{"id":"895dc92c-3134-451c-8d8e-9154b2f228a8","arxiv_id":"2608.08627","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"UniMoMo merges functionally similar experts in MoE recommenders under a traffic-aware protection scheme, preserving ranking quality while cutting expert count and inference latency.","lead":"This paper introduces UniMoMo, a post-training method that shrinks the number of experts in a trained mixture-of-experts recommendation model by merging similar experts while protecting high-traffic ones. It reports that the compressed four-expert model keeps 99.92% to 102.30% of the original ranking quality while running 1.28x to 1.63x faster on an A100 GPU.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 and Table 3 report inconsistent NDCG and latency for the same Amazon Beauty 8-expert source; the claimed 99.92% lower bound may actually be 99.37%.","rationale":"The reader's conditional verdict is well-founded, but the calibration limitation is acknowledged by the authors and affects generalizability rather than the internal validity of the reported numbers. The Table 1 vs Table 3 discrepancy is a more fundamental threat: the source against which preservation ratios are measured is not consistently defined. Table 3's 8-expert row is described in Section 4.4 as the comparison point for UniMoMo's expert-count sensitivity, and if it represents the same Amazon Beauty six-block model as Table 1's Origin MoE, the numbers cannot both be correct. The discrepancy of 0.0054 NDCG (2.0%) is larger than the claimed four-expert preservation loss/gain range for Amazon Beauty; indeed the four-expert ratio relative to the Table 3 value is 99.37%, below the abstract's stated minimum of 99.92%. This means the headline quantitative claim is not reproducible from the paper as written. A single controlled rerun with a fixed seed can determine which number is the source reference. If Table 3's value is reproduced after adaptation, then the method still preserves quality relative to an adapted origin, but the abstract's ratios must be recomputed; if Table 1's value is reproduced, Table 3 contains an error. Either resolution should be reported with per-seed error bars, which the paper currently omits. I therefore leave the conditional verdict unchanged, with the inconsistency as the primary condition to resolve.","tokens_in":79,"tokens_out":7413,"duration_ms":89346,"concrete_test":"Reproduce the Amazon Beauty six-block condition from one saved seed: (a) the frozen Origin checkpoint, (b) Origin after the paper's standard adaptation (5 epochs, lr 5e-4), and (c) UniMoMo with target 8 experts plus adaptation. Compare NDCG@10 and latency against Table 1 and Table 3. If (b) or (c) yields 0.2703/5.243 ms, recompute all source-relative ratios and speedups using that reference; if the 0.2703 value cannot be reproduced, Table 3's 8-expert row is erroneous and the abstract's four-expert 99.92% lower bound is unsupported. Also report per-seed values to resolve whether the 0.2649 vs 0.2703 difference is a seed artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is an internal inconsistency in the source checkpoint that defines every source-relative ratio. Section 4.4's Table 3, described as comparing UniMoMo with 2/4/8 experts on Amazon Beauty, reports the 8-expert (no-reduction) condition at NDCG@10 = 0.2703 and latency = 5.243 ms. Table 1's Origin MoE, also the Amazon Beauty six-block 8-expert checkpoint, reports NDCG@10 = 0.2649 and latency = 4.833 ms. These differ by 2.0% in NDCG and 8.5% in latency. If Table 3's 8-expert row is the correct source reference, then UniMoMo's four-expert ratio is 0.2686/0.2703 = 99.37%, below the paper's stated four-expert lower bound of 99.92%. If Table 1's Origin is the correct reference, Table 3's 8-expert row has no explained provenance. Either way, the headline source-relative preservation numbers and speedups (e.g., 4.833/3.127 = 1.55x vs 5.243/3.127 = 1.68x for the four-expert Amazon Beauty point) are not reproducible from the paper as written. The five-run means have no error bars, so a 2% source discrepancy dwarfs the 0.08-2.30% margins claimed. This is more load-bearing than the calibration-stream limitation because even a perfect calibration set cannot resolve an undefined source reference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniMoMo, a post-training compression method that converts a trained MoE-based recommendation checkpoint (RankMixer-style, 8 experts per MoE block) into a smaller standard MoE with a user-specified expert budget (4 or 2 experts). The method builds an expert-affinity graph from functional output summaries computed on unlabeled calibration tokens, protects high-exposure experts using a routing-entropy-based schedule, greedily coarsens the graph, reconstructs merged experts via traffic-weighted averaging plus a least-squares activation correction, and then applies a short supervised