{"id":"5c8fd1f3-a940-409e-a6b9-72f77ac80fac","arxiv_id":"2508.18376","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DualSparse-MoE cuts about 25% of MoE computation after training using expert partition and static neuron reconstruction, with 0.08-0.28% average accuracy loss and up to 1.41x module speedup.","lead":"This paper describes a way to speed up Mixture-of-Experts language models by cutting roughly a quarter of the computation after training, while keeping accuracy loss under about 0.3%. It works by reorganizing each expert's network into smaller pieces and reconstructing skipped parts from the remaining neurons.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Offline neuron-level reconstruction may not transfer to dynamic serving-time routing; the 0.08–0.28% accuracy bound needs direct out-of-distribution evidence.","rationale":"The reader's weakest-assumption framing identifies exactly the load-bearing condition: the offline neuron-level reconstruction must generalize to dynamic serving-time routing. My read agrees. The strongest claim is a Pareto claim about accuracy retention under 25% computation dropping; for that claim to hold, the reconstruction must be robust to the input distribution actually seen by the router. The abstract provides no evidence of robustness beyond the reported average numbers, and the corrupted full text prevents inspection of the reconstruction's construction, the calibration procedure, or the speedup measurements. Because this is an unverified empirical premise rather than an internal inconsistency, it does not justify moving from the reader's CONDITIONAL verdict to REJECT; it does justify keeping acceptance conditional on a direct out-of-distribution evaluation and per-task results. I am not making an ad hominem argument; the text-corruption header mismatch is weighed only as a verification obstacle, not as evidence about author intent.","tokens_in":23322,"tokens_out":4480,"duration_ms":56481,"concrete_test":"Hold out a disjoint distribution (e.g., a different domain, longer sequences, or a shuffled token-type mixture) from the data used to fit the neuron-level reconstruction and the dynamic-drop thresholds. Run the three MoE models at the enforced 25% drop rate and compare per-task accuracy and end-to-end latency to the no-drop baseline. If the average accuracy degradation exceeds the reported 0.08–0.28% range, or if any single task exceeds it by more than 0.5%, the static-reconstruction/transferability premise fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the dual sparsity identified offline in pre-trained MoE modules is stable enough that a static neuron-level reconstruction can compensate for dynamic, per-input tensor-level dropping. The abstract says the reconstruction is static and the model is not retrained, but it does not specify which calibration data were used to fit the reconstruction, whether that calibration distribution matches serving-time routing, or whether the mapping remains accurate when the router sends unusual token distributions to experts. If the dropped tensor-level entries fall outside the subspace the reconstruction was designed to recover, errors can compound across layers and the reported 0.08–0.28% average accuracy loss will not hold. The supplied full text is corrupted mojibake and its embedded header names arXiv:2508.18374v1 [astro-ph.GA], so the methods section that would justify 'mathematical consistency' and the speedup measurements cannot be audited. The concrete risk is not that the reported numbers are wrong on the authors' benchmarks, but that they fail to transfer to dynamic serving-time inputs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Drawing on the abstract alone: the paper claims pre-trained MoE modules contain dual tensor/neuron sparsity that can be exploited post-training via expert partitioning, without retraining. It proposes DualSparse-MoE combining dynamic tensor-level dropping with static neuron-level reconstruction, reporting ~25% compute drop with only 0.08%-0.28% average accuracy loss on three MoE models, proportional speedups, and a 1.41x module speedup with low degradation. The body text supplied is corrupted mojibake and contains an unrelated arXiv header, so no methods, equations, or experimental tables are available for evaluation.","tokens_in":23414,"tokens_out":3610,"duration_ms":40281,"significance":"The core idea---post-training expert partition plus static reconstruction for efficient MoE serving---is attractive and, if validated, would offer practical deployment benefits because it requires no retraining and applies to existing models. However, the claim is strong: it asserts that static offline reconstruction can compensate for dynamic per-input dropping generally. The evidentiary basis is absent from the submission: no derivations, implementation details, per-task numbers, or error bars. The 0.08-0.28% bounds and the proportional-speedup statement are not auditable. I therefore cannot credit the significance beyond the plausibility of the direction.","major_comments":[{"comment":"The body text is corrupted/unreadable mojibake; the embedded header reads arXiv:2508.18374v1 [astro-ph.GA], not the submitted paper. No methods, equations, algorithms, experiment settings, or tables are visible. None of the abstract's quantitative claims (0.08-0.28%, proportional speedups, 1.41x) can be verified. A complete, correct manuscript is required before further review.","section":"Full text"},{"comment":"No experimental table identifies which three MoE models, which tasks, or which baselines produce the average 0.08%-0.28% accuracy drop. With no per-task breakdown, standard deviations, or comparison to the unmodified model, it is impossible to determine whether the results are consistent across tasks or driven