{"id":"f882144b-a3e2-4651-8488-900f3f94de9b","arxiv_id":"2602.02685","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Sparse Top-2 routing beats full ensemble in decentralized diffusion models, and the paper attributes this to expert-data alignment rather than numerical stability — though much of the supporting evidence is circular.","lead":"This paper claims that in decentralized diffusion models, image quality is governed by routing each denoising step to experts trained on similar data, not by numerical stability. It reports that averaging all experts is the most stable strategy but gives the worst FID (47.9 vs 22.6 for Top-2), yet the evidence for the proposed mechanism is partly circular.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Causal role of alignment untested: paper's own Misaligned Top-2 control reports only Δrefine, not FID; headline gap may be due to sparsity/averaging rather than expert-data alignment.","rationale":"The reader's weakest assumption focuses on the bLeff proxy: Definition 6.1 defines L_eff with the full Jacobian, but Appendix D.3 reports an expert-only proxy, and the omitted router Jacobian is 2–4 orders of magnitude larger. This is a real internal inconsistency and weakens the stability-quality dissociation. However, I regard the missing FID for the Misaligned Top-2 control as more load-bearing because it targets the positive claim directly. Even if the bLeff ordering were corrected, the paper still lacks a causal test of alignment. Δrefine partially protects the dissociation (full ensemble has the lowest Δrefine), but no comparable independent metric protects the alignment claim. The verdict remains REJECT: the central claim is not established by the presented experiments.","tokens_in":18148,"tokens_out":9591,"duration_ms":89116,"concrete_test":"Compute FID for the paper's Misaligned Top-2 condition (random expert selection, same Paris experts, same seed set, n=1000, Heun-50) using the identical FID protocol as Table 4. Compare to aligned Top-2 FID (22.60) and Full ensemble FID (47.89). If random Top-2 FID is within ~2–3 points of aligned Top-2, the alignment principle is not supported; if it is much worse (near Full), alignment is the governing factor. Also report random Top-2 mean cluster rank to verify that alignment is actually broken.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—expert-data alignment governs generation quality—is not causally tested. The headline comparison (Top-2 FID 22.60 vs Full ensemble 47.89, Table 4) varies both alignment and the number of active experts. The paper contains the natural control: Misaligned Top-2, random expert selection, which preserves sparsity while breaking proximity-based alignment (Appendix F.2, Table 8). Yet Table 8 reports only Δrefine for this condition—no FID, no LPIPS, no perceptual quality metric. Without this number, the FID gap may reflect averaging dilution or ensemble size, not alignment. The other supporting analyses are not independent: cluster-distance alignment (Table 1) is nearly guaranteed because the router was trained post hoc to route inputs to the expert trained on the most similar data (§5.2); per-expert prediction 'quality' (Table 2) is measured as angular deviation from the blended velocity, which is defined as the weighted combination of the selected experts, so selected experts are partially aligned by construction; and the disagreement–LPIPS analysis (Figure 1) defines quality degradation as distance to the Top-2 output, the very reference being explained. Additionally, Top-1 achieves better cluster rank (1.54) than Top-2 (1.96) but worse FID (30.60 vs 22.60), so the alignment-quality relationship is not monotone. The causal mechanism is therefore unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies inference-time routing in Decentralized Diffusion Models (DDMs), where independently trained experts are combined by a router. It claims a stability–quality dissociation: full ensemble routing has the lowest trajectory sensitivity and step-refinement disagreement but the worst FID, while sparse Top-2 routing has the best FID despite higher sensitivity. The paper proposes that expert-data alignment — routing inputs to experts trained on similar data — is the governing principle of generation quality, and it attempts to validate this via cluster-distance analysis, per-expert prediction alignment, expert-disagreement analysis, and a separate MNIST-based DDM. It also formalizes a trajectory-local sensitivity measure and gives a conditional convergence argument. The central claim is plausible and counterintuitive, but the current evidence is weakened by several circular or incomplete validation steps.","tokens_in":18508,"tokens_out":6194,"duration_ms":59556,"significance":"If the dissociation and alignment principle were fully established, the result would be practically valuable: it would tell DDM practitioners to prioritize alignment-aware sparse routing over stability optimization, and it would challenge a common numerical-stability intuition for generative ODEs. The paper has real strengths: it works on two DDM systems, includes a formal conditional convergence proposition (Prop. B.1), openly acknowledges the retrospective