{"id":"c7034e5e-2ff9-4207-afa3-e335ddc0729d","arxiv_id":"2504.12920","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CSMF sequentially fine-tunes a two-tower EBR model with selective parameter masks, then serves a weighted linear combination of exposure, click, and conversion scores from one 64-dimensional index, improving offline and online metrics without extra retrieval cost.","lead":"An e-commerce retrieval team proposes CSMF, a way to train one retrieval model for three objectives by pruning and reusing different parts of the network for each objective. The method reports better retrieval accuracy than several multi-objective baselines while keeping online storage and latency nearly unchanged.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (8)'s single-dot-product fusion is not a consequence of the described magnitude-pruning CPP: the required block-decomposable embeddings and cross-tower mask alignment are never specified.","rationale":"The reader correctly identifies that mask alignment across towers is an unflagged assumption. I agree that this is a gap, but the more load-bearing problem is one level deeper: even with perfectly aligned masks, Eq. (8) requires the network output to be a concatenation of objective-specific blocks, and the magnitude-based element-wise pruning described in Section 4.2 does not produce such a decomposition. The paper's experiments and online A/B test are meaningful evidence that some version of the method works in practice, and a block-structured implementation would be compatible with the figures, so this is not grounds for rejection. However, the central algebraic claim is not derivable from the written procedure, and the preprint provides no code or formal derivation to resolve the ambiguity. A revision should either state and justify the block-structured mask (with synchronized user/item masks) or provide a corrected derivation of Eq. (8); until then the central claim should be treated as conditional. Since the reader's verdict was already CONDITIONAL, the verdict stays unchanged, though the justification is tightened.","tokens_in":16906,"tokens_out":17794,"duration_ms":185292,"concrete_test":"Implement the described CSMF on a small two-tower MLP (two hidden layers, 64-dim output) with tau=0.75, following Sections 4.1-4.2 exactly: pre-train on exposure, prune by cumulative percentile, recover accuracy, fine-tune on click, prune, recover, fine-tune on conversion. For a fixed batch, compute s_d, s_o, s_r by the three masked forward passes (theta_d, theta_d union theta_o, theta). Then compute the single dot product on the right-hand side of Eq. (8) using the block-weighted user vector and the item vector. If the two values differ by more than numerical precision, Eq. (8) does not follow from the described training. Separately, record whether the block partition of the user tower equals that of the item tower; if no shared mask is imposed, the identity also fails from misaligned block coordinates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's 'without increasing network parameters or online retrieval overhead' claim rests entirely on Eq. (8), which asserts that a weighted sum of three objective scores equals one dot product of a block-scaled user vector with the item vector. The equality is valid only if the final embedding is literally the concatenation [e_d; e_o; e_r], with the click model's output equal to [e_d; e_o] and the conversion model's output equal to [e_d; e_o; e_r] (no cross terms). The paper never establishes this block structure. Section 4.2 defines CPP as PackNet-style magnitude pruning: within each layer the scalar weights |w_kj| are sorted and those below a cumulative percentile are pruned, then retrained for the downstream objective. Element-wise masks of this kind do not make the output of a multi-layer network equal the concatenation of per-objective subnetwork outputs; an output unit can still receive contributions from all weight subsets. The notation {·;·} is also overloaded: in Eq. (4) it is parameter-set union, while in Section 4.4 it is vector concatenation, and Eq. (8) silently switches between the two. Additionally, even if block-structured outputs were imposed, the user tower and item tower must share the same block partition; Section 4.2 gives no rule synchronizing masks across towers, so e_u_theta_d^T v_i_theta_d need not be the exposure score. Both conditions are unstated and unverified, and the central serving mechanism depends on them.