by one favorable setting. Include per-model/per-task tables and error bars.","section":"Abstract (accuracy claims)"},{"comment":"The statement that 'nearly all degrees of computation dropping... yield proportional computational speedups' is load-bearing. In MoE serving, FLOP reduction typically does not translate linearly to wall-clock speedup because of routing, expert loads, and memory bandwidth. The paper does not report hardware, batch size, sequence lengths, overhead, or measured end-to-end speedups vs drop rates. This claim needs direct measurement and a breakdown of compute vs memory vs communication.","section":"Abstract (proportional speedup claim)"},{"comment":"The main risk: the static neuron-level reconstruction is fit on some calibration distribution; the abstract gives no evidence that this mapping transfers to dynamic per-input tensor-level dropping under different routing patterns, input distributions, or sequence lengths. Please report calibration data, validation on OOD/shifted data, and error accumulation through layers. Without this, the 0.08-0.28% accuracy retention bound is not justified as a general claim.","section":"Abstract (generalization of reconstruction)"},{"comment":"The 'approximate 25% drop rate' is an enforced operating point, not derived from accuracy or cost. The paper does not state how the rate is selected or whether the accuracy/speedup findings are robust around it. Clarify whether this is a tuned hyperparameter or a fixed policy, and add sensitivity analyses.","section":"Abstract (25% drop rate)"}],"minor_comments":[{"comment":"Name the three MoE models; 'prevailing' is too vague.","section":"Abstract"},{"comment":"Define 'dual sparsity' and 'mathematical consistency' precisely; these terms are central but not defined.","section":"Abstract"},{"comment":"State what 'MoE module speedup' excludes (end-to-end, prefill, decoding, communication).","section":"Abstract"},{"comment":"Add citations for prior work on finer-grained expert design.","section":"Abstract"},{"comment":"Remove/correct the unrelated arXiv ID and ensure the submission renders as a readable PDF.","section":"Full text"}],"recommendation":"major_revision","confidential_remarks":"The corruption and wrong header may indicate a submission/toolchain issue, but it is essential that the editor verify the authors are aware. I recommend requesting the original source and confirming the arXiv ID. If the full text in circulation is really this corrupted, this should be treated as a return-to-authors concern until corrected. That said, I am recommending major_revision rather than reject because the underlying claims are checkable and could be fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on DualSparse-MoE: the abstract describes a credible, practically useful idea—post-training expert partitioning that induces tensor-level sparsity, paired with static neuron-level reconstruction, no retraining, plus load-imbalance-aware expert parallelism. If the headline numbers hold (roughly 25% less MoE compute at 0.08–0.28% average accuracy loss; 1.41x module speedup), that matters for LLM serving. Credit where earned: the positioning is distinct from work that increases tensor-level sparsity during pre-training, and the numbers are internally consistent.\n\nBut I could not audit the actual methods. The supplied full text is garbled mojibake and the embedded header names a different arXiv ID and category (astro-ph.GA), so I have no equations, no tables, no error bars, no related-work comparison. That is a hard limit on what I can endorse.\n\nThe soft spots beyond the text-integrity problem: \"mathematical consistency\" is asserted without derivation, and \"proportional computational speedups\" is broader than the abstract supports—in MoE serving, FLOP reductions routinely decouple from wall-clock speedup when routing, memory bandwidth, and expert utilization dominate. I also take the stress-test concern seriously: the static reconstruction is fit offline, and the paper does not say which calibration distribution it used or whether the mapping stays accurate when serving-time routing shifts. That is the key empirical question. I would not call it fatal—static reconstructions can be robust—but the authors need to show distribution-shift evidence, not just aggregate benchmark numbers.\n\nI do not see a deep circularity in the 25% drop being a chosen operating point; that is normal for systems papers. The risk is transfer, not definition.\n\nWho is this for? MoE serving practitioners, and readers working on inference-time efficiency. It is an in-subfield systems result, not a paradigm shift. Recommendation: if the submitted PDF is actually what I was given, desk reject and ask for a clean version—no referee can verify anything. If this is a pipeline artifact and the real PDF is clean, send it to peer review, with reviewers asked to scrutinize wall-clock methodology and the offline-to-dynamic transferability claim.","headline":"Post-training expert partitioning with static reconstruction is a plausible MoE-serving win on the abstract, but I could not audit the methods and the transferability claim needs direct evidence.","tokens_in":24022,"tokens_out":2314,"would_cite":false,"duration_ms":33000,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that pre-trained Mixture-of-Experts models can shed about a quarter of their computation at serving time with negligible accuracy loss.","keywords":["Mixture of Experts","sparsity","inference acceleration","LLM serving","expert partitioning","neuron reconstruction","load imbalance","computation dropping"],"falsifier":"Run the same 25% drop pipeline on a held-out set whose input distribution or routing pattern is shifted (e.g., long-document or out-of-domain prompts), and check whether the accuracy gap to