nature of its sensitivity diagnostic (Remark 6.3), and provides a useful Jacobian decomposition (Table 6). However, the main empirical validations are partly circular or lack the natural control, so the manuscript currently supports a weaker statement than its title and conclusions claim.","major_comments":[{"comment":"The cluster-distance validation is circular. The manuscript states that the Paris router was 'trained post-hoc on the full dataset, effectively learning to route inputs to the expert trained on the most similar data.' Therefore the finding that Top-1/Top-2 select clusters with mean ranks 1.54/1.96 is largely a restatement of the router's training objective, not an independent confirmation of expert-data alignment. This measurement cannot serve as direct evidence for the paper's central claim unless compared against a router that was not trained to match data proximity.","section":"§5.2, Table 1"},{"comment":"The paper's central causal claim is untested. The headline quality comparison (Top-2 FID 22.60 vs. Full ensemble 47.89, Table 4) varies both sparsity and alignment simultaneously. The natural control is Misaligned Top-2 (random expert selection), which preserves sparsity while breaking proximity-based alignment, but Table 8 reports only Δrefine for this condition (0.040±0.035), not FID or any perceptual quality metric. If Misaligned Top-2's FID is close to aligned Top-2's 22.60, the FID gap is due to sparsity/averaging rather than alignment; if close to 47.89, the alignment claim is supported. Without this number, the 'governs' formulation is unsupported. Additionally, Table 4 and Table 1 show that Top-1 achieves better mean cluster rank (1.54) than Top-2 (1.96) but worse FID (30.60 vs. 22.60), so the alignment-quality relationship is not monotone and the claimed mechanism needs a sharpe","section":"§F.2, Table 8; Table 4"},{"comment":"The reported bLeff in Table 4 does not match Definition 6.1. Definition 6.1 defines the effective Lipschitz constant for the full routed flow v_t = Σ_k w_k v_k, i.e., using ||J_x v_t||. Appendix D.3 reports only the expert term ||Σ_k w_k ∇_x v_k|| and states that the router term ||Σ_k v_k ∇_x w_k|| is 2–4 orders of magnitude larger (Top-2: 923±1.4K; Full: 1161±2.1K). With the full Jacobian, the Full ensemble does not have the lowest mean sensitivity; the router-term mean is higher for Full than for Top-2. The claimed stability ordering in Table 4 therefore depends entirely on omitting the dominant term. The authors need to justify why the expert-only Jacobian is the relevant sensitivity measure for cross-strategy quality comparisons, or report both quantities and show that the dissociation is robust to this modeling choice.","section":"§D.3, Table 6; Table 4; Definition 6.1"},{"comment":"The per-expert prediction-quality metric is partially circular. The 'blended velocity' v_t is defined as the weighted combination of the selected experts, so a selected expert is guaranteed to have nonzero influence on the reference vector. The finding that selected experts have smaller angular deviation from the blended velocity (3.6° vs. 5.1°) may reflect the weighting scheme rather than intrinsic prediction quality. An independent ground-truth reference, such as the clean-data velocity or per-expert contribution to final sample quality, is needed to support the claim that routing systematically identifies more accurate experts.","section":"§5.3, Table 2"}],"minor_comments":[{"comment":"The preliminary analysis of the router Jacobian term reports ρ = -0.07 with p = 0.62 and n = 50. This is severely underpowered and should not be used to conclude that router-term sensitivity is uninformative; either report a properly powered analysis or remove the claim.","section":"§6.3 footnote"},{"comment":"It is unclear whether the bLeff values in Table 4 are full Jacobian norms or expert-only norms; Appendix D.3 says expert-only, but Section 6.2.2 and Definition 6.1 suggest full Jacobian. Also, baseline bLeff for Full ensemble in Table 7 (15.97±5.51) differs from Table 4 (17.07±6.33) despite both being 'baseline'; with n=50 vs n=1000 this may be sampling noise, but the inconsistency should be explained.","section":"Table 4 and Tables 7–8"},{"comment":"The sentence 'this correlation provides causal evidence' overstates the design: the disagreement–quality analysis is correlational and uses LPIPS to the Top-2 output as a reference. Please soften the causal language or add an explicit causal manipulation (e.g., Misaligned Top-2 with FID).","section":"§5.6"},{"comment":"The y-axis label 'LPIPS to Top-2' makes clear that the reference is the Top-2 output. This is a reasonable relative-quality reference, but it should be stated as a relative degradation measure, not as an absolute quality metric, throughout the text.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"I am recommending major revision rather than rejection because the most damaging gaps are, in principle, addressable with the existing assets: FID for Misaligned Top-2 can be computed on the released Paris checkpoints or on the MNIST DDM, and the sensitivity analysis can be rerun with the full Jacobian or with a clearly justified expert-only variant. If the authors cannot supply these numbers, the paper's central claims should be substantially weakened to a descriptive finding about the specific pretrained router rather than a general governing principle."