{"id":"2ce79144-3f09-4c31-8461-d97d56daf734","arxiv_id":"2412.08147","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using more flexible Bayesian posterior approximations in model merging gives better previews of multitask finetuning weightings on vision and language transformers.","lead":"Model merging can preview how different task-weighting choices would perform in multitask finetuning, without retraining for each weight. This paper shows that Bayesian posteriors, especially mixtures of Gaussians, make those previews more accurate on vision and language transformers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'necessarily better' claim is not derived and is false in general; variational KL minimization does not guarantee monotonic improvement of the merged-posterior mode.","rationale":"The reader's weakest_assumption identifies the same gap: the paper assumes that minimizing KL[q_t || p_t] ensures the surrogate is accurate enough that the merged mode bθ_α tracks the true weighted-loss minimizer. My concern sharpens this: even if each q_t has low KL to p_t, the mode of the weighted product ∏ q_t^{α_t} is not controlled by the individual KL values, and increasing the flexibility of q_t can, in principle, move that mode in the wrong direction for some α. This is the core load-bearing issue because the paper's headline claim is a deterministic, universal statement ('necessarily'), while the evidence is empirical and the proof is absent. The concrete double-well test is the minimal experiment that would settle whether the claim can hold in a setting with multimodality, which is the most natural failure regime. The paper has real independent support: the generalized-Bayesian identification in Eq. 5–6 is exact for exact posteriors, the Hessian-weighted derivation in App. A.1 is sound, and the empirical trends in Tables 1–2 are consistent and across several model families. These strengths justify keeping the reader's CONDITIONAL verdict: the practical method is plausible and useful as a preview tool, but the theoretical claim must either be proven under explicit assumptions or softened to match the empirical scope.","tokens_in":22461,"tokens_out":8552,"duration_ms":90501,"concrete_test":"Construct a two-task 1D problem with an asymmetric double-well loss for task 1, ℓ1(θ) = ((θ−1)(θ+2))², and a quadratic loss for task 2, ℓ2(θ) = θ². For α ∈ {0, 0.05, …, 1}, compute the true weighted-loss minimizer θ*_α and its performance vector. Train single-task models from multiple random initializations (covering both wells) to obtain checkpoints. Build previews with simple isotropic merging (Eq. 9), full-Gaussian Hessian-weighted merging (Eq. 10), and mixture-of-Gaussians merging (Alg. 1 with K=10 and 20 EM restarts). Report the normalized MSE of predicted performances against true performances, averaged over the α grid and over 100 random seeds. If the mixture method does not have strictly lower MSE than the full-Gaussian method, or if for any α its predicted performance is farther from the truth, the 'necessarily better' claim is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central theoretical claim is that 'more flexible posteriors necessarily yield better estimates of Pareto fronts.' No theorem or proof is provided in the full text; the only support is the heuristic argument in Sec. 3.3 that minimizing KL[q_t || p_t] makes surrogates 'globally accurate.' This is insufficient because the merged parameter bθ_α is the mode of the weighted product ∏_t q_t(θ)^{α_t} (Eq. 8, App. A.2), and the mode location depends on the interaction of surrogate tails, not merely on each q_t's high-probability regions. As the variational family expands (isotropic Gaussian → full Gaussian → mixture, Sec. 3.4), the minimizer of KL[q_t || p_t] can shift probability mass into regions with low true posterior mass that, after exponentiation by α_t and multiplication across tasks, dominate the product's mode. This non-monotonicity is a known phenomenon in variational inference and is not excluded by any assumption in Sec. 3.3. The paper's toy evaluation (App. B.1) uses log-sum-exp losses, which are convex and hence give unimodal posteriors; for multimodal task losses typical in deep learning, the monotonicity can break. Therefore the 'necessarily' claim is an overclaim: it is not a proven theorem and is plausibly false without additional assumptions such as log-concavity or exact membership of the true posterior in the chosen exponential family.