{"id":"587eb925-c108-4ad6-9843-7d38e8f98a17","arxiv_id":"2502.01117","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Training a diffusion weight generator with meta-learning and a local-consistency loss lets it hit intermediate optimizer checkpoints on schedule and end at the global optimum.","lead":"Researchers propose Mc-Di, a method that trains a diffusion model to generate neural-network weights for new tasks by learning from intermediate optimization checkpoints, not just the final trained weights. It combines meta-learning with a local-consistency objective and reports accuracy and latency gains across few-shot learning, transfer learning, and LLM fine-tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof uses a different reverse update than the evaluation rule Eq. (2); the deployed inference rule is unanalyzed, so the local-consistency guarantee is not established.","rationale":"The reader's REJECT verdict is supported by the same load-bearing weakness: the inference rule in Eq. (2), which is used in the downstream evaluation and stated in Theorem 1, is not the update whose posterior mean is derived in Appendix A. My stress-test confirms the mismatch and adds a stronger observation: Eq. (2) is not a consistent denoising update even when the denoiser is perfect, so the theorem's guarantee cannot hold as stated for that rule. This is not a matter of empirical disagreement or missing code; it is an internal inconsistency in the theoretical foundation. The paper does present broad experiments and a plausible meta-learning diffusion pipeline, and the empirical claims might survive a corrected proof or a corrected implementation, but the central mechanism advertised in the abstract and Section 3.1 is currently unproven. A re-derived proof or a simple oracle simulation would settle the issue. Since the reader already recommends REJECT and my concern does not move that verdict, I mark the verdict unchanged. Credit is given where due: the ablation study, transfer/few-shot results, and latency measurements are extensive, and the SAM-based convergence analysis in Theorem 2 is a separate, more standard argument. However, the local consistency guarantee is the paper's distinguishing contribution, and it is exactly the part that fails scrutiny.","tokens_in":23445,"tokens_out":11237,"duration_ms":107380,"concrete_test":"Simulate a 1D Gaussian toy: set T=4, k=2, choose a concrete increasing schedule (e.g., beta_t linear from 0.02 to 0.2, alpha_t=1-beta_t), targets theta_2=1, theta_4=2, and a fixed Gaussian epsilon. Generate x_0 = sqrt(alphabar_0) theta_4 + sqrt(1-alphabar_0) eps. Run the evaluation rule Eq. (2) for t=0..3 with the oracle denoiser eps_phi(x_t,t)=eps and record x_2 and x_4. If x_2 != theta_2 or x_4 != theta_4, the stated theorem fails even under ideal denoising. Then repeat the same simulation with the standard DDPM update used in the Appendix A proof and check whether that update reaches the targets; this isolates the update mismatch from any learning effects and determines which reverse rule, if either, Theorem 1 actually analyzes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 1: training with the local consistency loss L_loc in Eq. (3) lets the denoiser recover theta_d, ..., theta_M at intervals of T/k under the inference rule in Eq. (2). The Appendix A proof does not establish this. The theorem statement (main Eq. (2) and Appendix A Eq. (1)) uses x_{t+1} = (1/sqrt(alphabar_{t+1}))(x_t - sqrt(1-alphabar_{t+1}) eps_phi), but in the proof the model update is written, after Eq. (10), as x_{t+1} = (1/sqrt(alpha_t))x_t - (1-alpha_t)/(sqrt(1-alphabar_t) sqrt(alpha_t)) eps_phi, with alphabar_t = product_{j=t}^{T-1} alpha_j. These are not algebraically equivalent. Moreover, Eq. (2) is not the standard DDPM/DDIM update even with a perfect denoiser: substituting x_t = sqrt(alphabar_t) theta_M + sqrt(1-alphabar_t) eps into Eq. (2) with eps_phi = eps does not yield sqrt(alphabar_{t+1}) theta_M + sqrt(1-alphabar_{t+1}) eps, and the target coefficient after T steps is not 1. Thus the inference rule actually used in evaluation does not define a valid reverse process for the forward noising defined in Section D.1. The proof also mixes cumulative products: the left posterior is computed with the segment schedule alphabar_t^{T'} = product_{j=t}^{T'-1} alpha_j, while the right model term uses the global alphabar_t, producing