{"id":"6cae780b-eed1-413f-a3dc-90ea86891676","arxiv_id":"2411.09263","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Weight-averaged model merging is reinterpreted as template matching and implicit regularization, with systematic experiments showing logits ensembling generally outperforms weight averaging and ViTs degrade sharply under naive weight averaging.","lead":"This paper tries to explain why averaging the weights of several trained models often works, by looking at the patterns inside the weights and comparing weight averaging with feature averaging. It finds that ensembling features usually beats weight averaging, that vision transformers are especially bad for naively averaged weights, and that averaging acts a bit like regularizing the model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The regularization explanation for weight-averaged merging is not established: Eq. 5–11 supply upper bounds, not the claimed monotonic effect, and the paper never measures actual magnitude/variance reduction in real merges.","rationale":"The paper has real value as an empirical study: the weight/feature averaging comparison across architectures and datasets, and the magnification robustness curves, are useful engineering observations. My concern is targeted at the explanatory contribution, which is the central reason the paper claims interpretability. The authors assert a theoretical explanation ('model merging implicitly serves as a form of regularization') and support it with bounds that are mathematically true but logically insufficient: an upper bound on output variance that grows with weight scale does not establish that increasing weight scale increases actual output variance, and a non-increase bound on the max norm does not establish that averaging reduces it. The missing covariance term in Eq. 8 is repairable, but as written the derivation is incorrect. The absence of a direct measurement of actual merged weight magnitude/variance means the proposed mechanism is not empirically anchored either. This is exactly the kind of soft spot that should be fixed before the theoretical claims are taken as established. Because the reader already flagged the flawed proof and the compatibility assumption, and because the empirical core can stand if the theory is reframed as a hypothesis, I do not move the verdict; conditional acceptance remains the right call.","tokens_in":18960,"tokens_out":8006,"duration_ms":97779,"concrete_test":"Train 10 VGG19 models on CIFAR-100. For the uniform soup and each constituent, measure per-layer max-abs weight norm and elementwise variance, and output norm/variance on a fixed test batch before and after merging. Separately scale constituents by c ∈ {0.1, 0.5, 1, 2, 10, 50, 100} and check whether empirical output variance is monotone non-decreasing in c. If the merged model is not smaller in weight magnitude/variance, or if output variance does not grow with c, the regularization mechanism of Proposal 3 fails. If both hold, the empirical conclusion survives, but a corrected proof (lower bound or exact formula) is still needed to support the causal claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that weight averaging implicitly regularizes by reducing weight magnitude and variance, and that this explains when merging helps generalization. This claim requires (i) averaging actually reduces parameter magnitude/variance and (ii) smaller magnitude/variance makes outputs more stable. Link (i) is only a non-increase bound: Eq. 5–6 give ∥(W1+W2)/2∥ ≤ max(∥W1∥,∥W2∥), so the text's 'typically reduces' is not proved. Eq. 8 derives Var((W1+W2)/2) = 1/4(σ1²+σ2²) without a covariance term, i.e., it assumes independence of W1 and W2; real trained models are not independent. Link (ii) is the more serious gap: Property 1 (Eq. 10) and Theorem 1 (Eq. 11) are upper bounds on output norm and variance. The statement that increments in weight magnitude/variance 'will also magnify outputs' magnitudes/variances' does not follow from upper bounds; cancellations can make actual output variance decrease as weights grow. The magnification experiments impose artificial factors (e.g., ×100) and compare merging vs ensembling, but they never measure whether the actual merged model has lower weight magnitude/variance than its constituents. Thus Proposal 3 and the abstract's regularization claim are unsupported by the supplied mathematics and data.