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper reports a credible empirical dissociation — full ensembling is numerically most stable but worst in FID — but the causal story (expert-data alignment) is not supported by the experiments as designed. The central validation is circular, and the one control that could save it lacks a FID number. Still, I'd send this to referees, not desk-reject it.\n\nWhat's genuinely useful: the dissociation itself, documented with several independent metrics (local truncation error, Δrefine, Jacobian norms). The finding that full ensemble has the best numerical convergence and worst FID (47.89 vs 22.60) is worth reporting. The MNIST replication is a reasonable sanity check. The paper is also transparent — Remark 6.3 and Appendix E are honest about limitations.\n\nThe soft spots are serious. Section 5.2 says the Paris router was trained post-hoc to route to the expert with the most similar data. Table 1 then shows that sparse routing selects the closest clusters. That's guaranteed by construction, not evidence. The per-expert analysis (Table 2) compares angular deviation from the blended velocity, which is a weighted sum of the selected experts — so selected experts are aligned by definition. The disagreement–LPIPS analysis (Figure 1) defines quality as distance to Top-2, which is partly tautological.\n\nThe most damaging omission: Appendix F.2 includes a Misaligned Top-2 control that preserves sparsity but breaks proximity-based alignment, yet only reports Δrefine, not FID. Without that number, the headline FID gap could be due to sparsity or averaging dilution rather than alignment. That is the obvious control, and it's missing.\n\nAlso, bLeff is computed on the expert-only Jacobian, while Table 6 shows the router term dominates by 2–4 orders of magnitude. The paper's argument that the router term is not discriminative across strategies may be right, but it's a modeling choice. The dissociation still holds via Δrefine, which is independent, so the numerical-convergence ordering is not the problem. But the sensitivity story is weaker than presented.\n\nThe principle itself is not new — the related-work section admits matching inputs to trained experts is well-established in MoE. What would be new is showing it governs DDM quality. That requires the missing control.\n\nWho is this for: people building routing strategies for decentralized diffusion. It's a useful empirical pointer, but not a clean mechanistic conclusion. I'd want the authors to give us the Misaligned Top-2 FID and to reframe the cluster-distance analysis as a sanity check, not validation. Then it would be much more convincing.\n\nMy call: deserves peer review, but major revision is needed. Recommend to engage.","headline":"Empirical dissociation likely real, but the alignment claim rests on a circular validation and a missing control.","tokens_in":18958,"tokens_out":4481,"would_cite":false,"duration_ms":41249,"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":"In decentralized diffusion models, routing inputs to experts trained on similar data determines generation quality, not numerical stability: the most stable ensemble strategy produces the worst samples (FID 47.9) while sparse Top-2 routing","keywords":["decentralized diffusion models","expert-data alignment","routing","trajectory sensitivity","stability-quality dissociation","mixture of experts","diffusion sampling"],"falsifier":"Compute the effective Lipschitz constant along sampled trajectories using the full Jacobian including the router term (rather than the expert-only term) for Top-1, Top-2, and full ensemble. If full ensemble is not the lowest, the dissociation evaporates. Alternatively, run FID on 'misaligned Top-2' (random expert selection preserving sparsity): if FID stays near 22.6, alignment is not the governing factor.","tokens_in":17986,"feed_emoji":"🧩","tokens_out":4837,"duration_ms":42125,"temperature":0.7,"pith_summary":"This paper asks what controls sample quality in decentralized diffusion models (DDMs), where several diffusion models are trained independently on disjoint data clusters and combined at inference by a router. The intuitive answer—that minimizing trajectory sensitivity should yield better samples—is shown to be wrong: the full ensemble of all experts is the most numerically stable and has the lowest step-refinement disagreement, yet produces the worst FID (47.9), while sparse Top-2 routing is less stable but yields the best FID (22.6). The authors argue instead that expert-data alignment is the governing principle: quality depends on sending each denoising state to experts whose training distribution covers that state. They support this with three analyses: cluster-distance ranking, per-expert velocity alignment, and a disagreement–quality correlation, replicated on a second MNIST-based system.","feed_headline":"Stable ensembling yields worst samples in decentralized diffusion","feed_subtitle":"Routing inputs to experts trained on similar data—not