{"id":"ad2334b7-cf15-481c-a914-bd3d0061d606","arxiv_id":"2506.12303","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"SPIRE adds per-client embeddings to a shared diffusion backbone, enabling parameter-efficient personalization in federated learning, with new-client KID improvements on MNIST, CIFAR-10, and CelebA.","lead":"This paper proposes SPIRE, a federated learning scheme that keeps one shared diffusion model and gives each client a tiny learnable identity embedding, so personalizing to a new user changes only a few hundred parameters. The authors report lower KID than fine-tuning baselines on three image datasets and present a two-Gaussian analysis connecting the diffusion loss to learning mixture weights.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.2's proof drops the positive factor tanh'(u) when passing to the critical-point condition, giving Eq. (12) with the wrong sign; at the true mixing weight the condition would force 2w−1 = −(2w−1), so the claimed dimension-free bound is not established.","rationale":"I read the paper as making two linked claims: (i) the SPIRE architecture with a shared backbone and per-client embeddings enables parameter-efficient personalization; (ii) in a two-component GMM, gradient descent on the DDPM loss with respect to the mixing weight provably recovers the optimal weight with a dimension-free error bound (Theorem 4.2), and this justifies the conditioning mechanism. The first claim is an empirical engineering contribution that could stand alone; the second is the theoretical anchor cited in the abstract and conclusion. The reader's weakest_assumption identifies the exact place where the anchor breaks: the proof of Theorem 4.2 derives a stationary-point condition by dropping the tanh'(u) factor and obtaining the wrong sign. I verified the algebra: with u = µ_t^T X_t − 1/2 log((1−w)/w), the derivative (using Stein's lemma for the Gaussian noise term) is proportional to E[tanh'(u)(||µ_t||^2 tanh(u) + µ_t^T X_t)]; setting this to zero does not imply Eq. (12), because tanh'(u) is a strictly positive random weight correlated with the bracket. Moreover, the resulting Eq. (12) is inconsistent with the moment identities E[tanh(u)] = 2w−1 and E[X_t] = (2w−1)µ_t unless w = 1/2. The estimator built from Eq. (12), \\hat w = 1/2(1 − µ_t^T \\hat X_t/||µ_t||^2), converges to 1−w, not w. Thus Theorem 4.2 is not merely missing a technical step; its central equation has the wrong sign, and the proof cannot be repaired without changing the estimator and re-deriving the bound. Theorem 4.3 inherits this failure, so the theoretical pillar of the paper is unsupported. The empirical results in Table 1 are reported as point estimates without error bars or code; while this is itself a reproducibility concern, the decisive issue is the invalid theorem. I therefore agree with the reader's verdict and see no basis to adjust it.","tokens_in":16227,"tokens_out":13468,"duration_ms":142956,"concrete_test":"Recompute the derivative in the proof of Theorem 4.2 keeping tanh'(u) and the correct sign of d/dw[−1/2 log((1−w)/w)] = +1/[2w(1−w)]. In d=1 with µ=1, take t=1 and true w=0.7; under q_t, E[tanh(u)] = 0.4 and µ^T E[X]/||µ||^2 = 0.4, so Eq. (12) asserts 0.4 = −0.4. A Monte Carlo evaluation of these two expectations at w=0.7, with 10^6 samples, would immediately falsify Eq. (12), and a direct check of the sample estimator shows \\hat w → 0.3 instead of 0.7.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2, proof of Theorem 4.2: the DDPM objective (11) has u = µ_t^T X_t − 1/2 log((1−w)/w). Differentiating and using Stein's lemma for the Gaussian noise term gives a stationarity condition proportional to E[tanh'(u)(||µ_t||^2 tanh(u) + µ_t^T X_t)] = 0. The proof replaces this with E[tanh(u)] = −µ_t^T E[X_t]/||µ_t||^2, i.e. it drops the strictly positive, random weight tanh'(u) and uses the wrong sign for the log-derivative. That step is invalid because tanh'(u) is correlated with both u and X_t. The sign is also inconsistent with the data: for the true weight w, E[tanh(u)] = 2w−1 and µ_t^T E[X_t]/||µ_t||^2 = 2w−1, so Eq. (12) would force w = 1/2. Correspondingly, the estimator \\hat w = 1/2(1 − µ_t^T \\hat X_t/||µ_t||^2) built from Eq. (12) converges to 1−w, not w, whenever w ≠ 1/2. Since Theorem 4.3 builds on Theorem 4.2, the theoretical claim that gradient descent on the DDPM objective recovers optimal mixing weights with dimension-free error is unsupported. The empirical claim also lacks error bars and code, but the decisive flaw is the invalid proof of the paper's central theorem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes SPIRE, a federated learning method for personalized diffusion models in which a shared U-Net backbone is trained