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CSMF, a multi-stage fine-tuning method for two-tower embedding-based retrieval. A backbone is pretrained on exposure data; its least important parameters (under a cumulative-percentile heuristic) are pruned and then fine-tuned on click data; the process is repeated for conversion data. The claimed contribution is that after training, the exposure, click, and conversion scores can be fused online by rescaling blocks of the user embedding and computing a single dot product (Eq. (8)), so that multiple objectives are served from one unchanged ANN index with no extra parameters or retrieval latency. Experiments on an industrial advertising dataset and the AliExpress dataset, plus an online A/B test, report consistent gains over baselines such as MOPPR, MVKE, and DMMP.","tokens_in":17244,"tokens_out":9644,"duration_ms":105140,"significance":"If the central serving claim holds, the paper makes a practically valuable contribution: multi-objective EBR with flexible objective weights at no additional ANN storage or latency. The experimental evidence is substantial, including industrial-scale logs, multiple strong baselines, component ablations, hyperparameter sensitivity, a serving-cost table, and an online A/B test, and the flat storage/latency result in Table 4 is a genuine strength. However, the mathematical derivation that guarantees this efficiency, Eq. (8), depends on a block-decomposition property of the embeddings that the paper neither states nor verifies. The significance is therefore conditional on an architectural clarification and a check of the deployed towers.","major_comments":[{"comment":"The derivation of the zero-overhead fusion score is not supported by the training procedure described in Sections 4.1 and 4.2. The equalities e_{θ_d;θ_o} = [e_{θ_d}; e_{θ_o}] and e_θ = [e_{θ_d}; e_{θ_o}; e_{θ_r}] presuppose that the final embeddings are blockwise concatenations of sub-embeddings computed from disjoint parameter blocks. The described CPP operates on individual weights/neurons within each layer and never specifies that the user and item towers are pruned into identical, positionally aligned blocks, nor does it define how the output of a multi-layer network using interleaved θ_d and θ_o decomposes into separate sub-embeddings. Without this architectural constraint, the second and third equalities in Eq. (8) do not follow, and the abstract claim of no additional network parameters and no online retrieval overhead is unsupported. The paper should state the block structure as an explicit construction (e.g., each final-embedding dimension is assigned to one objective block and all upstream connections respect the block partition) and verify it in the deployed system.","section":"Section 4.4, Eq. (8)"},{"comment":"The pruning rule P(w_kj) = f(c_j ≤ ind_k) is not well-defined unless the neurons are sorted by |w_kj| before forming the cumulative sums C_k. As written, c_j is a cumulative sum over an arbitrary neuron ordering, so pruning the prefix of that ordering has no clear relation to the stated goal of pruning redundant or low-information parameters. The method should either specify that neurons are sorted by magnitude (as in PackNet) or justify a different ordering; otherwise the comparison with fixed-ratio pruning in Table 3 is ambiguous.","section":"Section 4.2, CPP definition"}],"minor_comments":[{"comment":"The notation {·;·} is overloaded: it denotes parameter-set union in Eq. (4) but vector concatenation in Section 4.4 and Eq. (8). Using distinct notation would prevent confusion.","section":"Eq. (4) and Section 4.4"},{"comment":"The method name is written as DMTL in Section 2.1 but as DTML in Table 2; the naming should be made consistent.","section":"Related Work and Table 2"},{"comment":"The y-axis label in Figure 5 says 'Hitrate@50' while the text and Table 2 use Recall@50; these should be unified.","section":"Figure 5"},{"comment":"In the sentence describing the weights assigned to the user-side vector, 'e_i_{θ_d}' should likely be 'e_u_{θ_d}'; as written it appears to refer to the item vector.","section":"Section 4.4"},{"comment":"The figure labels in the PDF are difficult to read, with concatenated words and stray symbols (e.g., 'Objective1ExpertsObjective2ExpertsShareExperts'). Redrawing the figures would improve clarity.","section":"Figures 2 and 3"},{"comment":"The exposure, click, and conversion quantities s_d, s_o, and s_r are called 'probabilities', but Eq. (1) and Eq. (8) define them as dot-product scores; the normalized probability is p_θ in Eq. (2). This terminology should be clarified.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth a look: it applies PackNet-style selective masking to multi-objective EBR in a way I have not seen before, and it comes with an online A/B test and a concrete serving-cost table. The training procedure—pre-train on exposure, prune, recover, fine-tune on click, then conversion—is clearly described, and the offline gains on both industrial and AliExpress data are plausible. Credit where due: the authors correctly identify the dimensionality blowup of MoE-based retrieval models and propose an inference-time parameter-free alternative.\n\nThe central selling point, Eq. (8), has a structural problem that is never addressed. The formula rewrites a weighted sum of three objective scores as a single dot product, which is only valid if the user and item embeddings are literal concatenations [e_d; e_o; e_r] and each objective's score is computed from the corresponding block. The pruning method in Section 4.2 is element-wise magnitude pruning per layer; nothing about that procedure guarantees such a block decomposition. An output unit can still receive contributions from all parameter subsets. The paper also never says that the user and item towers are pruned with matched masks, which Eq. (8) requires. So the 'no online overhead' claim rests on an unstated architectural assumption. The notation {·;·} is used both for parameter-set union and vector concatenation, which hides the switch.