{"id":"96a28831-b48f-4b74-bc22-58420271eb23","arxiv_id":"2602.01728","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A shared expert plus routed, subject-group experts with mutual loss reweighting improves cross-subject EEG classification on seven benchmarks.","lead":"EEG signals vary between people, so models trained on some subjects often fail on new ones. This paper proposes a framework that runs a shared expert and a set of subject-group experts together, and reports gains on seven EEG benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 applies a sequential-alignment definition to a branching/MoE architecture without proof; the reducibility-cost criterion is unsupported as stated.","rationale":"The paper's strongest claim is that the choice between shared and routed experts depends on the reducibility cost of branch-specific to branch-invariant functions, and that MGEC implements this theory. I focused on the theoretical foundation rather than the missing error bars or code because the paper explicitly markets itself as theory-to-method mapping, and the reader's verdict already flags the lack of proofs. The specific weakness is that Theorem 2 uses Definition 1's alignment measure, which is designed for sequential compositions, on a branching/MoE architecture without proving the translation. This is an internally inconsistent use of the framework, not merely an empirical gap. If the theorem cannot be derived, the paper's central theoretical contribution collapses to an unsubstantiated heuristic, even though the empirical method might still work. The reader's weakest assumption (Assumption 4) is also important, but it is downstream of Theorem 2: the theorem is used to justify the per-sample assumption, and if the theorem itself is not valid, that justification is already broken. For this reason, I consider the alignment-measure issue the single most load-bearing concern. The recommended verdict remains CONDITIONAL: the authors should be required to either provide a valid proof of Theorem 2 under the stated definitions, or re-characterize the reducibility-cost claim as a heuristic. I partially agree with the reader because the reader also points at unproved theory, but we identify different specific steps as the load-bearing one.","tokens_in":12928,"tokens_out":15522,"duration_ms":160878,"concrete_test":"Ask the authors to supply a proof of Theorem 2 that explicitly maps the branching target g of Definition 2 onto the sequential-composition form g = f_n ∘ ... ∘ f_1 required by Definition 1, and compute Align(R,g) with a justified module count n. Verify whether the router/expert parallel combination can be expressed as such a composition and whether the module count is actually M+1. If the proof requires a new definition of alignment for branching or parallel networks, or if the module count differs, then Theorem 2—and the reducibility-cost criterion—is unsupported as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim (Theorem 2) is not actually derived from the framework the paper sets up. Definition 1 defines alignment only for functions that decompose as a sequential composition, g = f_n ∘ ... ∘ f_1, with Align(N,g) = n·max_i M(f_i,N_i). Definition 2 then introduces a branching target g(x) = Σ_j 1_{I_j}(h0(x))·h_j(x) and a routed-expert network R with a router R0 and experts R_j, and Theorem 2 asserts that Align(R,g) is governed by (M+1)·max(M(R0,h0), max_j M(R_j,h_j)). But a branching function is not a composition of M+1 sequential sub-functions: the router and experts operate in parallel and are combined by a weighted sum. The n·max_i formula is not justified for this parallel/gating architecture, and the assumption M(R0,h0) ≤ max_j M(R_j,h_j) is simply asserted. The 'reducibility cost' criterion, which is the headline insight of Section 2, is therefore not a theorem of the stated framework; it is an unproved analogy. The reader's Assumption 4 is a related but downstream issue: even if Theorem 2 were valid, it concerns aggregate sample complexity and does not imply a per-sample guarantee that at least one expert is close to g for every x. The more basic problem is that the alignment measure itself has not been extended to the branching/MoE setting, so the comparison between routed and shared experts lacks a rigorous foundation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses cross-subject EEG generalization by combining a shared expert (for invariant, reducible functions) with routed mixture-of-experts (for subject-specific, irreducible functions), plus a mutual-guided reweighting scheme. The authors claim a theoretical justification based on algorithmic alignment: the choice between shared and routed experts depends on a 'reducibility cost' (Theorems 1-3), and they instantiate this in an MGEC framework. Experiments are reported on seven EEG benchmarks across auditory