the mismatched scaling in Eq. (11)/Eq. (3). Consequently, the guarantee that local targets are reached on schedule is unproven for the deployed algorithm; the empirical gains could come from the meta-learning/diffusion pipeline rather than from the specific local-consistency mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Mc-Di, a meta-learning framework that uses a conditional diffusion model to generate neural-network weights for downstream tasks. The method combines REPTILE-style bi-level optimization with a diffusion denoiser, and its main novelty is 'local consistency diffusion': the denoiser is trained with a local consistency loss that supervises intermediate weight targets sampled uniformly from an optimization trajectory, with the aim of recovering the sequence θ_d, θ_2d, ..., θ_M at evenly spaced intervals of T/k diffusion steps. The paper gives Theorem 1 to justify this property, Theorem 2 to analyze convergence and motivate a SAM-based improvement, and experiments on transfer learning, few-shot learning, domain generalization, and LLM fine-tuning. The empirical results are positive and the paper is ambitious, but the theoretical guarantee as stated is not established because the proof in Appendix A uses a different denoising update than the one declared in the theorem and in Section 2.3.","tokens_in":23768,"tokens_out":14626,"duration_ms":142753,"significance":"If the local-consistency property were rigorously established, the paper would make a useful contribution to gradient-free weight generation: exploiting intermediate optimizer states could improve both generation accuracy and inference latency, and the meta-learning wrapper addresses cross-task transfer. The empirical study is broad, covering four downstream scenarios with consistent gains over several baselines, and the paper also attempts a convergence analysis. However, the central theoretical claim is load-bearing and currently unsupported: the inference rule stated in Eq. (2) is not the update analyzed in the proof, and the concatenation of local segments is asserted rather than proven. The SAM-improvement claim also rests on an unverified assumption about the Hessian spectrum. These issues do not necessarily invalidate the empirical observations, but they do mean the paper's main intellectual contribution is not yet demonstrated.","major_comments":[{"comment":"Theorem 1 is not proved for the inference rule it states. Main-text Eq. (2) and Appendix A Eq. (1) define the update as x_{t+1} = (1/sqrt(alphabar_{t+1}))(x_t - sqrt(1-alphabar_{t+1}) eps_phi), but the proof after Appendix A Eq. (10) substitutes the standard DDPM update x_{t+1} = (1/sqrt(alpha_t))x_t - ((1-alpha_t)/(sqrt(1-alphabar_t)sqrt(alpha_t))) eps_phi, with alphabar_t = product_{j=t}^{T-1} alpha_j. These are not algebraically equivalent. Moreover, the stated update is not a valid reverse process for the forward model q(x_t|x_T) = N(sqrt(alphabar_t)x_T, (1-alphabar_t)I): substituting a perfect denoiser eps_phi = eps and x_t = sqrt(alphabar_t)theta + sqrt(1-alphabar_t)eps does not yield sqrt(alphabar_{t+1})theta + sqrt(1-alphabar_{t+1})eps, and at t = T-1 (where alphabar_T = 1) the update gives x_T = x_{T-1}, so the final step performs no denoising. Thus the guarantee that local targets are reached on schedule is unproven for the deployment rule described in the paper. The authors should either prove the theorem for the update actually used in implementation (the one in Appendix D.1) or correct the stated inference rule, and re-derive the local consistency loss accordingly.","section":"Section 2.3 and Appendix A"},{"comment":"The proof analyzes a single segment i in isolation and then asserts that training with the sum L_loc = sum_i L_loc^i makes the different segments share one inference chain. This concatenation is not demonstrated. For any t < T/k, the loss receives contributions from every i, each with a different schedule alphabar_t^i and a different target theta_{i*d}; the same network must satisfy all these conditions simultaneously at the states visited by the shared chain. The proof does not rule out interference between segments, nor does it show that the fixed-point of the combined objective yields the claimed 'pass through theta_d at T/k, then theta_2d at 2T/k, ...' behavior. A rigorous argument for the concatenation property is required for the central claim of