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an interpretability-oriented analysis of weight-averaged model merging. The authors offer three perspectives: (1) interpreting learned weights as class templates, so that weight averaging acts like a Mixup operation in parameter space; (2) empirically comparing weight-space averaging (model merging) with feature-space averaging (model ensembling) across CNNs and ViTs on several datasets; and (3) arguing that merging implicitly regularizes by reducing weight magnitudes and variances, making predictions more robust. The paper concludes with three 'Proposals' that blend these observations into practical guidance. The empirical sections report accuracy comparisons for uniform and greedy soups versus logit/feature ensembles, plus experiments where parameter magnitudes are artificially scaled to study robustness.","tokens_in":19174,"tokens_out":5688,"duration_ms":56561,"significance":"If the central claims were established, the paper would offer a useful intuitive account of why model soups work, a broad empirical map of when merging versus ensembling is preferable, and a concrete robustness mechanism. The empirical scope is genuinely broad: multiple architectures (ResNet, DenseNet, VGG, ViT, DeiT), multiple datasets, and model counts from 2 to 10. The template visualizations are a compelling pedagogical device. However, the theoretical backbone is not sound as written: the variance derivation omits a covariance term, the magnitude argument proves only a non-increase bound rather than 'typical reduction', and the output-magnitude/variance conclusions are not implied by the stated upper bounds. The paper never directly measures whether real merged models have lower weight magnitude or variance than their constituents, so the regularization narrative—which is the paper's central explanatory claim—remains unsupported. The comparative empirical results would still be of interest, but the interpretability thesis needs substantial reworking.","major_comments":[{"comment":"The variance formula Var((W1+W2)/2) = 1/4(σ1² + σ2²) silently omits the covariance term. For random matrices the correct expression is 1/4(σ1² + σ2² + 2Cov(W1,W2)). Since independently trained networks typically have correlated weights, the subsequent claim that averaging 'always' reduces variance relative to the larger member is not derived. The inequality (9) can be salvaged for any covariance with correlation ≤1, but the derivation as printed is invalid and the assumptions must be stated explicitly, or the general result must be proved.","section":"The Model Predictions and Weight Magnitudes, Eq. (8)"},{"comment":"Equations (5)–(6) establish only the non-increase bound ∥(W1+W2)/2∥ ≤ max(∥W1∥,∥W2∥). The text then asserts that averaging 'typically reduces' the maximum weight magnitude. A non-increase bound does not support a typical-reduction claim; equality occurs when the maximum is attained at the same entry in both matrices, and even near-equal maxima give a reduction only by happenstance. If the authors intend an empirical 'typically', they should measure the actual L∞ norms of merged versus individual models across their experimental settings rather than rely solely on the bound.","section":"Analysis of Weight Magnitude After Averaging, Eqs. (5)–(6)"},{"comment":"Property 1 and Theorem 1 are upper bounds on output norm and variance under i.i.d. sub-Gaussian assumptions on the weights. The statement that 'the increment of model weights' magnitudes/variances will also magnify outputs' magnitudes/variances' does not follow from upper bounds; monotonicity is not a logical consequence, because cancellations in the weighted sum can make output variance decrease even when weight variance increases. The magnification experiments (Figure 7b, Table 10) artificially scale every parameter by ×90–×100 but never compare the actual weight magnitude/variance of the merged model against its constituents. Consequently, Proposal 3 and the abstract's claim that merging 'implicitly serves as a form of regularization' are not established by the supplied mathematics or experiments. To support this claim, the authors should directly measure norms and variances of real merged weights and test whether output stability actually improves in the un-scaled setting.","section":"How Weight Magnitudes and Variance Effects Model Outputs, Property 1 and Theorem 1 (Eqs. (10)–(11))"},{"comment":"The equivalence between weight averaging and feature averaging in Eq. (1) holds only under the assumption of 'no symmetric neuron mismatch issue.' The paper extends the 'meaningful linear combination' and Mixup-like template interpretation to deep networks without verifying that the trained models lie in the same loss basin or are permutation-compatible. The paper's own results show severe violations of this condition: uniform soups on PathMNIST collapse from 90.04% to 32.42% (Table 1), and DeiT-Tiny on TinyImageNet falls to 8.91% with 10 models (Supplementary Table 9). In these cases the template-matching explanation does not apply, yet the paper does not explicitly condition Proposal 1 on compatibility or discuss how its failures are predicted by the framework. Without an explicit compatibility condition and a discussion of the failure cases, the template narrative is not a falsifiable explanation of when merging will succeed.","section":"Linear Model Scenarios, Eq. (1), and the template narrative in Section 'The Patterns Contained in Model Weights'"},{"comment":"All reported accuracies are single runs without error bars or significance tests. Since the paper's central empirical