_out":4092,"duration_ms":40708,"significance":"If the proposed inference protocol is well-defined, the paper offers a conceptually appealing unification of retrieval and ranking in generative recommenders, with a lightweight enhancer and an adaptive weighting scheme. Strengths include the model-agnostic design, the two-architecture instantiation, the ablation study, and the scaling-law analysis. However, the central ranking claim depends on an underspecified candidate-scoring procedure, and the enhancer introduces a self-training loop whose safeguards are not analyzed. The significance is therefore conditional on resolving these points.","major_comments":[{"comment":"The paper never specifies how the ranking head f_phi scores arbitrary candidate items at inference. Eq. (2) trains f_phi on contextualized latents e'_b_k produced by the Transformer for items in the user's history, but Eq. (3) applies f_phi to eS, described as 'the embedding of the negative sample set S,' and no test-time procedure is given for obtaining a contextualized representation for a candidate not in the history. As written, the reported AUC in Table 3 and the enhancer's hard-negative scores rest on an undefined or distributionally mismatched scoring path; please specify the exact inputs (raw embeddings vs. Transformer outputs), the additional forward passes if any, and reconcile the procedure with the claim of no additional computational cost.","section":"Sec. 4.1.2 and Sec. 4.2, Eqs. (2)-(3), Table 3"},{"comment":"The potential favorite generator relabels items as positive solely because the model's own ranking score exceeds the threshold alpha; because the same model (through shared parameters) then trains on these labels, the pseudo-labels are not independent of the model and can reinforce systematic errors. The paper provides no analysis of label precision, no comparison with alternatives such as using future interactions or a held-out oracle, and no sensitivity study for alpha; please add such analysis or explicitly justify why this self-training loop cannot lead to confirmation bias.","section":"Sec. 4.2.2"},{"comment":"The relative score defined in Eq. (5) can be negative when the ranking score exceeds the retrieval score, and the text's characterization of selecting 'high retrieval, low ranking' items is not guaranteed by taking the top-m of this quantity. Please clarify whether negative-scoring items can enter the hard-negative set, whether the formula's behavior for sigmoid-valued scores matches the intended meaning, and how the score is made stable when the ranking score is near zero.","section":"Sec. 4.2.1, Eq. (5)"}],"minor_comments":[{"comment":"The indexing in Eq. (1) and the retrieval loss appears inconsistent: the Transformer output is written as e_i'_{k+1} for the predicted next item, while L_retrieval is written in terms of e_i'_k; please align the notation.","section":"Sec. 4.1, Eq. (1)"},{"comment":"The abstract claims 'without additional computational costs,' while Sec. 4.2 states the overhead is 'minimal'; please reconcile these statements and state the exact computation added by the enhancer.","section":"Abstract and Sec. 4.2"},{"comment":"Table captions and column headers mix 'MovieLens-1M' and 'ML-1M' and state 'p-value < 0.05' without describing the test, the alternative hypothesis, or the number of runs; please clarify.","section":"Tables 2 and 3"},{"comment":"The parameter settings state that ranking-only models train for 20 epochs with early stopping while other models train for 100 epochs, but it is unclear whether early stopping is applied to the unified models and whether the same candidate protocol is used for the ranking AUC evaluation of all baselines.","section":"Sec. 5.1.4"},{"comment":"The module is called 'Gradient-Guided Adaptive Weighter' in the heading and 'weightier' in one sentence in Sec. 4.3; please use a single consistent name.","section":"Sec. 4.3"}],"recommendation":"major_revision","confidential_remarks":"The unification idea is timely and the experiments are broad, but the missing inference-time scoring protocol is load-bearing for the central ranking claim. Adding precise algorithmic details and pseudo-label safeguards, plus a reproducibility appendix with code or data splits, would make the contribution verifiable. The present manuscript is not ready for acceptance, but the issues appear addressable within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: if you work on generative recommender systems, read this for the framing and the adaptive weighter; don't cite the empirical numbers until they close the scoring gap.\n\nWhat's actually new here: UniGRF uses one autoregressive model with position-distinguished outputs—item positions predict the next item embedding (retrieval), behavior positions predict a click score (ranking). That is a sensible way to share parameters across stages, and they instantiate it on both HSTU and Llama to show it is model-agnostic. The gradient-guided adaptive weighter (Eqs. 6-10) is a neat practical trick: weight the two losses by recent loss-ratio to keep convergence speeds aligned. The enhancer's hard-negative mining via relative retrieval/ranking scores is a reasonable idea, though the \"potential favorites\" relabeling is riskier.\n\nWhere it gets wobbly: the paper never says how the ranking head is applied to arbitrary candidate items at inference. Eq. (2) trains f_phi on contextualized latents e_b'_k, but Eq. (3) computes ranking scores on eS, which reads as raw item embeddings. Those are different representations. If the implementation really feeds raw embeddings to f_phi, then the AUC and all enhancer scores are off-distribution. If it uses contextualized latents, the paper omits the extra forward passes and the cost. There is no code or data split to check which. This is not a minor clarity issue—it affects every ranking result and the enhancer's negative construction.