{"id":"a1c22978-d483-4731-b711-00816390fb67","arxiv_id":"2502.09375","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"FARM combines discrete Fourier frequency splitting with contrastive preference alignment to improve cross-domain live-streaming recommendation at Kuaishou.","lead":"This paper presents FARM, a recommendation model that uses Fourier transforms to split users' behavior histories into low- and high-frequency parts and blends them with cross-domain attention. The model is deployed at Kuaishou and reports small offline and online gains in live-streaming recommendation metrics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DFT frequency indexes variation along sequence positions, not behavior rarity; the §2.1/3.2 mapping of sparse behaviors to high-frequency components is asserted rather than derived, so the paper's central mechanism is unsupported unless a direct test ties high-frequency bands to like/comment/gift…","rationale":"I agree with the reader's weakest-assumption analysis and sharpen it: the DFT index is about positional variation, not behavior rarity, and an impulse-like rare event is broadband. The paper's deployed-system evidence and online A/B gains are real, and the ablations confirm that the added modules contribute; I am not calling the empirical results fraudulent or dismissing the engineering. But the central interpretive claim—that the frequency split makes the model 'perceive sparse yet valuable behaviors'—is not supported by the current analyses. The needed test is a probe/control experiment; if it fails, the paper should be reframed as an engineering contribution with a learned gated filtering module rather than a frequency-theoretic mechanism. Since the reader already conditioned acceptance on this issue, the verdict remains CONDITIONAL (unchanged).","tokens_in":16122,"tokens_out":7171,"duration_ms":74794,"concrete_test":"Take a trained FARM checkpoint on the Kuaishou dataset. For a held-out sample, compute the band-limited reconstructions Low[V_video], High[V_video], Low[V_live], High[V_live] from Eqs. (4)-(5) using the same cut-off c. For each position in the user sequences, the sparse-behavior labels (like, comment, gift) are available in the training logs. Train a linear probe on each band to predict these labels and compare AUC; include a control where the DFT basis is replaced by a random orthogonal split of the same band sizes. If the high-frequency probe does not clearly beat both the low-frequency probe and the random-split control, the §2.1/3.2 semantic mapping is not the operative mechanism, and the central novelty claim needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 asserts that 'low-frequency information reflects user's dense behaviors... high-frequency information represents user's sparse behaviors,' and Eq. (9) builds the model on this. The assertion is not derived, and the DFT index over a user's ordered sequence does not naturally encode behavior-type rarity. An isolated rare event (a like or gift at one position) is localized in time; its DFT magnitude is flat across all frequencies (an impulse is broadband), so it is not concentrated in the high-frequency band. The high band contains more DFT coefficients, so it carries most of the impulse energy for trivial bandwidth reasons, but it also carries all other rapid variation, including alternating dense behaviors. The ablations in Tables 1-2 and the case study in §4.3 show that removing the module hurts and that learned weights β,δ can be large; they do not show that the high-frequency band specifically encodes like/comment/gift. The empirical gains are therefore equally consistent with an alternative explanation: the module is a learned gated residual branch that adds capacity, and the frequency semantics are a post-hoc interpretation. This is not an internal inconsistency, but it makes the stated novelty load-bearing and currently unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FARM, a cross-domain live-streaming recommendation model for Kuaishou. FARM applies a Discrete Fourier Transform (DFT) to user behavior sequences in both short-video and live-streaming domains, splits each sequence into low- and high-frequency components, and claims that low-frequency components reflect dense behaviors (click, long-view) while high-frequency components represent sparse valuable behaviors (like, comment, gift). It then aligns user preferences across domains with a contrastive loss and fuses them with attention mechanisms before feeding a multi-task model (PLE/MMoE). Offline experiments on two Kuaishou applications compare FARM against a PLE base and four ablations; online A/B testing reports gains in interaction and core metrics. The central claim is that the frequency-aware module enables perception of sparse behaviors, leading to the reported improvements.","tokens_in":16395,"tokens_out":4341,"duration_ms":41755,"significance":"If the frequency-to-rarity mapping were validated, FARM would offer a novel angle on sparse-behavior modeling in industrial live-streaming recommendation. The paper's strengths include a large-scale real deployment with online A/B testing, a clear modular architecture, and ablation evidence that each module contributes to offline metrics. The main scientific claim, however, rests on an asserted rather than derived association between DFT frequency and behavior rarity, and the current experiments do not provide a direct test of that association. The empirical gains over the chosen base are small and are not compared against published state-of-the-art cross-domain or frequency-based recommenders. The significance of the work is therefore currently limited by the lack of support for its central mechanism, though the engineering contribution is credible.","major_comments":[{"comment":"The statement that low-frequency DFT components reflect