message is that ensembling consistently outperforms merging and that gaps widen with more models (Figures 3 and 7a), the lack of repeated-seed statistics makes it impossible to assess whether these differences are reliable. Additionally, many training details (optimizer hyperparameters, epochs, augmentation, pretraining initialization, and the exact construction of the 'Ens Features' variant) are deferred to supplementary material without enough specification in the main text to reproduce the comparisons. At minimum, the authors should report mean±std over multiple seeds for the headline tables and figures.","section":"Experiments and Supplementary Tables 1, 3–7, 9"}],"minor_comments":[{"comment":"The abbreviations 'Grd Ens Lgt', 'Ens Fts', 'Perf Ave' are used inconsistently across captions, tables, and text; please define each abbreviation at first use and use it uniformly.","section":"Throughout"},{"comment":"The caption of Figure 3 refers to 'Performance Average' while the text and Table 1 use 'Perf Ave'; please unify the terminology.","section":"Figure 3 and Table 1"},{"comment":"The proof section refers to 'Eq. 14' but the main text numbers equations differently; cross-references between main text and supplementary equations should be corrected.","section":"Supplementary 'Theoretical Proofs'"},{"comment":"In Figure 4, the scale of the axes and the statement that points are 'computed on the same data batch' are unclear; please describe how the scatter points are generated.","section":"Section 'The Model Predictions and Weight Magnitudes'"},{"comment":"The entry '81.95' for 'Grd Ens Fts' on CIFAR-10 with 5 models appears inconsistent with the rest of the column (~92.6) and should be verified.","section":"Supplementary Table 5"},{"comment":"The citation for Property 1 and Theorem 1 to Wang et al. 2024 should be cited prominently in the main text, not only in the proof appendix, and the applicability of its security-oriented assumptions to trained vision models should be discussed.","section":"References"},{"comment":"The reference to model arithmetic uses Mikolov 2013 (word2vec) as a catch-all; please cite the specific model-arithmetic works (e.g., Ilharco et al. 2022, Zhou et al. 2024b) in that sentence.","section":"Related Work and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical scope is a strength, but the interpretability thesis is currently overclaimed relative to the evidence. The variance formula, the 'typical reduction' claim, and the monotonicity of output variance with weight scale all need correction or explicit empirical verification. The authors should be pushed to measure the actual weight magnitudes and variances in their merge experiments and to condition the template explanation on permutation/linear-mode compatibility. If those gaps are addressed, the paper could become a useful interpretability study; as it stands, the central mechanism is unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a useful empirical study wrapped in an overclaimed theoretical story. The cross-architecture and cross-dataset comparison of weight averaging vs feature averaging is the real contribution; the regularization explanation, as written, is not established by the paper's own math.\n\nWhat's actually new: the systematic sweep is broader than prior work—CNNs and ViTs, seven datasets, 2 to 10 models, uniform and greedy soups versus logit/feature ensembles. The finding that uniform soups collapse on PathMNIST and CelebA (90% to 32%, 92% to 53%) and that DeiT-Tiny degrades to 8.9% with 10 models is a concrete, reproducible cautionary result. The template visualizations are nice didactic material, and the Mixup analogy is evocative, though it was already floating around in earlier work.\n\nThe soft spots are in the third section. Eq. (8) drops the covariance between W1 and W2; that is only valid under independence, which trained weights don't satisfy. Eq. (9) is an upper bound, not a proof that averaging reduces variance. Property 1 and Theorem 1 are upper bounds on output norm/variance imported from a security paper; the text's claim that they 'show that the increment of model weights' magnitudes/variances will also magnify outputs' magnitudes/variances' does not follow. Cancellations can make output variance decrease even as weights grow. The magnification experiment is interesting, but it never measures whether the actual merged model has lower weight norm or variance than its constituents—which is the direct link Proposal 3 needs. And the compatibility assumption (no neuron mismatch) is stated for the linear case and silently carried into deep networks; the catastrophic soup failures on PathMNIST and DeiT are precisely cases where that assumption fails, so the template explanation doesn't cover the interesting negative cases.\n\nThe missing error bars and training details (optimizer, epochs, seeds) are a real reproducibility concern for a paper making empirical claims. The code is public, which helps, but 10 models with no variance reporting is thin.