together with per-client identity embeddings, and only the embedding is fine-tuned for unseen clients. The paper pairs the method with a theoretical analysis: for a symmetric two-component Gaussian mixture model with a shared mean and per-client mixing weight, it claims that gradient descent on the DDPM objective with respect to the mixing weight recovers the optimal weight with a dimension-free error bound (Theorem 4.2), and that a two-stage estimation procedure yields a score-error bound (Theorem 4.3). Empirically, SPIRE is compared with FedAvg plus fine-tuning, a shared-representation baseline, and a meta-learning baseline on MNIST, CIFAR-10, and CelebA, reporting KID values for both collaborative pretraining and new-client fine-tuning.","tokens_in":16599,"tokens_out":11626,"duration_ms":130081,"significance":"If the theoretical claims were valid, the paper would make a useful contribution: it would connect conditional diffusion training to Gaussian-mixture estimation, provide a principled explanation of embedding-based conditioning, and show a parameter-efficient path to federated personalization for generative models. The paper has clear strengths: the architectural idea is simple and practically motivated, the new-client adaptation scenario is a real federated learning problem, and the authors use KID rather than FID for small datasets and explicitly state a limitation of their approach. However, the central theorem's proof is invalid: the chain-rule computation drops a positive correlated factor and has an inconsistent sign, and the theorem statement claims a gradient-descent result while the proof actually bounds a different moment estimator. The headline empirical claims also rest on single KID values without error bars or multiple seeds. Because the advertised contributions depend on the invalid theorem and the empirical evidence is not statistically grounded, the paper is not acceptable in its current form.","major_comments":[{"comment":"The derivative of the DDPM loss (11) with respect to w is computed incorrectly. The correct chain rule gives the stationarity condition E[tanh'(u_t)(||μ_t||^2 tanh(u_t) + μ_t^T X_t)] = 0, where u_t = μ_t^T X_t - 1/2 log((1-w)/w). The proof replaces this by Eq. (12), E[tanh(u_t)] = - μ_t^T E[X_t] / ||μ_t||^2, thereby removing the strictly positive and correlated factor tanh'(u_t) and changing the sign. The resulting condition is inconsistent with the data: for the true mixing weight, E[tanh(u_t)] = 2w - 1 and μ_t^T E[X_t] / ||μ_t||^2 = 2w - 1, so Eq. (12) forces w = 1/2. Consequently, the estimator \\w defined from Eq. (12) targets 1-w rather than w whenever w ≠ 1/2, and the dimension-free bound stated in Theorem 4.2 is not established.","section":"Section 4.2, proof of Theorem 4.2, Eq. (12)"},{"comment":"The theorem states that a gradient method on the DDPM loss with respect to the mixing weight recovers the optimal mixing weight with the displayed mean-squared error, but the proof never analyzes a gradient trajectory. Instead, it defines \\w as the first-order moment estimator obtained from Eq. (12) and computes the variance of that estimator. No argument shows that gradient descent on the objective (11) converges to this \\w, nor that the empirical minimizer of the DDPM loss is close to the moment estimator. Thus even a corrected stationarity computation would not support the theorem as stated. Since Theorem 4.3 (Eq. (13)) uses the dimension-free mixing-weight bound from Theorem 4.2, its O(1/n) term is also unsupported.","section":"Section 4.2, Theorem 4.2 statement versus proof"},{"comment":"The paper's central empirical claim that SPIRE 'vastly outperforms' baselines on new clients rests on single KID values with no error bars, no number of seeds, and no significance test. KID estimates are stochastic and depend on the choice of reference batches, so the reported margins (e.g., 0.012 versus 0.017 on MNIST, or 0.038 versus 0.048 on CIFAR-10) are not interpretable without variance information. The absence of code and of a precise experimental configuration further limits verification of the empirical findings.","section":"Section 5, Table 1"}],"minor_comments":[{"comment":"The sentence beginning 'F orm of θ_j' at the start of Section 3 contains a formatting artifact and should read 'Form of θ_j'.","section":"Section 3"},{"comment":"The lemma and its proof disagree on the sign of the log-odds term: the lemma writes tanh(μ_t^T x - 1/2 log(w/(1-w))), while the proof derives tanh(μ_t^T x + 1/2 log(w/(1-w))) = tanh(μ_t^T x - 1/2 log((1-w)/w)); this notation needs to be made consistent throughout the paper.","section":"Lemma 4.1 and its proof"},{"comment":"The noise variance is written inconsistently as β_t in Eq. (2), as sqrt(1-e^{-2t}) in