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a Bayesian model-merging framework for generating cheap “previews” of multitask finetuning behavior across task-weight vectors α. Single-task finetuned models are converted into exponential-family or mixture-of-Gaussian posterior surrogates q_t, and the weighted multitask solution is approximated by the mode of the product ∏_t q_t(θ)^{α_t}. The paper recovers Task Arithmetic and Hessian-weighted merging as Gaussian special cases, adds an EM-based mixture-of-Gaussians merging algorithm, and evaluates preview quality on logistic regression, ResNet-20, ViT-B/32, RoBERTa, and GEMMA-2B experiments. The advertised main theoretical result is that more flexible posteriors necessarily yield better estimates of Pareto fronts.","tokens_in":22761,"tokens_out":7341,"duration_ms":70392,"significance":"The practical contribution is real and useful: the paper demonstrates with diverse vision and language models that merging can produce fast previews of reweighting behavior, and the hyperparameter details in the appendices make the empirical setup unusually transparent. The Bayesian interpretation also provides a clean unification of several existing merging methods and a principled route to new mixture-based merges. However, the headline theoretical claim is not established; the manuscript's actual support is empirical, and the claim as stated is both underdefined and not derived. If the authors rework the central claim into a proven statement under explicit assumptions or an explicitly empirical statement, the paper would be a solid contribution to the model-merging and multitask-finetuning literature.","major_comments":[{"comment":"The abstract's statement that “more flexible posteriors necessarily yield better estimates of Pareto fronts” is not a proven theorem in the manuscript, and the discussion in Sec. 3.3 does not establish it. Equation (7) minimizes D_KL[q_t || p_t], which controls each surrogate marginally, while the preview θ̂_α in Eq. (8) and Alg. 1 is the mode of the weighted product ∏_t q_t(θ)^{α_t}. A reduction in KL can shift probability mass or change the tails of q_t in ways that move the product's mode away from the true weighted minimizer θ_α for some α. The toy validation in App. B.1 uses log-sum-exp losses, which are convex and give unimodal posteriors, so it does not test the multimodal failure mode. The claim should be replaced either by a theorem proved under explicit assumptions (for example, log-concavity or exact membership of p_t in the chosen exponential family) or by an empirical/expected statement.","section":"Abstract; Sec. 3.3, Eqs. (7)–(8)"},{"comment":"The paper never defines the error of a “Pareto-front estimate” formally. The abstract's theoretical claim is therefore not attached to a precise quantity: Tables 1 and 2 report the MSE of predicted accuracies and the best-preview accuracy, not a Pareto-front metric under a stated definition. Without such a definition, the claimed comparison between posterior families is unfalsifiable, and the empirical “consistently” wording should be tied to the metric actually reported.","section":"Sec. 2–4; Tables 1 and 2"},{"comment":"The empirical results do not support “necessarily better” even informally. In the ResNet-20 row of Table 2, the Hessian-weighted preview finds a best α whose true multitask accuracy (64.6%) is below the value found by simple merging (67.4%), and the preview accuracy itself decreases from 62.4% to 62.3%. The paper's statement in Sec. 4 that “we see a consistent trend that the performance improves as we use better posteriors” is too strong for this row; it should be reported as mixed or approximately equal and reconciled with the theoretical claim.","section":"Table 2, Fig. 3 row; Sec. 4.1"},{"comment":"For mixture posteriors, the closed-form argument in App. A.2 does not apply, because App. A.3 explicitly states that the EM iterates converge only to a local maximum. The mode returned by Alg. 1 can therefore depend on initialization, number of components, and convergence tolerance. The claim that more flexible mixtures necessarily improve previews is especially unsupported without additional analysis of these local-mode effects, and the paper should state clearly that the mixture previews are heuristic rather than guaranteed.","section":"Sec. 3.4; App. A.3; Alg. 1"}],"minor_comments":[{"comment":"The manuscript contains two different titles—“Variational Model Merging for Pareto Front Estimation in Multitask Finetuning” in the arXiv header and “How to Weight Multitask Finetuning? Fast Previews via Bayesian Model-Merging” in the running text—and they should be aligned.","section":"Title page"},{"comment":"There are small typographical errors: “Hessian-Weigthed” and “Mixture-Weigthed” appear in Fig. 3, and App. C.1 contains “capture more the the exact solution.”","section":"Fig. 3; App. C.1"},{"comment":"The pairing