\n\nBottom line: the empirical findings deserve to be in the literature, and practitioners will get value from the comparison tables. The theoretical section should be reframed as conjecture, with the upper-bound issue fixed and the covariance term corrected. I'd send it to peer review and ask for major revision. It's not a desk reject.","headline":"Worth engaging for the empirical comparison, but the regularization explanation is a conjecture, not a proof.","tokens_in":19755,"tokens_out":3115,"would_cite":true,"duration_ms":31349,"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":"Weight-averaged model merging works because weights are class-specific templates and averaging acts as an implicit regularizer that shrinks weight magnitude and variance, explaining both its successes and its failures.","keywords":["model merging","weight averaging","model soups","interpretability","template matching","regularization","feature ensembling","vision transformers"],"falsifier":"Train the same architecture on the same task from different random seeds, measure the linear-mode-connectivity barrier between every pair, and check whether the benefits of uniform weight averaging disappear exactly when the barrier is high. A sharper test is to permute the neurons of one model before averaging, which preserves its function but destroys neuron compatibility, and verify that merging accuracy falls to the same catastrophic levels as the PathMNIST and DeiT-Tiny collapses.","tokens_in":18717,"feed_emoji":"🧩","tokens_out":8273,"duration_ms":72722,"temperature":0.7,"pith_summary":"This paper tries to establish why weight-averaged model merging—combining the parameters of several trained models into one—works as well as it does. It argues that weights act as class-specific templates, so averaging weights is a meaningful linear combination akin to Mixup applied in weight space, and that averaging shrinks weight magnitude and variance, acting as an implicit regularizer. From these two mechanisms the paper derives concrete predictions: merging helps when models are neuron-compatible, is cheaper than ensembling, and fails for vision transformers and mismatched tasks, where uniform averaging collapses. A sympathetic reader would care because it turns a heuristic trick into a testable account of when untrained model combination is reliable.","feed_headline":"Averaging model weights blends templates and stabilizes predictions","feed_subtitle":"Two mechanisms explain it: weights act as class templates, and averaging shrinks weight variance.","key_machinery":"The load-bearing objects are: (1) the template interpretation of weights, where a row of the weight matrix is a class-specific pattern matched to an input by the inner product $w^\\top x = \\lVert w\\rVert\\lVert x\\rVert\\cos\\theta$, so merging two weight matrices is a Mixup-like linear combination of templates; (2) the linear-equivalence identity $h = \\tfrac{1}{2}(W_1+W_2)x = \\tfrac{1}{2}(W_1x + W_2x)$, which holds only without non-linearities and breaks under activations such as ReLU; (3) the norm bound $\\lVert (W_1+W_2)/2\\rVert \\le \\max(\\lVert W_1\\rVert,\\lVert W_2\\rVert)$ from the triangle inequality and the variance identity $\\operatorname{Var}((W_1+W_2)/2) = \\frac{1}{4}(\\sigma_1^2+\\sigma_2^2) \\le \\max(\\sigma_1^2,\\sigma_2^2)$, which together show that averaging shrinks weight magnitude and variance; and (4) the Lipschitz bounds (Property 1 and Theorem 1) that connect smaller weight magnitudes and variances to smaller, stabler outputs. These pieces carry the argument from template structure to regularization to the observed architecture- and dataset-dependent behavior.","core_discovery":"The paper's central claim is that weight-averaged model merging is not arbitrary interpolation but a structured operation on learned templates. Each row of a classifier's weight matrix encodes a class prototype, and inner products match inputs to those prototypes; averaging two weight matrices therefore blends prototype sets the way Mixup blends images, giving the merged model a broader set of patterns to match. At the same time, averaging provably reduces the maximum weight magnitude and the variance of weights, and via the paper's Lipschitz bounds (Property 1 and Theorem 1) this shrinks the magnitude and variance of the network's outputs, making predictions less sensitive to input perturbations. The paper concludes that merging is implicitly a regularizer that trades expressiveness for stability, which is why it can beat individual models on compatible same-task models but degrades sharply for ViTs, for models trained on mismatched tasks, and on datasets like PathMNIST and CelebA where weight conflicts dominate.","pith_inferences":["I infer that the regularization effect of merging could be repurposed as a no-training defense: averaging weights of models trained with different augmentations should suppress input sensitivity, a prediction that could be tested under adversarial or label-noise perturbations.","The