Eq. (9), and as both in the proof of Theorem 4.3; the relationship between these notations should be stated explicitly.","section":"Equations (2), (9), and the proof of Theorem 4.3"},{"comment":"The derivative computation in the proof contains duplicated lines and a sign error in the final expression, which suggests an editing mistake; the entire derivation should be rewritten and checked.","section":"Proof of Theorem 4.2"},{"comment":"The paper does not state how many random seeds or runs produced the reported KID values, nor does it provide code; this information is essential for evaluating the robustness claims.","section":"Experimental details"}],"recommendation":"reject","confidential_remarks":"The theoretical core of this paper is not salvageable by minor edits: the proof of Theorem 4.2 miscomputes the chain rule, the theorem statement claims a gradient-descent result that the proof does not deliver, and the estimator derived from Eq. (12) estimates 1-w rather than w. The empirical section is also under-powered, with single KID values and no significance statements. A resubmission would need either a valid gradient-descent analysis of the DDPM objective or a substantial reframing of the theoretical contribution as a moment-estimator bound, together with a more rigorous empirical evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: SPIRE has a sensible empirical idea, but the theory section's main theorem has a sign/chain-rule error, so the paper's headline claim doesn't hold as written.\n\nThe genuinely new bit is the conditional-embedding architecture for federated diffusion: a shared backbone plus per-client identity embeddings, with new clients fine-tuning only the embedding. That's a natural and practical way to do on-device personalization, and the new-client fine-tuning scenario is not something I see in the cited related work. The experimental section compares SPIRE against FedAvg+FT, a shared-representation baseline, and meta-learning on MNIST, CIFAR-10, and CelebA, and SPIRE does best or tied on most of the reported KID numbers. The robustness to fine-tuning epochs (Figure 4) is also a useful practical point. I also appreciate the frank limitation statement in the conclusion: the method requires changing the pretraining recipe, so it won't drop into an already-trained Stable Diffusion.\n\nThe soft spot is load-bearing. In the proof of Theorem 4.2, when they take the derivative of the DDPM loss with respect to the mixing weight, they drop the tanh'(u) factor and mis-handle the log-derivative sign. The correct stationarity condition involves E[tanh'(u)(||mu||^2 tanh(u) + mu^T X)] = 0, not E[tanh(u)] = -mu^T E[X]/||mu||^2. The sign is not cosmetic: at the true mixing weight w, the paper's equation forces w=1/2, and the estimator they define from it converges to 1-w. So the dimension-free error bound is not established, and Theorem 4.3 inherits the problem. This is an internal contradiction, not a missing detail.\n\nThe empirical evidence is also thinner than the abstract suggests: Table 1 gives single KID values with no error bars, no seeds, and no code. I can't tell whether the 'vastly outperforms' claim is robust across runs. That's a moderate issue, but it matters when the performance gap is the second big claim.\n\nIf you ask me, the right verdict is reject the current version. The practical idea is worth serious referee time, though. I'd recommend the authors either fix the theory (maybe a corrected moment condition still yields a useful bound) or explicitly demote the theory to a heuristic remark and lean on a properly replicated empirical study. A major revision could be publishable; the current submission overclaims.","headline":"SPIRE's empirical idea is plausible, but the theory has a sign/chain-rule error that invalidates the dimension-free bound, so the current version overclaims.","tokens_in":17110,"tokens_out":4887,"would_cite":false,"duration_ms":53098,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A shared diffusion backbone plus per-client identity embeddings lets new clients personalize by updating under 0.01 percent of weights.","keywords":["federated learning","diffusion models","personalization","conditional generation","Gaussian mixture models","parameter-efficient fine-tuning","score matching","client embeddings"],"falsifier":"Train a two-component symmetric Gaussian mixture with known means and mixing weight $w$, then run gradient descent on the DDPM loss in the one-layer parameterization of Lemma 4.1 and check whether the converged $w$ satisfies the critical-point condition $\\mathbb{E}[\\tanh(\\mu_t^\\top X_t-\\frac12\\log((1-w)/w))]=-\\mu_t^\\top\\mathbb{E}[X_t]/\\|\\mu_t\\|^2$; a systematic mismatch between the