between preview scores and the gray true scores is difficult to read in monochrome; an explicit column grouping or legend should indicate which values are previews and which are the corresponding multitask-finetuning results, and the table should state which α produced each value.","section":"Table 2"},{"comment":"The sentence “We can just use such optimizers to compute the posterior q_t” is vague about which optimizer (Adam, Adam-like, or IVON) solves Eq. (7) and under what approximation; please specify the exact relationship.","section":"Sec. 3.3"},{"comment":"The phrase “The pretrained and finetuned checkpoints of ViT-B-32 a model based on CLIP” contains a stray article and should be rewritten.","section":"App. B.4"}],"recommendation":"major_revision","confidential_remarks":"The paper sits between major revision and reject: the empirical study is broad and the framework is genuinely useful, but the advertised theoretical result is not derived and is contradicted by at least one of the paper's own table entries. I lean major_revision because the overclaim is fixable by weakening the theorem to an empirical claim or by proving it under explicit assumptions, and because the practical contribution would remain valuable after such a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the practical contribution is real and probably useful. The paper casts model merging as weighted minimization of surrogate losses derived from variational posteriors, recovers simple averaging, task arithmetic, and Hessian-weighted merging as special cases, and adds a mixture-of-Gaussians merging scheme with an EM mode-finding update. Evaluating merged models as cheap previews over the whole alpha simplex, rather than at a single best weight, is a genuinely useful framing. The experiments span logistic regression, ResNet-20, ViT-B/32, RoBERTa, and Gemma-2B with LoRA, and the trend is consistent: richer posteriors give previews that track joint finetuning better.\n\nThe soft spot is exactly where the stress test points. The abstract's 'necessarily better' claim is not proved anywhere. Section 3.3 only argues that minimizing KL makes surrogates globally accurate, and then the paper jumps to the conclusion. That inference doesn't hold: the merged parameter is the mode of a weighted product of approximate posteriors, and mode location can worsen as the variational family expands, especially with multimodal task losses. The toy example uses convex log-sum-exp losses, so it does not exercise the failure mode. To make the claim rigorous you would need assumptions like log-concavity or exact posterior membership in the family. Otherwise the honest statement is 'empirically better or comparable in our experiments.'\n\nOther concerns are addressable and moderate. No error bars anywhere; Tables 1 and 2 report point estimates only. The mixture method is demonstrated on small models but not on ViT/RoBERTa/GEMMA, so the scaling story is incomplete. No code release. And the two abstract versions disagree: the arXiv abstract promises a theorem that the full text does not state. That mismatch needs to be resolved.\n\nI want to be clear that the paper is not weak overall. The unification is clean, the connection to Bayesian committee machines and posterior merging is well placed, and the preview evaluation is a practical step forward. The citation pattern is fine; reliance on Daheim et al. and the Khan line of work is natural since the paper builds directly on those methods.\n\nRecommendation: send to serious peer review. The authors should either prove a precise theorem with stated assumptions or soften the abstract, add variance estimates, and release code. The framework is worth engaging with; the overclaim is fixable.","headline":"Useful Bayesian unification of model merging with a practical preview tool, but the headline 'necessarily better' claim is unproven and likely false without assumptions; the empirical trend is consistent, if under-reported.","tokens_in":23291,"tokens_out":2176,"would_cite":true,"duration_ms":23627,"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":"Bayesian model merging with flexible posteriors makes merged models track joint multitask training more faithfully across task weightings.","keywords":["multitask finetuning","model merging","Pareto front estimation","variational inference","posterior merging","exponential family","mixture of Gaussians","task weighting"],"falsifier":"Take a two-task problem with known losses, compute the true Pareto front by joint training on a fine $\\alpha$ grid, then compute isotropic, full-Gaussian and mixture previews from single-task models. If the full-Gaussian or mixture preview has higher mean squared