\n\nThis is a load-bearing gap, but it is fixable: the authors could clarify that the network is designed with disjoint sub-networks whose outputs are concatenated, and that masks are aligned across towers. As written, a reviewer cannot verify the main claim from the preprint.\n\nOther soft spots: no code, no public data, and the significance claim ('p < 0.05') is given without details on variation or tests. The hyperparameter sensitivity study is reasonable, and the ablation is helpful.\n\nWho is this for? Researchers and engineers working on retrieval, especially with multi-objective serving constraints. You will get value from the training recipe and the online results, but you should treat Eq. (8) with suspicion until clarified.\n\nVerdict: it deserves serious peer review, but needs major revision to state and verify the block-alignment condition. I would send it to referees.","headline":"CSMF is a promising PackNet-style approach to multi-objective EBR, but the no-overhead fusion formula rests on an unstated block-structure assumption that the paper never verifies.","tokens_in":17760,"tokens_out":2438,"would_cite":false,"duration_ms":25518,"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":"A single embedding model can retrieve for several objectives at once by splitting its parameters into blocks and reweighting them at serving time, with zero extra storage.","keywords":["recommendation systems","embedding-based retrieval","multi-objective optimization","efficient fine-tuning","parameter pruning","two-tower retrieval","softmax loss","online serving"],"falsifier":"Take the trained CSMF model, choose any weight triplet, and compute the left and right sides of Eq. (8) over a few thousand random user-item pairs; if the two numbers differ by more than floating-point round-off on a single pair, the serving claim breaks. A cheaper probe is to inspect the pruning masks of both towers and check whether every layer keeps the same neuron indices in both towers.","tokens_in":1761,"feed_emoji":"🎯","tokens_out":2166,"duration_ms":101929,"temperature":0.7,"pith_summary":"The paper claims that a single two-tower embedding-based retrieval model can be trained to serve several objectives at once, such as exposure, click, and conversion, and then be adjusted online by reweighting objectives with zero additional computation or storage. It proposes CSMF, which sequentially fine-tunes the model on each objective while pruning away redundant neurons from earlier stages to make room for later ones, plus a cumulative percentile pruning method and an adaptive margin loss to manage conflicts. The payoff is Eq. (8): a weighted sum of three objective scores equals a single dot product of a block-weighted user vector with the unchanged item vector, so one approximate-nearest-neighbor index continues to serve arbitrary objective weightings. If true, industrial systems can switch retrieval objectives by changing three scalar weights rather than re-embedding or adding experts.","feed_headline":"One dot product serves click, conversion, and more","feed_subtitle":"CSMF partitions a two-tower model into objective blocks so weighted scores stay a single lookup, with no added storage or latency.","key_machinery":"The load-bearing identity is Eq. (8), which rewrites the weighted sum of three objective-specific dot products as one dot product of a block-concatenated user vector with the item vector, where the user-side block scalars are accumulated sums of the objective weights. The identity is made available by the CSMF training schedule: a pre-train, selective-mask, accuracy-recovery, fine-tune cascade that partitions the two-tower network into three mutually exclusive parameter sets, one for exposure, one for click, and one for conversion, so that each objective's score is computed by a disjoint prefix of the final embedding. The pruning step is Cumulative Percentile-based Pruning, which per layer prunes the neurons whose cumulative absolute value stays below a fraction $\\tau$ of the layer's total, and the conflict-handling step is the Cross-Stage Adaptive Margin Loss, which adaptively widens or narrows the margin between positive and negative items depending on whether upstream objective scores agree.","core_discovery":"The paper's central discovery is that multi-objective weighted fusion need not increase the embedding dimension or retrieval cost if the model's parameters are partitioned into disjoint blocks, each responsible for one objective. By pre-training on exposure data, then fine-tuning on click data after pruning, then fine-tuning the further-pruned remainder on conversion data, CSMF yields three scores $s_d$, $s_o$, and $s_r$ from the same forward pass. Because each score