attention, motor imagery, and sleep staging, plus a synthetic domain-generalization experiment, with claims of consistent state-of-the-art performance.","tokens_in":13324,"tokens_out":5615,"duration_ms":54949,"significance":"The problem is important and the proposed architecture is a reasonable, well-motivated combination of two common paradigms. The breadth of the empirical evaluation — seven public datasets spanning three tasks and a synthetic verification — is a genuine strength. If the theoretical criterion were rigorously established, it would provide a principled design rule for when to use shared versus routed experts under distribution shift. However, the theory as stated is not currently load-bearing: Theorem 2 is an algebraic consequence of an assumed alignment formula that is never derived for the branching/MoE setting, and Theorem 3 is asserted without proof. The empirical claims also lack uncertainty quantification, and several gains are within 0.1-0.6 accuracy points. The contribution is therefore plausible but not yet established at the standard claimed in the paper.","major_comments":[{"comment":"Definition 1 defines alignment only for sequential compositions g = f_n ∘ ... ∘ f_1, with Align(N,g)=n·max_i M(f_i,N_i). Theorem 2 applies the same n·max_i formula to a branching target g(x)=Σ_j 1_{I_j}(h0(x))·h_j(x) and a routed-expert network R. A branching/gating function is not a composition of M+1 sequential sub-functions: the router and experts operate in parallel and are combined by a weighted sum. No proof is given that Align(R,g) = (M+1)·max(M(R0,h0), max_j M(R_j,h_j)), and the assumption M(R0,h0) ≤ max_j M(R_j,h_j) is simply asserted. Consequently, the 'reducibility cost' criterion is not derived from the stated framework; at present it is an unproved analogy. This is load-bearing for the paper's central theoretical claim.","section":"Section 2, Definition 2 / Theorem 2"},{"comment":"Theorem 3 claims that, under Assumption 4, the total alignment error E[ε_S^(t)+ε_R^(t)] is non-increasing. No proof or training-dynamics model is provided. Assumption 4 — that for every sample at least one expert is ε-close to the target — is an unverified per-sample guarantee. The remark that it 'follows from Theorem 2' is not justified, since Theorem 2 concerns aggregate sample complexity and does not imply a per-sample bound. The mutual-guided losses in Section 3.3, Eqs. (10)-(11), are heuristic reweightings based on cross-entropy losses; if both branches are wrong on a hard/noisy sample, the weighting does not correct the worse branch. Without a proof or a direct empirical check of the learnability-gap premise, the claim that mutual-guided learning prevents over- and under-reduction is not supported.","section":"Section 2, Theorem 3 / Assumption 4"},{"comment":"All comparison tables report point accuracies only; there are no standard deviations, confidence intervals, number of runs, or significance tests. Several reported improvements are very small: AVED is 53.9 vs 53.2 for the best baseline, and ISRUC-3 is 79.8 vs 79.4. Because the central empirical claim is that MGEC 'consistently outperforms state-of-the-art methods', these comparisons need repeated runs with paired significance tests, or at least error bars, before the claim can be evaluated. This is especially important given the small gains on some datasets.","section":"Section 4.2, Tables 2-5"},{"comment":"The synthetic experiment is presented as empirical validation of the theoretical analysis, but it reports a single qualitative curve without error bars, and the mapping from the varying proportion of reducible components to the formal quantity α/β in Theorem 2 is not specified. As the only direct evidence for the theoretical criterion, the experiment is too weak to support the quantitative M+1 threshold in Theorem 2. Please provide repeated runs and a precise description of how the synthetic setup instantiates the reducibility cost.","section":"Section 4.5, Synthetic Verification"}],"minor_comments":[{"comment":"The paper's metadata and the first abstract refer to 'Shared-Routed Expert Alignment (SREA)', while the full-text title and abstract refer to 'Mutual-Guided Expert Collaboration (MGEC)'. Please harmonize these names.","section":"Title/Abstract"},{"comment":"In the last row, 'CbraMod Proposed' reads '76.579.8' — likely a missing column separator. Also, please clarify which backbone is used for the proposed method in the multi-task comparison and whether the improvement over CBraMod is statistically meaningful.","section":"Table 5"},{"comment":"The text states 'the six experts reduce to three groups', but Section 4.1 specifies M=5 experts. Please clarify this inconsistency.","section":"Section 4.4"},{"comment":"The term 'reducibility cost' is never formally defined; it is only the ratio α/β appearing in Theorem 2. A precise definition would help the reader connect the theoretical statement to the method description.