template/Mixup analogy suggests a data-free probe for class similarity: interpolating two class templates in weight space and measuring which test images the merged template activates could reveal semantic overlap between classes.","One extension the paper leaves implicit is that permutation-alignment should be a mandatory pre-processing step before averaging ViT weights; the observed DeiT collapse predicts that aligned ViTs would recover soup gains.","The merging–ensembling gap peaks at a particular magnification factor (×100 on CIFAR-100, ×110 on CIFAR-10), suggesting a sweet spot where weight shrinkage still helps without destroying expressiveness; a practical follow-up is to rescale averaged weights to a target norm before deployment."],"forward_implications":["Because averaging shrinks weight magnitude and variance, merged models should be more stable than ensembles under weight perturbation, which the paper's scaling experiments (up to ×100) confirm.","Merging should be attempted only among neuron-compatible models; greedy soups, which select models by validation performance, recover much of the loss that uniform soups suffer on incompatible models such as PathMNIST (from 32.42% back to 88.48%).","For vision transformers, uniform weight averaging is ineffective and worsens as more models are added, so logit or feature ensembling is the safer strategy for those architectures.","Because merging acts as a regularizer, its benefits should be largest when individual models are overfit or noisy and smallest when they are already well-regularized.","The template view implies that merging across tasks produces composite templates (for example bicycle plus bottle), which is why cross-task soups can help on one task while sacrificing fine-grained class structure."],"supporting_citations":[{"why":"Introduces the weight-averaged merging (model soups) procedure whose behavior this paper reinterprets.","marker":"Wortsman et al. 2022"},{"why":"Supplies the Mixup operation that the paper uses as the analogy for weight-space interpolation of templates.","marker":"Zhang 2017"},{"why":"Provides the CIFAR-100 dataset used for the template visualization and merging experiments.","marker":"Krizhevsky 2009"},{"why":"Provides the Tiny ImageNet dataset used for template visualizations and for the ViT merging failure experiments.","marker":"Le and Yang 2015"},{"why":"Source of Property 1 and Theorem 1, the Lipschitz bounds connecting weight magnitude/variance to output magnitude/variance.","marker":"Wang et al. 2024"},{"why":"Provides the triangle-inequality framework used to prove that averaging shrinks maximum weight magnitude.","marker":"Tversky and Gati 1982"}],"fun_headline_variants":["Weight averaging blends class templates, stabilizes predictions","Why weight merging works: template interpolation plus variance shrink","Merged weights mix object templates and stabilize via variance shrink","Model merging explained: weights as templates, averaging as stabilizer","Weight-averaged merging blends prototypes and reduces output variance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire template and regularization story assumes the models being averaged are neuron-compatible, meaning they sit in the same loss basin without permutation mismatch; the paper's own experiments show that when this fails, uniform averaging collapses, with accuracy dropping to 32.42% on PathMNIST and 8.91% for DeiT-Tiny on TinyImageNet.","fun_headline_variants_meta":{"raw":{"variants":["Weight averaging blends class templates, stabilizes predictions","Why weight merging works: template interpolation plus variance shrink","Merged weights mix object templates and stabilize via variance shrink","Model merging explained: weights as templates, averaging as stabilizer","Weight-averaged merging blends prototypes and reduces output variance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1559,"prompt_tokens":925,"completion_tokens":634,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":554}},"tokens_in":541,"tokens_out":634,"duration_ms":7463,"temperature":1.0,"reasoning_tokens":554,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:50:19.747822+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture on the same task from different random seeds, measure the linear-mode-connectivity barrier between every pair, and check whether the benefits of uniform weight averaging disappear exactly when the barrier is high. A sharper test is to permute the neurons of one model before averaging, which preserves its function but destroys neuron compatibility, and verify that merging accuracy falls to the same catastrophic levels as the PathMNIST and DeiT-Tiny collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of Property 1 and Theorem 1, the Lipschitz bounds connecting weight magnitude/variance to output magnitude/variance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the triangle-inequality framework used to prove that averaging shrinks maximum weight magnitude."}],"review_version":1}