adaptation. Experiments on Amazon Beauty, KuaiRec, and TenRec with 2, 4, and 6 MoE blocks report source-relative five-run mean NDCG@10 ratios of 99.92%--102.30% for the four-expert conversion and 98.36%--104.24% for the two-expert operating point, with measured A100 speedups of 1.28x--2.21x. An appendix reports results on four additional recommendation backbones, where UniMoMo is claimed to rank first in all 24 reported metrics.","tokens_in":19172,"tokens_out":6818,"duration_ms":71245,"significance":"If the central results hold, the paper makes a useful contribution: it defines a checkpoint-level expert-budget conversion problem for recommendation MoEs, solves it with a deterministic, unlabeled-calibration-driven coarsening procedure, and preserves standard top-k inference with no method-specific online module. Strengths include the explicit serving-budget framing, the joint use of functional affinity and routing exposure, the included complexity analysis, measured latency, and a matched comparison against MergeMoE. However, the manuscript currently contains an internal inconsistency in the source checkpoint that defines every source-relative ratio, and the five-run means are reported without any per-run or variance information. These issues must be resolved before the stated preservation ratios and speedups can be considered reproducible.","major_comments":[{"comment":"The 8-expert source is reported inconsistently. Table 3, Amazon Beauty, 8 experts, gives NDCG@10 = 0.2703, HR@10 = 0.4492, AUC = 0.6815, latency = 5.243 ms. Table 1, Amazon Beauty Origin MoE, which is the same six-block eight-expert source with the same 24.97M parameters, gives NDCG@10 = 0.2649, HR@10 = 0.4391, AUC = 0.6778, latency = 4.833 ms. Section 4.4 explicitly uses the Table 3 value to state that four experts retain 99.37% of NDCG@10, while the Abstract, Section 1, Section 4.2, and Section 6 report the lower bound as 99.92%. If Table 3 is the correct reference, the four-expert ratio is 0.2686/0.2703 = 99.37% and the latency speedup is 5.243/3.127 = 1.68x; if Table 1 is the correct reference, Table 3's 8-expert row has no explained provenance. This discrepancy is larger than the 0.08%--2.30% margins being claimed. The authors must identify the correct source reference, recompute all source-relative ratios and speedups, and report per-run values so that the corrected numbers can be checked.","section":"§4.4, Tables 1 and 3"},{"comment":"All headline claims are based on five-run means without standard deviations, ranges, or per-run values. The matched gains over MergeMoE are 0.0004, 0.0028, and 0.0051 in NDCG@10, and the ablation differences are 0.0018 and 0.0034, so the central claim rests on differences at the third decimal place. Without any measure of run-to-run variability, it is not possible to determine whether these differences are stable or whether the source-relative ratios are within noise. Please report the per-run numbers or at least standard deviations for the main tables.","section":"§4.1--§4.2, Tables 1 and 2"},{"comment":"The functional affinity is computed from an isotropic Gaussian summary consisting of a mean vector and a single scalar variance per expert. The memory argument for using a scalar variance is reasonable, but the manuscript provides no evidence that this summary preserves the merge decisions that would be obtained from a full-covariance or otherwise richer affinity measure. Since the affinity graph is the primary input to the coarsening algorithm, this is a load-bearing modeling choice. Please add a validation on at least one dataset/layer comparing the isotropic summary against a full-covariance or alternative affinity measure, or demonstrate that the resulting partition is stable under perturbations of the summary statistics.","section":"§3.3, Eqs. (4)--(6)"},{"comment":"The protection schedule uses hand-set hyperparameters gamma_min = 0.1, gamma_max = 0.4, and beta = 1, with no sensitivity analysis. Because the ablation study (Table 2) is used to claim that the exposure constraint contributes to the result, the reader needs to know how sensitive the final partition and NDCG@10 are to these choices. Please report a sensitivity sweep over gamma_min, gamma_max, and beta for at least one dataset, or state explicitly that the conclusions are robust across a tested range.","section":"§3.4, Eq. (14)"}],"minor_comments":[{"comment":"The 8-block row in Table 4 repeats NDCG@10 = 0.2703 from Table 3's 8-expert row. Please clarify whether this is a copy error or whether the depth sweep and the expert-count sweep share the same observation.","section":"Table 4"},{"comment":"The Abstract and Section 1 say the experiments cover 2, 4, and 6 MoE blocks, but Table 4 and its surrounding text include 8 blocks. Please make the scope statement consistent.","section":"Abstract and §1"},{"comment":"In the SDIM row, the two numbers '69.6369.03' are concatenated