is a dot product of the user and item sub-vectors of the corresponding parameter block, the linearly weighted score $k_d s_d + k_o s_o + k_r s_r$ collapses, via Eq. (8), to a single dot product in which the user-side block weights are simply $(k_d+k_o+k_r)$, $(k_o+k_r)$, and $k_r$. This lets deployment change the objective mix by reweighting three scalars while the item vectors and the approximate-nearest-neighbor index stay exactly as they were.","pith_inferences":["The block-reweighting trick generalizes to any number of cascaded objectives: each additional objective adds one parameter block and one more ring of accumulated weights in Eq. (8), so the serving index never grows.","Because the weights act only on the user vector, an online system could vary the objective-weight triplet per user, per scene, or even per request without touching the approximate-nearest-neighbor index; the paper only demonstrates fixed global weights in its experiments.","A structured variant that fixes the pruning masks of the user and item towers to the same neuron indices at every layer would make Eq. (8) an identity rather than an assumption on the paper's part; checking whether the current masks already align is a cheap and high-value numerical test."],"forward_implications":["One unchanged approximate-nearest-neighbor index serves any weighting of exposure, click, and conversion objectives; changing $k_d$, $k_o$, and $k_r$ requires no re-embedding or reindexing.","Mixture-of-experts style multi-objective retrieval models can be replaced by CSMF while keeping storage and latency flat; the paper reports 1019.95 MB of vector storage versus 1020.11 MB for a baseline, and ANN time 1.22 ms versus 1.21 ms.","On the industrial dataset, CSMF gains 3.51 and 6.61 percent in nDCG@50 and Recall@50 on clicks and 2.23 and 7.51 percent on conversions over the best baseline, with online A/B gains of 0.42 percent RPM, 0.57 percent CTR, and 0.67 percent CVR.","Because the weighted fusion uses only the final embeddings, the online serving cost of rebalancing objectives is effectively zero, enabling quick adaptation to changing business priorities.","Because cascade training uses the largest dataset first, downstream objectives with sparse data inherit a strong backbone rather than requiring separate expert modules."],"supporting_citations":[{"why":"Supplies the selective-pruning plus accuracy-recovery paradigm (PackNet) that CSMF adapts to multi-objective retrieval.","marker":"[24]"},{"why":"The mixture-of-experts multi-objective retrieval baseline (MVKE) whose increased vector dimension and serving latency CSMF is designed to avoid.","marker":"[33]"},{"why":"The listwise multi-objective retrieval model (MOPPR) used as the strongest online baseline and the control in the A/B test.","marker":"[39]"},{"why":"Establishes the two-tower embedding-based retrieval plus ANN setup and in-batch negative sampling that CSMF inherits.","marker":"[14]"},{"why":"Provides the YouTubeDNN two-tower architecture used as the base model and as the separate-model baseline.","marker":"[35]"},{"why":"Supplies the contrastive softmax objective that the EBR training loss and the adaptive margin loss build on.","marker":"[25]"}],"fun_headline_variants":["Three objectives, one dot product, zero extra latency","Cascaded masks fuse multiple objectives in one dot","Multi-objective ranking with a single lookup","CSMF: one dot product, no added parameters","Weighted objectives without extra retrieval cost"],"cache_read_input_tokens":19840,"weakest_assumption_plain":"The paper's weighted-fusion identity assumes the user tower and item tower are pruned into matching block structures, so the same neuron positions are alive in both towers; if the two towers mask different neurons, the block-weighted dot product no longer equals the summed objective scores.","fun_headline_variants_meta":{"raw":{"variants":["Three objectives, one dot product, zero extra latency","Cascaded masks fuse multiple objectives in one dot","Multi-objective ranking with a single lookup","CSMF: one dot product, no added parameters","Weighted objectives without extra retrieval cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001181,"raw_usage":{"total_tokens":4873,"prompt_tokens":937,"completion_tokens":3936,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":3865}},"tokens_in":553,"tokens_out":3936,"duration_ms":27001,"temperature":1.0,"reasoning_tokens":3865,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:19:43.987817+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained CSMF model, choose any weight triplet, and compute the left and right sides of Eq. (8) over a few thousand random user-item pairs; if the two numbers differ by more than floating-point round-off on a single pair, the serving claim breaks. A cheaper probe is to inspect the pruning masks of both towers and check whether every layer keeps the same neuron indices in both towers.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive softmax objective that the EBR training loss and the adaptive margin loss build on."}],"review_version":1}