the paper.","section":"Theorem 1 proof, Appendix A"},{"comment":"The claimed convergence benefit of SAM is not established. Theorem 2 assumes an l-smooth, mu-strongly convex loss with Hessian eigenvalues bounded by lambda, and the bound involves lambda; the paper then states that SAM improves convergence by reducing lambda. However, no theorem or measurement in the paper shows that SAM reduces lambda in the setting of interest, and Figure 7 plots accuracy versus GPU hours rather than Hessian spectra. In addition, Lemma 1 analyzes gradient descent with step size 1/l, while the weight preparation stage in Algorithm 1 uses Adam. As written, the statement in Section 3.2.3 that 'the convergence of the weight generation paradigm can be improved by introducing SAM' is an empirical observation, not a consequence of the presented analysis. The authors should either provide a proof that SAM reduces the relevant lambda, or expressly label the SAM motivation as empirical.","section":"Theorem 2 and Section 3.2.3"}],"minor_comments":[{"comment":"The loss L_loc in Eq. (3) drops the positive coefficient (1-alpha_t)^2 / (alpha_t (1-alphabar_t^i)(1-alphabar_t)) that appears in the KL derivation in Eq. (11) of Appendix A. The proof only establishes proportionality, not equality; if the theorem's guarantee depends on the exact objective, this discrepancy should be addressed or the loss definition should be revised.","section":"Theorem 1 statement, Eq. (3)"},{"comment":"The inference update in Eq. (2) is described as 'aligns with the vanilla diffusion algorithm', but it does not match the vanilla update given in Appendix D.1. These two formulas should be reconciled, and the indexing convention for alphabar_t (including the empty product at t = T) should be stated explicitly.","section":"Section 2.3 and Appendix D.1"},{"comment":"The Taylor expansion in Eq. (14) evaluates the Hessian at a point xi on the segment between hat-theta and theta*, while the assumption bounds eigenvalues only 'around the optimum theta*'; the bound in Eq. (15) therefore needs an additional uniformity assumption on the Hessian along that segment.","section":"Appendix C, proof of Theorem 2"},{"comment":"Several references are incomplete or malformed, e.g., '[23] et al. Jang', '[37] et al. Madan', '[41] et al. Pan', and '[51] et al. Zhang'; these should be completed with full author lists.","section":"References"},{"comment":"The axis labels in Figures 4 and 5 appear garbled in the submitted PDF (the text is rendered as unicode fragments). Please ensure that all figures use readable, standard labels.","section":"Figures 4 and 5"}],"recommendation":"major_revision","confidential_remarks":"The empirical results are extensive and the idea of using local optimizer trajectories to guide weight generation is attractive. However, the main theorem's proof is inconsistent with the stated inference rule, and without a correct proof of the local-consistency property the paper's central claim is not established. I recommend major revision rather than rejection because the mismatch appears correctable: the authors can restate Theorem 1 for the actual reverse update used in implementation and provide a rigorous proof of the segment-concatenation property. They should also clarify the status of the SAM claim. If the authors cannot align the theory with the deployment rule, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is genuinely new: train a diffusion weight generator on intermediate optimizer checkpoints with a local-consistency loss, and wrap it in REPTILE-style bi-level meta-learning. The experiments are broad, the ablations are informative, and the latency numbers are consistent across transfer, few-shot, domain generalization, and LLM fine-tuning. The ablation Tw-Di vs. Mv-Di vs. Mc-Di does a good job isolating the contribution of the consistency loss, and the k>1 local-target scheme with T/k spacing is a real departure from prior diffusion weight generators.