dense behaviors and high-frequency components represent sparse behaviors is asserted without derivation, and it is in tension with standard DFT properties. A DFT coefficient's index corresponds to the rate of variation along the sequence positions, not to the rarity of a behavior type. A single isolated like or gift is a localized impulse whose spectrum is broadband, so its energy is not concentrated in the high-frequency band; conversely, alternating dense behaviors can produce high-frequency variation. This mapping is load-bearing because Eq. (9) builds the entire intra-domain frequency-aware module on it. Please provide a direct empirical test: for example, compute the distribution of high-frequency energy contributed by positions labeled as like/comment/gift versus other positions, or train the model with only the high-frequency branch and report like/comment/gift prediction metrics. Without such evidence, the stated mechanism remains a post-hoc interpretation rather than a validated design.","section":"Section 2.1, Eq. (9)"},{"comment":"The offline evaluation compares FARM only against the PLE base model. Since the paper claims superiority over prior live-streaming and cross-domain methods, the evaluation should include at least one or two representative published baselines from the related work cited by the authors, such as Moment&Cross [3], LCN [14], eLiveRec [40], or a frequency-based sequential model like FMLP-Rec [43] or BSARec [28]. Additionally, all reported numbers are point estimates with no standard deviations or significance tests. The improvements are small (e.g., +0.18% AUC for Effective-view on Kuaishou), so statistical reliability matters. Please report multiple seeds or confidence intervals and clarify whether the 'Improve' column is percentage-point difference or relative improvement.","section":"Tables 1-2, Section 4.1.2"},{"comment":"The frequency-aware analysis is a case study of a single user and a single candidate author, with two predicted probabilities and two learned weights. This is anecdotal evidence. The large values of β and δ only show that the learnable gates are active; they do not show that the high-frequency band specifically encodes like/comment/gift behaviors, since a gated residual branch could equally well be amplifying useful signal for independent reasons. Please provide a population-level study: for instance, slice users by the density of their sparse behaviors and show that the high-frequency component contributes more for low-density users, or correlate the model's like/gift prediction improvements with the proportion of high-frequency energy in the corresponding behavior sequence positions.","section":"Section 4.3 (RQ2)"},{"comment":"Figure 3 reports that FARM has higher cosine similarity between h_video and h_live than the w/o C-PA variant, and this is presented as evidence that the preference align module works. However, Eq. (14) directly optimizes exactly this cosine similarity through the contrastive loss, so the observed increase is partly by construction. This analysis is self-confirmatory and should be framed as a sanity check, not as independent evidence of effectiveness. The stronger evidence is the degradation of w/o C-PA in Tables 1-2; please rely on that and, if possible, add a downstream metric such as online click or gift rate for the align module.","section":"Section 4.4, Eq. (14)"},{"comment":"The offline results are reported to three or four decimal places with no variance, despite the statement that 'we report the average performance of all models on testing set over hours.' If this averaging is over hourly test sets, the standard deviation across hours should be available and should be reported. Given the very small absolute improvements (e.g., +0.18% AUC for Effective-view on Kuaishou), the lack of uncertainty makes it impossible to assess whether the differences are meaningful. Please add standard deviations or statistical tests for the main comparisons in Tables 1 and 2.","section":"Section 4.2, Tables 1-2"}],"minor_comments":[{"comment":"The text says 'γ and δ control the strength of high-pass components,' but γ already denotes the live-domain low-pass weight in Eq. (9); this should be β and δ.","section":"Section 3.2, Eq. (10)"},{"comment":"The phrase 'a serious of tailor-designed attention mechanisms' appears multiple times; it should be 'a series of'.","section":"Abstract and Section 3.3"},{"comment":"The definition of the DFT matrix and basis vectors is underspecified: the matrix F is defined via rows f_j, but the dimension N and the indexing convention are not stated precisely, and writing f_j in R^N is inconsistent with complex exponentials. Please clarify with a standard definition.","section":"Section 2.1, Eq. (1)"},{"comment":"The data-collection description is terse. The phrase 'first-only strategy [3]' and the 30s sliding window need a brief explanation (e.g., what is a 'segment sample' and how negatives are defined) for reproducibility.","section":"Section 4.1.1"}],"recommendation":"major_revision","confidential_remarks":"This is an industry paper with deployment evidence and a clear ablation study, but the central conceptual claim—that DFT high-frequency components encode sparse valuable behaviors—is not grounded in the signal-processing properties of the behavior sequence and is not directly tested. The comparison only against PLE, without published baselines, further weakens the academic contribution. The paper may be better suited for an industrial or systems-oriented venue, or the authors need to substantially rework the evidence for the frequency mechanism. The online A/B gains are a strong practical signal, but they do not by themselves validate the stated interpretation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a credible industrial cross-domain live-streaming recommender with small, internally consistent gains, but its central 'frequency-aware' narrative is not backed by the evidence in the paper. The DFT split with learned weights may well be doing useful work, but the claim that high-frequency components encode sparse, valuable behaviors like like/comment/gift is asserted without derivation.