error to the true front, or finds a best $\\alpha$ whose true performance is worse than the isotropic version's, the claimed monotone improvement is falsified for that setting.","tokens_in":1732,"feed_emoji":"⚖️","tokens_out":2395,"duration_ms":77806,"temperature":0.7,"pith_summary":"The paper wants to make the search over task weights in multitask finetuning cheap by estimating the whole performance tradeoff surface without retraining. Its proposal, Variational Model Merging, treats each separately finetuned model's posterior as a surrogate for that task's loss, then forms merged parameters for any weighting $\\alpha$ by taking the mode of the weighted product of posteriors. The central claim is that the more flexible the posterior family used per task, the better the merged preview tracks the performance of models actually finetuned jointly, across the full range of $\\alpha$ values. Simple averaging and Task Arithmetic fall out as isotropic-Gaussian special cases, and the paper validates the ordering on vision and language transformers. If the claim is right, practitioners can screen task weightings in seconds to minutes instead of running many joint-training experiments.","feed_headline":"More flexible Bayesian mergers yield better task-weight previews","feed_subtitle":"Variational model merging makes single-task finetunes trace joint-training tradeoffs without hours of retraining.","key_machinery":"The engine is posterior merging: the identity $p_\\alpha \\propto p_0^{\\gamma}\\prod_t p_t^{\\alpha_t}$ means weighted-multitask inference can be obtained by combining per-task posteriors instead of retraining. Per-task posteriors are fitted variationally as exponential-family or mixture-of-Gaussian distributions, and the merged parameter is the mode of the weighted product (Eq. 8), which is closed form for exponential families. Algorithm 1 iterates an EM-style fixed point for mixture-of-Gaussians that generalizes a classical mode-finding procedure. Replacing each task loss by $-\\log q_t$ is what makes the surrogates accurate over a region rather than at a single point, and the flexibility of the posterior family controls how much of the relevant region is covered.","core_discovery":"The paper's claim is that Pareto-front estimation by model merging is a Bayesian posterior-merging problem, and that moving to more expressive posterior families strictly helps. Given task losses $\\ell_t$ and a prior $p_0$, each single-task posterior $p_t$ is approximated variationally by $q_t$ in an exponential family; merging with weights $\\alpha$ gives $\\hat{\\theta}_\\alpha$ as the mode of the product $p_0^{\\gamma}\\prod_t q_t^{\\alpha_t}$. For Gaussian $q_t$ this reduces to Hessian-weighted merging, and for mixtures the paper derives an EM fixed-point algorithm. The theoretical result stated is that more flexible posteriors necessarily yield better estimates of the Pareto front: a full-Gaussian merge is expected to beat an isotropic one, and a mixture to beat a single Gaussian. Experiments on logistic regression, ResNet-20, ViT-B/32, RoBERTa and LoRA-finetuned GEMMA-2B show previews improving in this order while staying far cheaper than joint training.","pith_inferences":["A concrete test of the 'necessarily better' claim would be to measure whether lower variational KL for each task always translates into lower preview error on the $\\alpha$ grid; the paper reports MSE but does not tie the two numbers together per $\\alpha$, so a systematic correlation study would separate the theorem from the approximation.","The monotonicity may fail for utility functions that care about one task's worst-case behavior, because a more flexible posterior can concentrate mass in a region that is globally accurate but misses the exact minimizer for some $\\alpha$; worst-case Pareto error is the sharper quantity.","If the posterior family is made too flexible, each task's surrogate could overfit its own training loss and produce previews that look good in-sample but mislead on validation, suggesting the method should be paired with held-out preview evaluation.","The same posterior-merging identity suggests a route to task-weight search during continual pretraining: previews could warm-start the last fraction of pretraining where the task mix is rebalanced."],"forward_implications":["Task Arithmetic and simple parameter averaging become the isotropic-Gaussian special case of one framework, so any improvement in posterior approximation translates directly into a new merging rule.","Previews over a grid of $\\alpha$ values cost seconds to minutes after single-task finetuning, versus hours for each joint finetuning run, making broader weight searches practical for