\n\nThe soft spot is load-bearing. Theorem 1's statement and the evaluation rule in Eq. (2) use the update x_{t+1} = (1/sqrt(alphabar_{t+1}))(x_t - sqrt(1-alphabar_{t+1}) epsilon_phi). The proof in Appendix A, however, derives the loss using the standard DDPM update x_{t+1} = (1/sqrt(alpha_t))x_t - (1-alpha_t)/(sqrt(1-alphabar_t) sqrt(alpha_t)) epsilon_phi. These are not algebraically equivalent, and the proof's posterior calculation mixes segment-level and global cumulative products. So the central claim that the trained denoiser reaches theta_{i*d} exactly at step i*T/k is unproven for the inference rule actually deployed. The empirical gains could come from the meta-learning/diffusion pipeline rather than from the specific local-consistency mechanism. This is not a cosmetic gap; it is the main theoretical justification for the method.\n\nTheorem 2 relies on strong convexity, which the authors admit is unrealistic for deep networks. That is a minor concern because the bound's structure plausibly holds, and they only use it to motivate SAM. More annoying is the abstract's 'superior accuracy' claim: Table 5 shows Mc-Di below LoRA on QNLI, so the honest phrasing should be 'comparable accuracy with large speedup.' No code or data are released, which limits verification of the empirical results.\n\nWho should read this: people working on weight generation and hypernetworks will want the empirical results and the local-target idea, even though the theory needs repair. A serious referee should see it, because the empirical claim is important if the proof gets fixed. My recommendation: send to peer review, but with a clear note that Theorem 1's proof must be reconciled with the inference rule, and the LLM claims toned down. If those land, this could be a solid paper.","headline":"A promising empirical recipe for diffusion-based weight generation whose central consistency theorem, as stated, isn't proved by the appendix — worth engaging but not ready as-is.","tokens_in":24389,"tokens_out":1995,"would_cite":false,"duration_ms":20443,"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":"A diffusion denoiser trained with a local consistency loss generates task weights without gradients, hitting intermediate optimizer checkpoints on schedule and ending at the global optimum, with higher accuracy and lower latency.","keywords":["diffusion models","weight generation","meta-learning","local consistency diffusion","few-shot learning","transfer learning","domain generalization","LoRA adaptation"],"falsifier":"Run the trained Mc-Di sampler with the deployed inference rule $x_{t+1} = \\frac{1}{\\sqrt{\\bar\\alpha_{t+1}}}(x_t - \\sqrt{1-\\bar\\alpha_{t+1}}\\epsilon_\\phi(x_t,t))$ and record $\\|x_{iT/k} - \\theta_{i\\times d}\\|^2$ for $i=1,\\ldots,k$. Theorem 1 predicts these distances go to zero at every checkpoint; if the final weight is accurate but the checkpoints do not track $\\theta_{i\\times d}$, the local-consistency mechanism is not what produces the reported gains. An algebraic check of whether this recursion equals the proof's update $x_{t+1} = \\frac{1}{\\sqrt{\\alpha_t}}x_t - \\frac{1-\\alpha_t}{\\sqrt{1-\\bar\\alpha_t}\\sqrt{\\alpha_t}}\\epsilon_\\phi$ would settle the mismatch without new experiments.","tokens_in":23157,"feed_emoji":"⚡","tokens_out":12952,"duration_ms":124942,"temperature":0.7,"pith_summary":"Diffusion models that generate neural-network weights are attractive because they replace gradient fine-tuning with a single sampling run, but existing methods generalize poorly across tasks and only model the final optimal weights. This paper argues that the fix is to treat the optimizer's intermediate checkpoints as local targets and train the denoiser with a local consistency loss. The proposed method, Mc-Di, wraps this diffusion model in a bi-level meta-learning update so it learns across tasks instead of per task. The key theoretical claim is that one inference chain can reach every local target at evenly spaced intervals and still finish at the global optimum, which is what makes the generated weights accurate and the inference fast. If correct, Mc-Di gives a gradient-free recipe for transfer learning, few-shot learning, domain generalization, and large language model adaptation.","feed_headline":"Local-consistency diffusion speeds weight generation 1.7-3.6x","feed_subtitle":"Training the denoiser on Adam checkpoint weights as local targets improves accuracy on unseen tasks and cuts latency.","key_machinery":"The load-bearing object is the local consistency loss $L^{\\mathrm{loc}} = \\mathbb{E}_{i\\in(0,k]} L_i^{\\mathrm{loc}}$, with $$$L_i^{{\\mathrm{loc}}$} = \\mathbb{E}_{t\\in[0,iT/k)} \\big\\|\\sqrt{1-\\bar\\alpha_t^i}\\,\\epsilon_\\$\\varphi$(x_t,t) - \\sqrt{1-\\bar\\alpha_t}\\,\\epsilon\\big\\|^2, \\quad x_t = \\sqrt{\\bar\\alpha_t^i}\\,\\theta_{i\\times d} + \\sqrt{1-\\bar\\alpha_t^i}\\,\\epsilon,$$ where $\\bar\\alpha_t^i = \\prod_{j=t}^{iT/k-1}\\alpha_j$ and $\\bar\\alpha_t = \\prod_{j=t}^{T-1}\\alpha_j$. The loss makes the denoiser predict, within each segment, the noise that maps the next state back to the local target, and Theorem 1 says the shared chain then hits $\\theta_{i\\times d}$ at step $iT/k$. Supporting machinery: a weight-preparation stage that runs Adam and samples $\\{\\theta_d,\\ldots,\\theta_{kd}\\}$ at uniform spacing $d=M/k$; a bi-level meta-learning inner loop that conditions the denoiser on a task embedding; and sharpness-aware minimization plus data augmentation moved into weight preparation so they improve the Theorem 2 bound without extra meta-training time.","core_discovery":"The central claim is that a denoiser $\\epsilon_\\phi$ trained with the local consistency loss $L^{\\mathrm{loc}}$ from Eq. (3) recovers the target sequence $\\theta_d,\\theta_{2d},\\ldots,\\theta_{M=kd}$ from standard Gaussian noise under the vanilla diffusion inference rule of Eq. (2), reaching $\\theta_{i\\times d}$ after exactly $iT/k$ steps and $\\theta_M$ after $T$ steps. The proof derives the loss from the DDPM denoising matching objective, using Bayes' rule and the reparameterization trick to compare the posterior expectation with the denoiser's prediction. Because the intermediate targets are placed on the same inference chain rather than trained as separate generation tasks, the local targets reinforce the global optimum instead of conflicting with it. The paper also proves a cumulative-error bound, $L_D(\\hat\\theta)-L_D(\\theta^*) \\le \\frac{\\lambda}{2}(c + \\frac{2\\psi}{\\mu}(1-\\mu/l)^M)$, and moves sharpness-aware minimization into weight preparation to shrink $\\lambda$. At $k=1$ the method reduces to vanilla diffusion, so Mc-Di is presented as the generalization of prior diffusion weight-generation methods.","pith_inferences":["The same local-consistency objective could be tested on other optimizer trajectories (e.g., SGD with momentum or AdamW) and on non-uniform target spacings, which would show whether the gains come from the trajectory itself or from the particular checkpoints chosen.","A direct comparison with a curriculum that trains each segment independently and stitches the chain at inference would isolate the value of the shared-chain consistency term; the paper's theory predicts the shared version should win at equal compute.","The limitation section points toward arbitrary sub-trajectories $\\theta_i \\to \\theta_j$ and names Flow Matching and trajectory-balance objectives as the next step; if the local-consistency idea carries over, weight generation could eventually start from any previous weight rather than from Gaussian noise.","If the guarantee holds on larger models, weight generation for foundation-model adaptation could become a single forward pass per task, changing how multi-task serving is scheduled."],"forward_implications":["Bi-level meta-learning makes diffusion weight generation transferable across tasks: after meta-training, a single denoiser can generate weights for unseen tasks without any gradient updates, which is what distinguishes Mc-Di from single-level diffusion methods like OCD and D2NWG.","Local targets add supervision without breaking global consistency, so the same inference chain reaches intermediate checkpoints and the final global optimum; the ablation attributes the accuracy gain to this consistency rather than to the extra targets alone.","Decomposing a $T$-step generation problem into $k$ shorter subproblems reduces the practical time to reach a given reconstruction error, so Mc-Di keeps accuracy with fewer diffusion steps.","Moving SAM and data augmentation into weight preparation improves the convergence bound through a smaller Hessian curvature term without adding time during meta-training.","Generated LoRA matrices reproduce gradient-based fine-tuning accuracy at roughly one quarter of the wall-clock latency, so Mc-Di offers a multi-task fine-tuning mode for LLMs."],"supporting_citations":[{"why":"Supplies the DDPM forward/reverse process and denoising matching objective from which local consistency diffusion is derived.","marker":"[20]"},{"why":"Provides the unified DDPM derivation (Bayes rule, reparameterization) used in the proof of Theorem 1.","marker":"[34]"},{"why":"Supplies