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears internally inconsistent in its naming (SREA vs MGEC), which may indicate a versioning issue; the editor may want to verify that the submitted version is coherent. The main technical concern is that the theoretical results are stated without proofs and the central criterion is not derived for the actual MoE architecture. The empirical contribution is broad but needs statistical grounding. I would consider the paper acceptable only after the theoretical claims are either proved, substantially reframed as heuristics, or removed, and after the empirical comparisons are supplemented with uncertainty quantification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper is a solid empirical proposal masquerading as a theory paper. The framework—shared expert plus prototype-routed MoE plus mutual reweighting—is new in this combination, and the evaluation across seven EEG benchmarks in three task families is a real effort. But the theoretical section, which the abstract sells as the main contribution, does not hold up. Definition 1 defines alignment only for sequential compositions; Definition 2 introduces a branching/MoE target, and Theorem 2 then applies the same alignment measure to that branching architecture without any proof or derivation. The reducibility-cost criterion, the headline insight, is essentially a definitional ratio of sample complexities, not a theorem. The stress-test note is right: the n·max formula doesn't extend to parallel/gated structures. Assumption 4, used for Theorem 3, is also unproved and stronger than anything Theorem 2 would give.\n\nWhat's good: the empirical work is honest in places—the case study shows one subject where the proposed method underperforms the ablations, and the synthetic verification in Fig. 5 at least tries to vary the reducibility proportion. The gains over strong baselines are consistent, though small and reported without standard deviations, so we can't tell if they're meaningful.\n\nSoft spots, in order: (1) no proofs for any of the three theorems; (2) no error bars or confidence intervals anywhere; (3) no code released; (4) abstract and full text disagree on the method name (SREA vs MGEC). That last one is sloppy but easy to fix.\n\nThis is a paper worth sending to peer review—the idea is plausible, the benchmarks are substantial, and a good reviewer could push the authors to either prove or soften the theory, add error bars, and release code. But I wouldn't treat the theoretical claims as established. If I were in EEG generalization I'd cite it as a method paper, not as a theory paper.\n\nRecommendation: engage with it as an empirical contribution; ask for proofs or explicit 'this is a conjecture' language, error bars, and code before accepting.","headline":"Plausible empirical combination, but the theory is asserted rather than proved and the evaluation lacks error bars; worth a serious look but not as-is.","tokens_in":13828,"tokens_out":3558,"would_cite":true,"duration_ms":33015,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that cross-subject EEG generalization depends on a reducibility cost, and proposes a framework that pairs a shared expert with routed experts to outperform prior models on seven benchmarks.","keywords":["cross-subject EEG","domain generalization","mixture-of-experts","shared expert","reducibility cost","algorithmic alignment","mutual-guided learning","EEG decoding"],"falsifier":"Train on a dataset where some samples are hard for both a shared model and a routed mixture-of-experts (e.g., synthetic data with overlapping class regions). If the mutual-guided loss makes training diverge or degrade compared to a fixed-weight ensemble, Assumption 4 is violated and the theory's guarantee breaks; if it stays stable, the mechanism is more forgiving than the assumption.","tokens_in":12827,"feed_emoji":"🧠","tokens_out":4828,"duration_ms":45042,"temperature":0.7,"pith_summary":"This paper is trying to establish a design principle for cross-subject EEG models: whether to use one shared network or many specialized expert networks is not a matter of taste but depends on a reducibility cost, the sample-complexity price of approximating subject-specific functions by a single shared function. The authors prove a theorem (Theorem 2) that says a shared expert aligns better when that cost is low, and routed experts align better when it is high. They then build the Mutual-Guided Expert Collaboration (MGEC) framework, which runs a shared branch and a routed mixture-of-experts branch simultaneously and uses a mutual-guided reweighting to let each branch learn from the other on samples where the other is confident. On seven public EEG datasets, MGEC consistently outperforms state-of-the-art models, including