converged weight and the moment-matching value would invalidate the claimed link between DDPM training and maximum likelihood.","tokens_in":16022,"feed_emoji":"🖼️","tokens_out":8135,"duration_ms":195494,"temperature":0.7,"pith_summary":"SPIRE proposes a way to personalize a federated diffusion model without retraining the huge shared part: keep a global backbone that learns the population-level score, and attach to each client a lightweight identity embedding that is trained locally and used as conditioning. The paper's central claim is that this split lets a new client—one that never joined pretraining—adapt with only a few hundred updated parameters, under 0.01% of the model, with no extra communication. On MNIST, CIFAR-10, and CelebA the method matches or beats baselines during collaborative pretraining and outperforms them on new-client fine-tuning. The theoretical part argues that for a two-component Gaussian mixture with shared means and per-client mixing weights, gradient descent on the DDPM objective recovers the optimal mixing weight with a dimension-free error bound, giving a first formal bridge between conditional diffusion training and maximum likelihood in Gaussian mixtures.","feed_headline":"New clients personalize a diffusion model with under 0.01% of weights","feed_subtitle":"A client-specific embedding steers a shared score network, so new users adapt on-device in a few hundred parameters.","key_machinery":"The central objects are the identity embedding and the one-layer score function of Lemma 4.1. The embedding is a low-dimensional vector, indexed by client id, that is injected into a U-Net backbone as conditioning; it carries all client-specific information while the backbone weights stay shared. The lemma shows that the exact score of a two-component symmetric Gaussian mixture is $\\tanh(\\mu_t^\\top x-\\tfrac12\\log((1-w)/w))\\mu_t-x$, which is a single tanh neuron with a residual connection, where the mixing weight $w$ appears only as a bias in the pre-activation. That makes $w$ the “embedding” of the theoretical model, and it lets the authors relate the stationary point of the DDPM loss to the EM algorithm's M-step, transferring standard GMM learning guarantees to conditional diffusion training. For the practical method, the same mechanism—a per-client bias steering a shared network—is what Algorithm 1 and Algorithm 2 implement.","core_discovery":"On the paper's own terms, per-client diffusion modeling is a conditional generation problem, not a fine-tuning problem. The authors claim that factorizing the score network into a high-capacity shared backbone plus per-client embeddings makes personalization sample-efficient: the embeddings encode client data statistics, and for a client joining after pretraining, the only learned parameters are those embeddings. The theory they offer is that in a symmetric two-component Gaussian mixture where clients share the means but differ in the component mixing weight, the score has the closed form\n$$\\nabla_x \\log q_t(x)=\\tanh\\!\\left(\\mu_t^\\top x-\\tfrac12\\log\\!\\tfrac{1-w}{w}\\right)\\mu_t-x,$$\nand gradient descent on the DDPM objective with respect to $w$ reaches a point satisfying first-order moment matching, so the sample-based estimate obeys\n$$\\mathbb{E}[(w-\\hat w)^2]\\le \\frac{w(1-w)}{n}+\\frac{d}{4\\|\\mu_t\\|^2 n}.$$\nThe claim is that this bound is dimension-free in the sense that both terms scale the same way with dimension, and that it explains why a tiny conditioning signal can steer a shared score network toward a personalized distribution.","pith_inferences":["Beyond the paper's claims: if the embedding is read as a learned log-odds term, then any conditioning interface that biases a score network's pre-activations—class labels, text prompts, style tokens—could inherit a similar sample-efficiency argument, not just client IDs.","Beyond the paper's claims: the dimension-free weight bound suggests that the main bottleneck for personalization is the quality of the shared backbone rather than the client's local data size; this could be tested by fixing the client distribution and varying only the pretraining corpus.","Beyond the paper's claims: the paper states the limitation that SPIRE requires pretraining with client conditioning from scratch, which points to a direct extension—add a small trainable conditioning adapter to a publicly pretrained unconditional diffusion model and check whether new-client gains survive while the backbone stays frozen.","Beyond the paper's claims: since the theory is built on the exact tanh score form, a natural stress test is whether real U-Net embeddings actually behave like learned log-odds, for example by training SPIRE on a synthetic two-component GMM and