the REPTILE bi-level meta-learning update that gives Mc-Di cross-task transferability.","marker":"[39]"},{"why":"Defines the OCD baseline that existing single-level diffusion weight generation relies on and that Mc-Di generalizes.","marker":"[35]"},{"why":"Provides the conditional diffusion architecture and task-conditioning scheme Mc-Di adopts, and the Meta-Diff baseline it extends.","marker":"[50]"},{"why":"Introduces D2NWG, the fastest prior diffusion weight-generation method, used as the latency comparison baseline.","marker":"[47]"},{"why":"Supplies sharpness-aware minimization, moved into weight preparation to reduce the Hessian eigenvalue in Theorem 2.","marker":"[14]"},{"why":"Provides Adam, the real-world optimizer whose trajectory is sampled to produce local target weights.","marker":"[26]"},{"why":"Supplies GHN3, a parameter-prediction baseline compared across transfer, few-shot, and domain generalization.","marker":"[29]"}],"fun_headline_variants":["Diffusion weight generation with local consistency boosts speed up to 3.6x","Mc-Di: meta-learning plus local consistency diffusion for fast weight updates","Denoiser trained on local target weights yields 3.6x faster weight generation","Local consistency diffusion recovers intermediate targets on one chain","Faster weight generation via local-consistency diffusion with meta-learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire local-consistency argument depends on the update rule used in the proof being algebraically the same as the update rule the model runs at inference; the paper writes them differently, so if they are not equivalent, the promised on-schedule hitting of local targets is not proved.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion weight generation with local consistency boosts speed up to 3.6x","Mc-Di: meta-learning plus local consistency diffusion for fast weight updates","Denoiser trained on local target weights yields 3.6x faster weight generation","Local consistency diffusion recovers intermediate targets on one chain","Faster weight generation via local-consistency diffusion with meta-learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000884,"raw_usage":{"total_tokens":3826,"prompt_tokens":958,"completion_tokens":2868,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":2772}},"tokens_in":574,"tokens_out":2868,"duration_ms":18356,"temperature":1.0,"reasoning_tokens":2772,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T16:32:40.145712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained Mc-Di sampler with the deployed inference rule $x_{t+1} = \\frac{1}{\\sqrt{\\bar\\alpha_{t+1}}}(x_t - \\sqrt{1-\\bar\\alpha_{t+1}}\\epsilon_\\phi(x_t,t))$ and record $\\|x_{iT/k} - \\theta_{i\\times d}\\|^2$ for $i=1,\\ldots,k$. Theorem 1 predicts these distances go to zero at every checkpoint; if the final weight is accurate but the checkpoints do not track $\\theta_{i\\times d}$, the local-consistency mechanism is not what produces the reported gains. An algebraic check of whether this recursion equals the proof's update $x_{t+1} = \\frac{1}{\\sqrt{\\alpha_t}}x_t - \\frac{1-\\alpha_t}{\\sqrt{1-\\bar\\alpha_t}\\sqrt{\\alpha_t}}\\epsilon_\\phi$ would settle the mismatch without new experiments.","supporting_citations":[{"cited_title":"OCD: learning to overfit with conditional diffusion models","cited_arxiv_id":null,"evidence_quote":"Defines the OCD baseline that existing single-level diffusion weight generation relies on and that Mc-Di generalizes."},{"cited_title":"Metadiff: Meta-learning with conditional diffusion for few-shot learning","cited_arxiv_id":null,"evidence_quote":"Provides the conditional diffusion architecture and task-conditioning scheme Mc-Di adopts, and the Meta-Diff baseline it extends."},{"cited_title":"Sharpness-aware mini- mization for efficiently improving generalization","cited_arxiv_id":null,"evidence_quote":"Supplies sharpness-aware minimization, moved into weight preparation to reduce the Hessian eigenvalue in Theorem 2."},{"cited_title":"Adam: A method for stochastic optimization","cited_arxiv_id":null,"evidence_quote":"Provides Adam, the real-world optimizer whose trajectory is sampled to produce local target weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies GHN3, a parameter-prediction baseline compared across transfer, few-shot, and domain generalization."}],"review_version":1}