without a separator. This should be formatted as two separate values.","section":"Table 5"},{"comment":"Figure 3 has no colorbar or axis labels, and the text does not define the displayed similarity scale. Since the figure is presented as a routing-pattern diagnostic, please add a quantitative legend or explicitly state what the cells and shading represent.","section":"Figure 3"},{"comment":"Several references in the recommendation-MoE acceleration discussion, such as [22, 23, 24, 52, 59], concern AIGC detection, image/video interpretability, or document intelligence and are not connected to recommendation MoE acceleration. Please remove them or explicitly connect them to the topic.","section":"§2.2"},{"comment":"The sentence 'an internal variable named BPR does not denote pairwise BPR loss' suggests a misleading variable name in the code or protocol description. Please rename the variable or clarify the actual objective function being used.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The related-work section contains several references from AIGC and document-intelligence papers that appear unrelated to the recommendation-MoE topic; I would encourage the editor to ask the authors to justify or remove these citations. The internal inconsistency between Tables 1, 3, and 4 must be fixed before the paper can be accepted; if the corrected source-relative ratios fall below the stated ranges, the authors should reframe the contribution accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take: UniMoMo is a genuinely useful post-training compression recipe for MoE recommenders, and the core idea is sound — but the paper has an internal numeric inconsistency in its own source checkpoint that currently makes the headline preservation ratios unverifiable. That has to be fixed before I'd trust the deployment claims.\n\nWhat's new: the paper formulates expert-bank reduction as a budgeted constrained graph coarsening, coupling two signals that earlier expert-merging work treats separately — functional affinity estimated on calibration states (isotropic Gaussian KL) and routing-exposure protection derived from routing entropy. The least-squares activation correction for SwiGLU merging is handled cleanly, and Section 3.5's error bound honestly states when the reconstruction is exact. The serving-budget framing is a nice reframing: target expert count is a deployment constraint, not a metric-searched hyperparameter. The matched comparison to MergeMoE with the same adaptation budget is the right experimental design, and the appendix shows portability across four backbones. The limitations section is honest about calibration dependence.\n\nThe soft spots, in order of seriousness. First, the Table 1 / Table 3 discrepancy is load-bearing. On Amazon Beauty six-block, Table 1's Origin MoE is 0.2649 NDCG@10 and 4.833 ms. Table 3's 8-expert row, described as the uncompressed checkpoint, is 0.2703 and 5.243 ms. Those are the same configuration. If Table 3 is the reference, the four-expert 0.2686 gives 99.37%, not the stated 99.92%; if Table 1 is the reference, Table 3's row has no explained provenance. Either way the source-relative ratios and the speedups (1.55x vs 1.68x for that point) are not reproducible as written. Five-run means with no error bars make this worse — a 2% source drift dwarfs the 0.08–2.3% margins claimed. Second, there is a paragraph in Related Work (2.2) that jumps to fake-image detection papers (Ivy-Fake, Fake-hr1, DAATI, TabDSR, DeFakerOne) with zero relevance to recommendation MoE acceleration; it reads like an insertion artifact and should be cut. Third, no code is released, which for a systems-flavored paper matters. Fourth, the calibration-stream and isotropic-Gaussian summaries are acknowledged limitations but not deeply validated; that is acceptable for a first report, but the paper should be explicit that the affinity measure is an assumption rather than a validated choice.\n\nBottom line: the method is coherent, the contributions are real, and the comparison is fair. This deserves a serious referee, but the numeric inconsistency must be resolved before acceptance. I'd engage with it once the source-reference question is answered.","headline":"Useful post-training MoE compressor for recommenders with a sound core idea, but an internal Table 1 vs Table 3 inconsistency in the source checkpoint makes the headline preservation ratios unverifiable until fixed.","tokens_in":19732,"tokens_out":3285,"would_cite":false,"duration_ms":30411,"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":"UniMoMo converts a trained eight-expert MoE recommender to a four- or two-expert checkpoint that retains 98–104% of source NDCG@10 while running 1.28–2.21x faster.","keywords":["mixture of experts","expert merging","model compression","recommender systems","graph coarsening","post-training compression","routing exposure","adaptive