\n\nThe enhancer also has a self-training flavor: items are relabeled positive purely because the model's own ranking score exceeds alpha. Without external validation or leakage analysis, the gains could be partly self-confirming. Hyperparameters alpha, T, lambda_a, lambda_b, and m are not fully reported for all datasets, so reproduction is impossible. The p-value caption is uninterpretable without seeds or error bars.\n\nWhat it does well: the experiments are wide, the ablations show the enhancer and weighter each matter, and the scaling-law figure is a nice bonus. The baseline set is appropriate, and self-citation is not egregious. There are two duplicate references ([40]/[41] and [72]/[73]) that should be cleaned up. The authors also honestly note that industrial scaling is still future work.\n\nBottom line: this is a solid framework idea that deserves a serious referee. The missing inference protocol and the pseudo-labeling are fixable—with those, it could be a useful contribution. For now, I would not cite the numbers.\n\nRecommendation: send to peer review, but require the authors to specify the candidate-scoring procedure and release code or at least the split and hyperparameters before acceptance.","headline":"A plausible unification of retrieval and ranking into a single generative model, but the undefined inference-time scoring path for the ranking head and the self-referential enhancer need fixing before the reported AUC can be trusted.","tokens_in":21093,"tokens_out":3858,"would_cite":false,"duration_ms":36123,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that retrieval and ranking in recommender systems can be unified inside one generative model, with a ranking-driven enhancer and a gradient-guided adaptive weighter, and that this design beats separately trained…","keywords":["generative recommendation","retrieval","ranking","unified framework","hard negative sampling","adaptive loss weighting","sequential recommendation","model-agnostic"],"falsifier":"Take UniGRF trained on MovieLens-1M, fix a held-out user, and for a candidate item that does not appear in that user's history, feed its embedding through the same ranking-scoring path used in Eq. (3) to obtain a score; then compute AUC over the full item pool. If the scoring head cannot be applied to unseen candidates, or if the AUC on those candidates is near chance while the reported AUC only uses history items, the claim that one shared generative model ranks arbitrary candidates is falsified; a complementary ablation on Amazon-Books, removing the potential-favorite relabeling and the hard-negative miner one at a time, would show whether the enhancement loop is genuinely joint.","tokens_in":20001,"feed_emoji":"🎯","tokens_out":10001,"duration_ms":90725,"temperature":0.7,"pith_summary":"Recommender systems usually split work into a retrieval stage that narrows the item pool and a ranking stage that scores the survivors, and information is lost in the handoff. This paper tries to eliminate that loss by making one autoregressive generative model do both jobs at once: for each user, the same transformer predicts the next item (retrieval) and the click probability for the item just seen (ranking), distinguished only by output position. To make the two stages help each other, the paper adds a ranking-driven enhancer that turns ranking scores into harder negative samples and relabels likely-favorite items as positive, and a gradient-guided adaptive weighter that balances the two task losses so neither stage converges at the expense of the other. On MovieLens-1M, MovieLens-20M, and Amazon-Books, the framework reports consistent gains over separately trained retrievers and rankers, with the largest improvements in ranking accuracy. Because the design only specifies a position convention and two loss adjustments, the authors argue it can wrap any autoregressive generative backbone, demonstrating it with HSTU and Llama.","feed_headline":"Single generative model unifies retrieval and ranking for recommenders","feed_subtitle":"Shared parameters plus hard-negative and loss-balancing modules beat separate two-stage systems on three public datasets.","key_machinery":"The load-bearing object is the position-split sequence formulation: one autoregressive transformer is fed alternating item and feedback embeddings, and its output positions are read twice—feedback positions feed a small ranking head $f_\\phi$ with binary cross-entropy, item positions feed sampled softmax retrieval against the item pool. Two mechanisms do the work of making the stages collaborate: the ranking-driven enhancer uses Eq. (5) to convert ranking scores into hard-to-detect disliked negatives and an $\\alpha$ threshold to relabel potential favorite items as positive, and the gradient-guided adaptive weighter computes the ratio of successive losses for each stage and softmax-weights the two losses so retrieval's slower convergence does not starve ranking. The whole argument turns on these mechanisms operating inside one shared-parameter model, so any retrieval improvement feeds ranking and any ranking precision sharpens retrieval's training signal.","core_discovery":"UniGRF's central claim is that retrieval and ranking need not be two models or two training runs: they can be two positions in one sequence-generation output. The user history is encoded as alternating item and interaction-type embeddings; a transformer emits contextualized vectors at every position, and the framework reads the item-position vectors as next-item predictions for retrieval and the feedback-position vectors as click scores for ranking. Retrieval is constrained by sampled softmax loss against the true next item and randomly sampled negatives, while ranking is a small head $f_\\phi$ trained with binary cross-entropy on the feedback vectors. The ranking-driven enhancer then computes a relative score $s_{retrieval} \\cdot (s_{retrieval}/s_{ranking} - 1)$ to mine hard-to-detect disliked items and flags items with ranking score above $\\alpha$ as potential favorites, relabeling them positive in the next epoch, so the ranking stage shapes the retrieval training data. The gradient-guided adaptive weighter measures convergence speed by the ratio of successive losses for each stage and sets softmax weights to slow or fast tasks accordingly. The reported result is that this one-model design outperforms separate HSTU/Llama retrievers and rankers and traditional cascade frameworks on all three datasets, with ranking AUC showing the clearest gains.","pith_inferences":["If the ranking head can be applied to arbitrary candidate embeddings at test time, the same scoring function could in principle rank the whole item pool, turning the retriever's candidate generation into an optional shortcut rather than a necessity; the paper does not explore this.","The relative-score mining rule in Eq. (5) is a generic recipe: any pair of tasks where a precise scorer and a broad scorer share a model could use the same ratio to mine hard examples, so the mechanism may transfer beyond retrieval and ranking to prerank/rerank or other cascade settings.","Because the adaptive weighter uses only successive loss ratios, it is a crude proxy for optimization speed; a direct gradient-norm monitor would be a natural robustness check and could behave differently on noisy or small batches.","The sparse-dataset pattern suggests a testable prediction: the potential-favorite relabeling component (the $\\alpha$ threshold) should matter more on Amazon-Books than on MovieLens, since noisy negatives are more likely in sparse regimes; ablating the two enhancer components separately would reveal which module drives the reported gain."],"forward_implications":["One training run produces a retriever and a ranker from the same parameters, so serving both stages costs roughly half the time and space of two separately trained generative models.","Ranking accuracy improves more than retrieval accuracy, which matters because the ranking stage directly determines the final list the user sees.","Sparse datasets benefit most: Amazon-Books shows the largest gains, suggesting cross-stage information transfer is most valuable when per-user evidence is thin.","The framework inherits scaling-law behavior: increasing transformer layers lowers both losses and improves both retrieval and ranking metrics.","Cascade frameworks that glue separately trained retrieval and ranking models generally underperform the unified model, and on generative backbones they can even degrade performance."],"supporting_citations":[{"why":"Supplies the HSTU generative architecture and the sequence-generation formulation of retrieval and ranking that UniGRF unifies and instantiates.","marker":"[67]"},{"why":"Supplies the Llama transformer backbone used to show UniGRF is model-agnostic.","marker":"[45]"},{"why":"Supplies the sampled softmax loss that constrains the retrieval stage in UniGRF.","marker":"[22]"},{"why":"Provides the SASRec self-attentive sequential baseline for retrieval comparisons.","marker":"[23]"},{"why":"CoRR is a cascade retriever-ranker baseline based on KL distillation that UniGRF must outperform.","marker":"[20]"},{"why":"RankFlow is a joint-optimization cascade baseline that UniGRF compares against.","marker":"[34]"},{"why":"Defines the AUC metric and the DIN ranking baseline used to evaluate ranking performance.","marker":"[75]"}],"fun_headline_variants":["One generative model for retrieval and ranking in recommenders","Single generative model unifies two-stage recommender pipeline","Generative recommender: one model, both retrieval and ranking","UniGRF: unified generative model beats two-stage recommenders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"At the heart of the method is the assumption that the ranking head $f_\\phi$, trained on contextualized vectors for items that already appear in the user's history, can also score arbitrary candidate items from the pool in the same way; the paper never specifies how candidates are fed into the model at test time when computing AUC, so if that transfer fails the reported ranking gains do not follow from the stated design.","fun_headline_variants_meta":{"raw":{"variants":["One generative model for retrieval and ranking in recommenders","Single generative model unifies two-stage recommender pipeline","Generative recommender: one model, both retrieval and ranking","UniGRF: unified generative model beats two-stage recommenders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000339,"raw_usage":{"total_tokens":1908,"prompt_tokens":1015,"completion_tokens":893,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":825}},"tokens_in":631,"tokens_out":893,"duration_ms":8408,"temperature":1.0,"reasoning_tokens":825,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:03:28.143482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take UniGRF trained on MovieLens-1M, fix a held-out user, and for a candidate item that does not appear in that user's history, feed its embedding through the same ranking-scoring path used in Eq. (3) to obtain a score; then compute AUC over the full item pool. If the scoring head cannot be applied to unseen candidates, or if the AUC on those candidates is near chance while the reported AUC only uses history items, the claim that one shared generative model ranks arbitrary candidates is falsified; a complementary ablation on Amazon-Books, removing the potential-favorite relabeling and the hard-negative miner one at a time, would show whether the enhancement loop is genuinely joint.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RankFlow is a joint-optimization cascade baseline that UniGRF compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the AUC metric and the DIN ranking baseline used to evaluate ranking performance."}],"review_version":1}