comparing the inferred weight against the moment-matching value."],"forward_implications":["New clients can be personalized on-device after a single model download, since only their embedding is updated and no gradients need to be sent back.","Because the personalized part is tiny, fine-tuning is robust to learning-rate and epoch choices and does not catastrophically forget the shared backbone.","The theory predicts that clients with very few samples can estimate their own mixing weight reliably in high dimension, as long as the shared mean or backbone is well estimated.","Score-estimation error is dominated by the global mean estimate until the number of clients is very large (on the order of $d^6B^8$), after which the client's local sample size takes over (Theorem 4.3).","The same architecture achieves lower Kernel Inception Distance than whole-model fine-tuning or per-client bottleneck fine-tuning on all three tested datasets for new-client adaptation."],"supporting_citations":[{"why":"Defines the DDPM and score-matching objectives and supplies the mean-estimation procedure whose error enters Theorem 4.3.","marker":"[29]"},{"why":"Provides the EM updates for mixtures of two Gaussians that the proof uses to equate score-loss convergence with the M-step.","marker":"[18]"},{"why":"Establishes the equivalence between the score-matching objective and the DDPM objective used as SPIRE's training loss.","marker":"[4]"},{"why":"The meta-learning personalized federated learning baseline that SPIRE compares against empirically.","marker":"[12]"},{"why":"The shared-representation baseline, whose vertical split is adapted into a U-Net partitioning as a competing method.","marker":"[6]"},{"why":"The CelebA dataset used in the experiments for new-client personalization.","marker":"[21]"}],"fun_headline_variants":["One tiny embedding personalizes federated diffusion","Less than 0.01% weights adapt diffusion to new clients","SPIRE: conditional generation for on-device diffusion","Diffusion personalization with a few hundred parameters","Federated diffusion: share backbone, tweak embeddings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the learned score function has exactly the one-layer tanh form of Lemma 4.1, so that wherever the DDPM training loss levels off, the client parameter sits at the first-order moment matching point; if real U-Net scores deviate from that form, the dimension-free bound in Theorem 4.2 does not directly apply.","fun_headline_variants_meta":{"raw":{"variants":["One tiny embedding personalizes federated diffusion","Less than 0.01% weights adapt diffusion to new clients","SPIRE: conditional generation for on-device diffusion","Diffusion personalization with a few hundred parameters","Federated diffusion: share backbone, tweak embeddings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000631,"raw_usage":{"total_tokens":2952,"prompt_tokens":1023,"completion_tokens":1929,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":1852}},"tokens_in":639,"tokens_out":1929,"duration_ms":16590,"temperature":1.0,"reasoning_tokens":1852,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:53:34.157955+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a two-component symmetric Gaussian mixture with known means and mixing weight $w$, then run gradient descent on the DDPM loss in the one-layer parameterization of Lemma 4.1 and check whether the converged $w$ satisfies the critical-point condition $\\mathbb{E}[\\tanh(\\mu_t^\\top X_t-\\frac12\\log((1-w)/w))]=-\\mu_t^\\top\\mathbb{E}[X_t]/\\|\\mu_t\\|^2$; a systematic mismatch between the converged weight and the moment-matching value would invalidate the claimed link between DDPM training and maximum likelihood.","supporting_citations":[{"cited_title":"Learning mixtures of gaussians using the DDPM objective","cited_arxiv_id":null,"evidence_quote":"Defines the DDPM and score-matching objectives and supplies the mean-estimation procedure whose error enters Theorem 4.3."},{"cited_title":"The em algorithm gives sample-optimality for learning mixtures of well-separated gaussians, 2020","cited_arxiv_id":null,"evidence_quote":"Provides the EM updates for mixtures of two Gaussians that the proof uses to equate score-loss convergence with the M-step."},{"cited_title":"Personalized federated learning: A meta- learning approach","cited_arxiv_id":null,"evidence_quote":"The meta-learning personalized federated learning baseline that SPIRE compares against empirically."},{"cited_title":"Exploiting shared represen- tations for personalized federated learning","cited_arxiv_id":null,"evidence_quote":"The shared-representation baseline, whose vertical split is adapted into a U-Net partitioning as a competing method."}],"review_version":1}