routing"],"falsifier":"Take a trained eight-expert MoE checkpoint and build two UniMoMo conversions from disjoint calibration streams—one drawn from a narrow time window or a single user cohort, one from the full traffic distribution—while keeping the supervised adaptation stage fixed. If the two four-expert checkpoints differ in five-run mean NDCG@10 by more than the matched-baseline gap reported in the paper (about 0.0004–0.0051), the calibration-representativeness premise is falsified; alternatively, replacing the isotropic Gaussian affinity with a full-covariance affinity and showing a material NDCG shift would falsify the summary-statistics premise.","tokens_in":18605,"feed_emoji":"⚡","tokens_out":7624,"duration_ms":72579,"temperature":0.7,"pith_summary":"UniMoMo addresses the serving cost of trained mixture-of-experts (MoE) recommender systems by converting an existing eight-expert checkpoint into a smaller standard MoE with four or two experts per block, rather than retraining or adding a compression module. The conversion treats expert reduction as a constrained graph-coarsening problem: experts are grouped by how similarly they respond to shared recommendation states on a short unlabeled calibration stream, while experts that carry heavy routing traffic are protected from repeated merging. Reconstructed experts combine traffic-weighted parameter averaging with a least-squares correction of intermediate activations, and a brief supervised adaptation stage adjusts the resized router and experts. Across three datasets and MoE depths of 2, 4, and 6 blocks, the four-expert checkpoints keep 99.92% to 102.30% of source mean NDCG@10 with 1.28x to 1.63x measured A100 speedups, and the two-expert operating point keeps 98.36% to 104.24% with up to 2.21x speedup. The result matters because it turns a trained recommender checkpoint into a deployable smaller model after training, under an explicit expert budget.","feed_headline":"Halving experts keeps recommender ranking and cuts latency 1.3–2.2x","feed_subtitle":"Merging functionally similar experts yields a four- or two-expert MoE that keeps NDCG@10 near 100%.","key_machinery":"The load-bearing object is a calibration-conditioned constrained graph coarsening. For each MoE layer, every expert is evaluated on the same calibration states, and the mean/scalar-variance summaries yield affinity scores $s_{ij} = 1/(1+D_{ij})$ from a symmetrized KL divergence; simultaneously, routing mass defines a protected set via the normalized entropy deficit. The greedy average-link coarsening repeatedly merges the most similar admissible pair until the target expert count is reached, with protection acting as a deterministic merge priority rather than a hard constraint. Reconstruction then fuses each cluster into one SwiGLU expert using routing-mass-weighted projection averages plus a least-squares correction $T_m = P^+ Q$ folded into the down projection, so the compressed layer keeps standard top-$k$ routing and no extra online computation. This one machinery couples behavioral compatibility with routing exposure under the same expert-count budget.","core_discovery":"The paper's central claim is that a trained sparse MoE recommender can be resized to a smaller standard MoE at deployment time without a compression-specific online module, provided the merge plan is conditioned on recommendation traffic. Instead of comparing expert parameters, UniMoMo runs every expert on the same unlabeled calibration inputs and summarizes each expert's outputs as an isotropic Gaussian (mean vector and scalar variance); the symmetrized KL divergence between these summaries defines an expert-affinity graph. A normalized entropy deficit of routing mass marks high-exposure experts as protected, so the greedy average-link coarsening merges behaviorally similar clusters while limiting repeated modification of traffic-heavy experts. Each merged cluster is initialized by a routing-mass-weighted average of the up, gate, and down projections, corrected by a least-squares map on the intermediate SwiGLU activations, and the router rows are set to the arithmetic centroid before supervised adaptation. The paper reports that this complete conversion-and-adaptation workflow preserves source-relative five-run mean NDCG@10 at 99.92%–102.30% for four-expert checkpoints and 98.36%–104.24% for the aggressive two-expert, top-1 point, with measured A100 speedups of 1.28x–1.63x and 1.47x–2.21x respectively.","pith_inferences":["The same calibration-conditioned coarsening should transfer to MoE language models, since the procedure only needs shared hidden states, routing mass, and a target expert count; a testable extension is applying UniMoMo to a Transformer MoE checkpoint and checking whether the 8-to-4 quality retention carries over.","The isotropic Gaussian summary is a strong simplification; comparing merge plans built from full-covariance summaries on a smaller