large models.","Hessian-based merging methods become full-Gaussian special cases, giving a unified explanation of why they outperform naive averaging.","For mixture posteriors, increasing the number of components tightens the preview: on the CIFAR-10 ResNet experiment the best previewed region moves toward the joint-training optimum as the number of components grows.","Because the merged mode is closed form for exponential families, the same recipe extends to non-Gaussian posteriors, not just Gaussians, as the paper illustrates with a Beta-Bernoulli example."],"supporting_citations":[{"why":"Task Arithmetic is recovered as isotropic-Gaussian posterior merging and serves as a baseline preview method.","marker":"Ilharco et al., 2023"},{"why":"Simple averaging is cast as weighted-surrogate minimization, the starting point of the framework.","marker":"Wortsman et al., 2022"},{"why":"Fisher-weighted merging, previously heuristic, becomes a full-Gaussian special case.","marker":"Matena & Raffel, 2022"},{"why":"The existing Hessian-based merging method is recovered from full-Gaussian posteriors and compared in experiments.","marker":"Daheim et al., 2024"},{"why":"Supplies the variational learning rule and the Laplace-approximation link used to build surrogates.","marker":"Khan & Rue, 2023"},{"why":"IVON, the variational optimizer used to obtain per-task Gaussian posteriors without extra Hessian passes.","marker":"Shen et al., 2024"},{"why":"The Gaussian-mixture mode-finding algorithm that Algorithm 1 generalizes.","marker":"Carreira-Perpiñán, 2000"},{"why":"The posterior-merging identity for weighted products of Bayesian posteriors is the underlying mechanism.","marker":"Tresp, 2000"}],"fun_headline_variants":["Flexible Bayesian merges sharpen Pareto tradeoff previews","Richer Bayesian posteriors give better Pareto fronts","Expressive posteriors unlock sharper Pareto previews","Posterior flexibility directly improves Pareto front estimates"],"cache_read_input_tokens":25344,"weakest_assumption_plain":"The argument rests on the variationally fitted per-task posterior $q_t$ being an accurate stand-in for the true task loss everywhere the merged solutions travel; if a more flexible $q_t$ lowers the KL divergence but moves its mode away from the regions that matter for some $\\alpha$, the preview for those weights can get worse instead of better.","fun_headline_variants_meta":{"raw":{"variants":["Flexible Bayesian merges sharpen Pareto tradeoff previews","Richer Bayesian posteriors give better Pareto fronts","Expressive posteriors unlock sharper Pareto previews","Posterior flexibility directly improves Pareto front estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00056,"raw_usage":{"total_tokens":2668,"prompt_tokens":960,"completion_tokens":1708,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":1648}},"tokens_in":576,"tokens_out":1708,"duration_ms":10245,"temperature":1.0,"reasoning_tokens":1648,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:09:38.925352+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-task problem with known losses, compute the true Pareto front by joint training on a fine $\\alpha$ grid, then compute isotropic, full-Gaussian and mixture previews from single-task models. If the full-Gaussian or mixture preview has higher mean squared error to the true front, or finds a best $\\alpha$ whose true performance is worse than the isotropic version's, the claimed monotone improvement is falsified for that setting.","supporting_citations":[{"cited_title":"Editing models with task arithmetic","cited_arxiv_id":null,"evidence_quote":"Task Arithmetic is recovered as isotropic-Gaussian posterior merging and serves as a baseline preview method."},{"cited_title":"Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time","cited_arxiv_id":null,"evidence_quote":"Simple averaging is cast as weighted-surrogate minimization, the starting point of the framework."},{"cited_title":"Merging models with F isher-weighted averaging","cited_arxiv_id":null,"evidence_quote":"Fisher-weighted merging, previously heuristic, becomes a full-Gaussian special case."},{"cited_title":"The B ayesian learning rule","cited_arxiv_id":null,"evidence_quote":"Supplies the variational learning rule and the Laplace-approximation link used to build surrogates."},{"cited_title":"Mode-finding for mixtures of G aussian distributions","cited_arxiv_id":null,"evidence_quote":"The Gaussian-mixture mode-finding algorithm that Algorithm 1 generalizes."},{"cited_title":"A B ayesian committee machine","cited_arxiv_id":null,"evidence_quote":"The posterior-merging identity for weighted products of Bayesian posteriors is the underlying mechanism."}],"review_version":1}