numerical stability—governs generation quality.","key_machinery":"The central object is the routed velocity field v_t(x_t) = Σ_k w_k(t)(x_t) v_k(t)(x_t), where w_k are router weights over independently trained diffusion experts. The paper's key conceptual identity is a decomposition of the flow Jacobian into an expert term Σ_k w_k ∇_x v_k and a router term Σ_k v_k ∇_x w_k; since the router term is two to four orders of magnitude larger but nearly identical across routing strategies, the authors use the expert term as the sensitivity measure. The governing notion is expert-data alignment, operationalized by three measurable quantities: the rank of a selected expert's training-cluster centroid in a learned visual-embedding space, the angular deviation betwee","core_discovery":"The central claim is a dissociation: numerical stability does not govern generation quality in decentralized diffusion. Full ensemble routing—which averages all expert velocity predictions—has the lowest effective Lipschitz constant along trajectories and the smallest refinement disagreement, yet the worst generation quality; sparse Top-2 routing, with higher sensitivity, produces the best samples. The governing factor is expert-data alignment: routing inputs to experts trained on similar data. Evidence includes cluster-distance analysis (selected experts' training clusters rank close to the current input, mean rank 1.96 vs 4.5 random), per-expert analysis (selected experts are 29% closer in","pith_inferences":["A direct testable extension would train the router with an explicit alignment objective (e.g., maximizing predicted per-expert velocity alignment) and check whether FID improves independently of sparsity; the paper's analysis suggests it should.","The stability–quality dissociation may generalize to other ensemble generative models (e.g., mixture-of-experts diffusion transformers), where the same mechanism—forcing all experts to process every token—could explain quality losses despite smooth activations.","A practical diagnostic for decentralized serving: compute trajectory-integrated expert disagreement at sampling time and use it to trigger a fallback (e.g., more steps or a different router) when disagreement spikes.","The paper's counterfactual 'misaligned Top-2' is reported only for refinement error; testing FID under random expert selection would separate the effect of sparsity from alignment, a clean next experiment."],"forward_implications":["Practitioners should route toward experts whose training distribution covers the current denoising state; optimizing numerical stability metrics alone can actively hurt sample quality.","Sparse routing (Top-2) delivers the best quality with 4× fewer active experts, so alignment-based sparsity yields both better FID and lower compute.","Full ensemble underperforms not because it is unstable but because it forces most experts to process out-of-distribution inputs; improving expert OOD robustness may rescue ensemble strategies.","Within a fixed routing strategy, trajectory-local sensitivity can still flag numerically sensitive samples, though it is not a cross-strategy quality predictor.","Expert disagreement can serve as a cheap quality monitor during decentralized inference: when selected experts disagree, expected quality drops."],"fun_headline_variants":["Expert-data alignment beats stability in diffusion routing","Stability doesn't predict sample quality in decentralized diffusion","Routing to trained experts, not stability, drives diffusion quality","Why ensemble stability fails in decentralized diffusion","Decentralized diffusion: alignment trumps stability"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The dissociation rests on measuring trajectory sensitivity with the expert-only Jacobian norm; if the full Jacobian (whose router-gradient term is 2–4 orders larger and slightly favors Top-2 over full ensemble) is the proper sensitivity measure, the full ensemble may no longer be the most stable and the stability–quality dissociation collapses.","fun_headline_variants_meta":{"raw":{"variants":["Expert-data alignment beats stability in diffusion routing","Stability doesn't predict sample quality in decentralized diffusion","Routing to trained experts, not stability, drives diffusion quality","Why ensemble stability fails in decentralized diffusion","Decentralized diffusion: alignment trumps stability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1187,"prompt_tokens":748,"completion_tokens":439,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":382}},"tokens_in":492,"tokens_out":439,"duration_ms":3998,"temperature":1.0,"reasoning_tokens":382,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:18:20.835059+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the effective Lipschitz constant along sampled trajectories using the full Jacobian including the router term (rather than the expert-only term) for Top-1, Top-2, and full ensemble. If full ensemble is not the lowest, the dissociation evaporates. Alternatively, run FID on 'misaligned Top-2' (random expert selection preserving sparsity): if FID stays near 22.6, alignment is not the governing factor.","supporting_citations":[],"review_version":1}