\n\nWhat's genuinely new is the package: applying DFT-based low/high splitting from the sequential-rec literature to cross-domain live-streaming, with learned per-domain weights, contrastive align-before-fuse, and multi-task prediction. The paper is honest about being an industrial deployment. The offline ablations are consistent, the online A/B shows believable gains on engagement and gift count, and the case study of a low-active user is a reasonable sanity check. The authors also document the data sparsity problems specific to their setting clearly.\n\nThe soft spot is real and load-bearing. In Section 2.1 and Eq. (9), the paper maps DFT frequency index to behavior rarity. That mapping is not derived, and the stress-test note is right: an isolated rare event is a time-localized impulse, whose DFT magnitude is broadband, not concentrated in the high band. The high band also captures rapid alternations of dense behaviors. So the ablations show that removing the module hurts, and the case study shows that the learned weights on the high band are large, but both observations are consistent with the module acting as a learned gated residual branch. The stated mechanism is therefore unsupported. Figure 3 is also self-confirmatory, because the contrastive loss directly optimizes the cosine similarity being plotted.\n\nOther, smaller issues: only PLE is compared; no published SOTA baselines. No error bars or variance on any reported metric. The final hyperparameters (cut-off frequency, contrastive weight) are chosen from ranges but the selected values aren't disclosed. None of these are fatal, but they should be fixed.\n\nWho this is for: practitioners building cross-domain live recommender systems will get value from the architecture and deployment details. Researchers should treat the frequency mechanism with skepticism. It deserves a serious referee because it is a deployed system with real data and a falsifiable claim; the referee should ask for a direct test that ties high-frequency components to sparse behavior labels. That would make the paper much stronger.\n\nMy recommendation: send to peer review, but flag the frequency-behavior mapping as the key issue that needs a direct experiment.","headline":"Solid industrial cross-domain recommender with believable small gains, but the 'high-frequency = sparse behaviors' story is asserted, not shown.","tokens_in":16938,"tokens_out":2155,"would_cite":false,"duration_ms":20063,"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":"FARM claims that splitting a user's behavior history into low- and high-frequency components lets a live-streaming recommender perceive sparse high-value behaviors such as likes, comments, and gifts, and that this lifts cross-domain…","keywords":["live-streaming recommendation","cross-domain recommendation","Discrete Fourier Transform","frequency-aware modeling","sparse user behaviors","multi-task learning","contrastive learning","attention mechanism"],"falsifier":"Compare the high-frequency DFT components of behavior sequences that end in a gift or comment with those that end in a click: if the spectra do not separate, or if ablating only the high-frequency path hurts click prediction as much as gift prediction, then the asserted frequency semantics are not supported and the gains could come from added parameters instead.","tokens_in":15950,"feed_emoji":"🎯","tokens_out":9833,"duration_ms":90615,"temperature":0.7,"pith_summary":"The paper identifies two data-sparsity problems in live-streaming recommendation: high-value behaviors such as like, comment, and gift are rare compared with clicks, and short-video exposure is about nine times larger than live-streaming exposure, so the live-streaming domain alone under-models user preference. To address this, it proposes FARM, which applies the Discrete Fourier Transform to each domain's historical behavior sequence and recombines low- and high-frequency components with learned per-domain weights, allowing the model to attend to sparse behaviors. It then aligns the short-video and live-streaming preference representations with contrastive learning before fusing them with attention. Across the two deployed apps, offline results over the PLE base model show average gains of 0.42% in AUC, 1.06% in UAUC, and 1.05% in GAUC, and online A/B tests show increases in click, like, follow, and gift count. The paper argues that these gains come from the frequency-aware module perceiving sparse yet valuable high-frequency behaviors.","feed_headline":"Frequency-split histories lift live-streaming recommendations","feed_subtitle":"Separating sparse like, comment and gift signals from dense clicks lifts offline metrics and online gift count.","key_machinery":"The central object is the frequency split of a user behavior sequence via the Discrete Fourier Transform: given a sequence embedding $V$, the spectrum $\\bar{x} = Fx$ is partitioned at a cut-off $c$ into low components $Low[V]$ and high components $High[V]$, and the module returns the learned weighted sum $\\alpha\\,Low[V] + \\beta\\,High[V]$. The load-bearing identity is the unitary DFT decomposition of a sequence into Fourier basis components, together with the paper's semantic mapping that high-index basis components carry sparse valuable behaviors. This machinery converts sequence position into frequency and is what allows the model to attend to behaviors that occur rarely in the raw sequence.","core_discovery":"The central claim is expressed in Eq. (9): $F_{video} = \\alpha \\, Low[V_{video}] + \\beta \\, High[V_{video}]$ and $F_{live} = \\gamma \\, Low[V_{live}] + \\delta \\, High[V_{live}]$, where the $Low$ and $High$ operators reconstruct a sequence from the low-index and high-index parts of its Discrete Fourier Transform spectrum. The paper asserts that low-frequency information reflects dense behaviors such as click and long-view, while high-frequency information represents sparse valuable behaviors such as like, comment, and gift; the trainable sigmoid weights $\\alpha, \\beta, \\gamma, \\delta$ let the model upweight the high-frequency part. After reweighting, a multi-head target-attention using the candidate live-streaming author as query extracts the behavior signals and feeds the multi-task prediction head. The supporting evidence is that removing either frequency module degrades all six prediction tasks, that a case study shows a low-active user's predicted like and gift probabilities rise from about 0.4 to above 0.88 when the high-frequency weights are high, and that online A/B tests show gains in like, comment, follow, and gift count. In the paper's account, this is what it means for the model to perceive sparse user preference.","pith_inferences":["An implication the paper leaves implicit is that the same low/high frequency split could be applied to other multi-behavior domains, such as purchase or bookmark, where the sparse behavior is the one the platform most wants to predict.","Because the DFT is applied along the sequence-position axis, the 'high frequency equals rare behavior' mapping is not guaranteed by the transform itself; a direct test would compare the frequency spectra of sequences ending in a gift with those ending in a click.","The contrastive alignment runs only from video to live, so a bidirectional or live-to-video variant could be tested to see whether the trade-off between the two domains changes.","The learned frequency weights could serve as a per-user interpretability signal, flagging users whose sparse behaviors carry the most predictive weight."],"forward_implications":["In both deployed apps, FARM's offline AUC, UAUC, and GAUC beat the base model across all six prediction targets, with the largest relative gains on long-view, comment, and gift metrics.","Removing the cross-domain preference fuse module causes the largest degradation among the ablations, so the transfer of short-video preference into the live-streaming domain is load-bearing for the result.","The case study and the learned high-frequency weights indicate that the frequency module specifically repairs prediction of sparse like and gift behaviors for low-active users.","Online A/B testing shows increases in click, like, follow, and gift count, implying that frequency-aware perception of sparse behaviors converts into engagement and revenue metrics."],"supporting_citations":[{"why":"Defines the first-only real-time reporting strategy that produces the live-streaming interaction dataset used for training and evaluation.","marker":"[3]"},{"why":"Supplies the align-before-fuse principle that motivates the cross-domain preference align module.","marker":"[18]"},{"why":"Provides the multi-gate mixture-of-experts multi-task architecture used as the prediction head.","marker":"[23]"},{"why":"Prior sequential-recommendation work that integrates low- and high-frequency information via the Fourier transform, the approach FARM adapts.","marker":"[28]"},{"why":"Defines PLE, the multi-task base model whose offline metrics FARM is compared against.","marker":"[30]"},{"why":"Provides the multi-head attention mechanism used in the target-attention, self-attention, and cross-attention modules.","marker":"[33]"},{"why":"Supplies the DFT formulation and low/high component decomposition equations used by the intra-domain frequency-aware module.","marker":"[38]"}],"fun_headline_variants":["DFT splits signals to boost sparse live-streaming likes","High-frequency sparse gifts get DFT boost in live recsys","Frequency-aware model lifts live recs via DFT split","Cross-domain DFT captures sparse like, gift signals","Sparse live behaviors amplified by frequency split in model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the frequency index in the Discrete Fourier Transform of a user's behavior sequence corresponds to behavior rarity, so the high-frequency components of the history carry the sparse behaviors like, comment, and gift.","fun_headline_variants_meta":{"raw":{"variants":["DFT splits signals to boost sparse live-streaming likes","High-frequency sparse gifts get DFT boost in live recsys","Frequency-aware model lifts live recs via DFT split","Cross-domain DFT captures sparse like, gift signals","Sparse live behaviors amplified by frequency split in model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1552,"prompt_tokens":1130,"completion_tokens":422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":746,"completion_tokens_details":{"reasoning_tokens":345}},"tokens_in":746,"tokens_out":422,"duration_ms":5646,"temperature":1.0,"reasoning_tokens":345,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T21:42:23.526592+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the high-frequency DFT components of behavior sequences that end in a gift or comment with those that end in a click: if the spectra do not separate, or if ablating only the high-frequency path hurts click prediction as much as gift prediction, then the asserted frequency semantics are not supported and the gains could come from added parameters instead.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior sequential-recommendation work that integrates low- and high-frequency information via the Fourier transform, the approach FARM adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DFT formulation and low/high component decomposition equations used by the intra-domain frequency-aware module."}],"review_version":1}