EEG foundation models, which suggests the principle is more than theoretical. If correct, this gives practitioners a principled basis for choosing between invariant and subject-specific representations instead of guessing.","feed_headline":"Shared plus routed experts beats both alone on EEG","feed_subtitle":"Theory says the choice hinges on reducibility cost; the framework realizes it and wins on seven benchmarks.","key_machinery":"The central object is the reducibility cost, defined through the sample-complexity bounds M(.,.) and the ratio alpha/beta in Theorem 2. The proof machinery is algorithmic alignment: a network aligns with a target function if its modules can be replaced by sub-functions with bounded sample complexity. The framework itself is carried by three mechanisms: (1) a shared expert trained with a joint embedding loss over augmented temporal neighbors to control transfer error; (2) a routed mixture-of-experts with prototype-based sparse routing and a subject-level entropy loss for specialization; (3) a mutual-guided reweighting that adjusts each branch's loss by the cross-branch gap exp(l_i^other - l_i","core_discovery":"The paper's central claim is that the divide between shared and routed architectures in cross-subject generalization is governed by a reducibility cost: the sample-complexity ratio of learning a branch-invariant function versus branch-specific ones. The authors formalize this in Theorem 2, which says a shared expert has better alignment when reducibility is cheap (small beta) and routed experts win when it is expensive (large alpha). They then instantiate the insight in MGEC, a framework that trains a shared branch and a routed mixture-of-experts branch jointly, with a mutual-guided reweighting that upweights samples where the better branch is confident and the other is not. Experiments on s","pith_inferences":["A sharper test of the theory than the paper's synthetic experiment would be to measure per-subject reducibility cost on a real EEG dataset and verify that the optimal expert count tracks the predicted regime boundary.","The mutual-guided reweighting could be applied to any pair of models, not just EEG; this would make the framework a general recipe for domain generalization beyond brain signals.","If Assumption 4 fails on noisy samples, a confidence-gated reweighting that downweights samples where both branches disagree should be strictly more robust; the paper does not test this variant."],"forward_implications":["Practitioners can decide a priori whether to invest in a shared model, an expert mixture, or both, by estimating the reducibility cost on a validation set.","The mutual-guided reweighting is a plug-in regularizer that can be added to any shared/expert pair, not just EEG models, making it a general domain-generalization recipe.","The theory provides a principled explanation for why ensemble methods sometimes beat single models on heterogeneous subjects and sometimes hurt.","MGEC achieves consistent gains over strong backbones, including foundation models, across seven datasets spanning three EEG tasks."],"fun_headline_variants":["Reducibility cost decides shared vs routed EEG experts","Shared+routed experts beat either alone on EEG","EEG experts: don't pick, combine shared and routed","Reducibility cost: when to share vs route EEG experts","Theory-driven mix of shared and routed experts wins on EEG"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper assumes that for any sample, at least one of the two branches is already close to correct; if both branches are wrong on hard or noisy samples, the mutual-guided reweighting will upweight the worse branch's error instead of correcting it.","fun_headline_variants_meta":{"raw":{"variants":["Reducibility cost decides shared vs routed EEG experts","Shared+routed experts beat either alone on EEG","EEG experts: don't pick, combine shared and routed","Reducibility cost: when to share vs route EEG experts","Theory-driven mix of shared and routed experts wins on EEG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001793,"raw_usage":{"total_tokens":6889,"prompt_tokens":719,"completion_tokens":6170,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":6090}},"tokens_in":463,"tokens_out":6170,"duration_ms":38343,"temperature":1.0,"reasoning_tokens":6090,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:33:09.101609+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train on a dataset where some samples are hard for both a shared model and a routed mixture-of-experts (e.g., synthetic data with overlapping class regions). If the mutual-guided loss makes training diverge or degrade compared to a fixed-weight ensemble, Assumption 4 is violated and the theory's guarantee breaks; if it stays stable, the mechanism is more forgiving than the assumption.","supporting_citations":[],"review_version":1}