the full model remains within the reported 0.08–0.28% band; if the gap widens to about 1% or more, the reconstruction mapping is not transferable.","tokens_in":23094,"feed_emoji":"⚡","tokens_out":4270,"duration_ms":48305,"temperature":0.7,"pith_summary":"The paper argues that pre-trained Mixture-of-Experts LLMs contain two layers of unused computation: tensor-level sparsity (many of the fine-grained sub-FFNs an input could use are not needed) and neuron-level sparsity (many individual neurons inside a layer are inactive for a given token). It shows these can be exploited after training, not during it, by partitioning experts into finer pieces and by learning a static reconstruction that patches up the neurons that are dropped. If the claim holds, a deployed MoE module can skip about a quarter of its computation and lose only a few tenths of a percent of average accuracy, with speedup roughly proportional to the amount dropped. Separately, making expert-parallel scheduling aware of load imbalance adds another 1.41x speedup at 0.5% accuracy cost.","feed_headline":"25% less MoE compute costs only 0.08-0.28% accuracy","feed_subtitle":"Post-training expert partition and neuron reconstruction cut MoE serving cost with no retraining—if the sparsity holds.","key_machinery":"The load-bearing mechanism is the pair of expert partition and neuron reconstruction. Expert partition re-cuts each pre-trained expert into finer sub-FFNs, so tensor-level sparsity can be exploited without changing the model's mathematical output; neuron reconstruction uses a static, offline-built mapping to rebuild the information lost when neurons are dynamically dropped at serving time. The system then does dynamic tensor-level dropping and static neuron-level reconstruction together, and optionally applies load-imbalance-aware expert-parallel scheduling.","core_discovery":"The central claim is that the dual sparsity present in pre-trained MoE modules—tensor-level sparsity among fine-grained expert sub-FFNs and neuron-level sparsity of individual units—is a stable resource that can be harvested at inference time. The paper introduces DualSparse-MoE, which does this by post-training expert partitioning (splitting existing experts without retraining, preserving the mathematical form of the transformation) and then combining dynamic per-input tensor-level computation dropping with static neuron-level reconstruction. On three prevailing MoE models, enforcing an approximately 25% drop rate costs only 0.08–0.28% average accuracy, and the speedups scale almost proport","pith_inferences":["Beyond the paper's measured benchmarks, the strongest test is whether the offline reconstruction mapping transfers to shifted input distributions, unusual router token distributions, or longer sequence lengths; if it does not, the reported accuracy retention will not hold in production.","The 1.41x load-imbalance-aware speedup is likely hardware- and model-dependent, and could be combined with quantization or kernel-level optimizations for even larger serving gains.","If the dual sparsity is a general property of pre-trained MoE models, it suggests current training recipes systematically over-provision experts, and future training could aim for smaller footprints from the start."],"forward_implications":["A 25% drop in MoE computation yields only 0.08–0.28% average accuracy loss, so deployed LLMs can serve more tokens per unit of compute.","Because the sparsity is induced post-training, existing pre-trained MoE models can adopt the method without retraining or modifying the pretraining objective.","Higher drop rates yield roughly proportional speedups, giving operators a dial to trade accuracy for latency.","Adding load-imbalance awareness to expert parallelism gives a 1.41x speedup on the MoE module at 0.5% average accuracy cost, on top of the dropping gains.","The method treats MoE as tensor-level sub-FFNs plus neuron-level units, so it can be layered onto standard MoE serving stacks."],"supporting_citations":[],"fun_headline_variants":["MoE compute cut 25% for under 0.3% accuracy loss","Post-training sparsity: no retrain, 0.08-0.28% accuracy cost","Expert partitioning unlocks inference-time MoE speedup","Cut MoE compute by 25%, keep accuracy within 0.3%","Dual sparsity enables 25% compute cut with minimal accuracy loss"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The dual sparsity found in pre-trained modules is stable and general enough that the offline, static neuron reconstruction continues to repair the outputs after dynamic per-input tensor dropping, even when serving-time inputs, routing patterns, or sequence lengths differ from what was seen when the mapping was built.","fun_headline_variants_meta":{"raw":{"variants":["MoE compute cut 25% for under 0.3% accuracy loss","Post-training sparsity: no retrain, 0.08-0.28% accuracy cost","Expert partitioning unlocks inference-time MoE speedup","Cut MoE compute by 25%, keep accuracy within 0.3%","Dual sparsity enables 25% compute cut with minimal accuracy loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000845,"raw_usage":{"total_tokens":3555,"prompt_tokens":823,"completion_tokens":2732,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":2631}},"tokens_in":567,"tokens_out":2732,"duration_ms":22673,"temperature":1.0,"reasoning_tokens":2631,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:27:13.566805+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 25% drop pipeline on a held-out set whose input distribution or routing pattern is shifted (e.g., long-document or out-of-domain prompts), and check whether the accuracy gap to the full model remains within the reported 0.08–0.28% band; if the gap widens to about 1% or more, the reconstruction mapping is not transferable.","supporting_citations":[],"review_version":1}