layer would show whether the scalar-variance assumption hides expert differences that matter for ranking.","The dependence on a 20-batch calibration stream suggests a stress test: draw the calibration stream from a narrow time window or a single user cohort, and measure how far the resulting NDCG ratios move from the reported 99.92%–102.30% range.","Because routing-exposure protection sets per-layer protection from entropy alone, a natural extension is making the target expert count per layer an output of the same entropy signal rather than a fixed global budget."],"forward_implications":["A trained eight-expert MoE recommender can be exported to a four-expert, top-2 serving shape that runs 1.28x–1.63x faster while keeping 99.92%–102.30% of source mean NDCG@10.","An aggressive two-expert, top-1 deployment is also viable, retaining 98.36%–104.24% of ranking quality with 1.47x–2.21x speedup.","Because the compressed model retains standard top-$k$ routing and folds all corrections into stored parameters, it can drop into existing serving pipelines without a compression-specific module.","The conversion generalizes across depths of 2, 4, and 6 MoE blocks and across three datasets, and the portability results on four other backbones indicate the partition does not depend on the RankMixer architecture.","The ablation shows both calibrated behavior affinity and routing-exposure protection contribute to final ranking, meaning neither signal alone reproduces the full conversion quality."],"supporting_citations":[{"why":"MergeMoE is the matched 8-to-4 compression baseline that UniMoMo must beat under the same adaptation budget; it supplies the routing-statistics plus least-squares reconstruction approach UniMoMo extends.","marker":"[37]"},{"why":"HC-SMoE provides hierarchical output-based expert clustering, the group-by-output idea UniMoMo adapts to recommendation states with exposure constraints.","marker":"[3]"},{"why":"Merging Experts into One approximates the combined output of activated experts, a reduction operation that motivates the traffic-weighted reconstruction.","marker":"[16]"},{"why":"RankMixer is the industrial-scale ranking backbone whose MoE blocks define the conversion target.","marker":"[74]"},{"why":"MoS is the recommendation-specific MoE architecture used for the four-backbone portability comparison.","marker":"[33]"},{"why":"KuaiRec is the fully-observed interaction dataset supplying one of the three evaluation settings.","marker":"[12]"},{"why":"TenRec is the large-scale benchmark dataset supplying the largest of the three evaluation settings.","marker":"[61]"}],"fun_headline_variants":["Merge experts by response similarity, keep NDCG, speed up 1.3–2.2x","Graph coarsening trims recommendation experts, near-100% NDCG","UniMoMo: post-training MoE compression with traffic-aware merging","Shrink recommender MoE to 2 or 4 experts, preserve ranking","Functional expert merging: 2x speedup, no online module"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the 20 unlabeled training batches used for calibration being representative of the states and routing traffic the recommender will see at serving time, and on the mean-plus-scalar-variance summary capturing the expert differences that matter for ranking; behavior absent from that stream cannot influence the merge plan.","fun_headline_variants_meta":{"raw":{"variants":["Merge experts by response similarity, keep NDCG, speed up 1.3–2.2x","Graph coarsening trims recommendation experts, near-100% NDCG","UniMoMo: post-training MoE compression with traffic-aware merging","Shrink recommender MoE to 2 or 4 experts, preserve ranking","Functional expert merging: 2x speedup, no online module"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001603,"raw_usage":{"total_tokens":6463,"prompt_tokens":1102,"completion_tokens":5361,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":5254}},"tokens_in":718,"tokens_out":5361,"duration_ms":40275,"temperature":1.0,"reasoning_tokens":5254,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:29:05.070205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained eight-expert MoE checkpoint and build two UniMoMo conversions from disjoint calibration streams—one drawn from a narrow time window or a single user cohort, one from the full traffic distribution—while keeping the supervised adaptation stage fixed. If the two four-expert checkpoints differ in five-run mean NDCG@10 by more than the matched-baseline gap reported in the paper (about 0.0004–0.0051), the calibration-representativeness premise is falsified; alternatively, replacing the isotropic Gaussian affinity with a full-covariance affinity and showing a material NDCG shift would falsify the summary-statistics premise.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MoS is the recommendation-specific MoE architecture used for the four-backbone portability comparison."}],"review_version":1}