{"total":14,"items":[{"citing_arxiv_id":"2605.19931","ref_index":38,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"StruMPL: Multi-task Dense Regression under Disjoint Partial Supervision and MNAR Labels","primary_cat":"cs.CV","submitted_at":"2026-05-19T14:51:12+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"StruMPL is a multi-task dense regression model that jointly addresses disjoint partial supervision, MNAR labels, and inter-task physical constraints for improved forest biomass estimation from Earth observation.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.15622","ref_index":95,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Position: Zeroth-Order Optimization in Deep Learning Is Underexplored, Not Underpowered","primary_cat":"cs.LG","submitted_at":"2026-05-15T05:11:43+00:00","verdict":"UNVERDICTED","verdict_confidence":"UNKNOWN","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Zeroth-order optimization is underexplored rather than underpowered in deep learning, with limitations stemming from full-space designs that can be addressed via subspace, spectral, and systems-aware approaches.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.14370","ref_index":64,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Deciphering Neural Reparameterized Full-Waveform Inversion with Neural Sensitivity Kernel and Wave Tangent Kernel","primary_cat":"physics.geo-ph","submitted_at":"2026-05-14T04:49:23+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Neural tangent kernel from neural reparameterization modulates sensitivity and wave tangent kernels to produce spectral filtering, wavenumber modulation, and frequency bias that improve NeurFWI convergence.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"rameters using networks and updates the neural parameters by minimizing the wavefield misfit. (b) Ground truth. (c) Smooth initial model. (d) ADFWI result (smooth). (e) NeurFWI result (smooth). (f) Constant initial model. (g) ADFWI result (constant). (h) NeurFWI result (constant). models [40]. However, existing INR implementations using tanh [71], sinusoidal [40], or Gabor [64] activations may suffer from slow high-wavenumber convergence, ren- dering them computationally prohibitive for large-scale datasets. As shown in Fig. 3 (b)-(h), comparisons between implicit FWI (IFWI) [40] and automatic differentiable FWI (ADFWI) [30] on the 2D Overthrust model 4 confirm that while NeurFWI re- duces initial model dependency, it struggles with slow high-wavenumber convergence."},{"citing_arxiv_id":"2605.13106","ref_index":40,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Hypernetwork-Conditioned WENO5 Conservative-Form CNNs for One-Dimensional Conservation Laws","primary_cat":"math.NA","submitted_at":"2026-05-13T07:16:12+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"A hypernetwork conditions a conservative-form CNN to predict WENO5 weights from mesh and initial-condition metadata, preserving conservation and generalizing across resolutions for 1D hyperbolic conservation laws.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.16398","ref_index":25,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Support-Safe Variational Hybrid Filtering for Contact-Mode and Sparse-Law Recovery","primary_cat":"cs.RO","submitted_at":"2026-05-12T18:13:35+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"VHYDRO is a support-safe variational hybrid filter that jointly recovers continuous latent states, discrete contact modes, and sparse port-Hamiltonian laws per regime while preventing loss of feasible transitions.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.09824","ref_index":60,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Geometric Pareto Control: Riemannian Gradient Flow of Energy Function via Lie Group Homotopy","primary_cat":"eess.SY","submitted_at":"2026-05-11T00:01:59+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Geometric Pareto Control embeds Pareto solutions in a Lie group submanifold and navigates via Riemannian gradient flow to achieve 100% feasibility and low suboptimality in control tasks without retraining.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"LetF(z,xt) =−∇zV(z,x t)denote the nominal field for the current observation. The raw first slope is: ˜k1 =F(z t,xt).(58) To enforce the bounded-velocity condition used in the drift analysis, the slope is radially capped: k1 = Π∥·∥≤Vmax(˜k1),Π ∥·∥≤Vmax(v) =vmin { 1, Vmax ∥v∥+δv } ,(59) with a smallδv >0preventing division by zero. The midpoint latent state is: zmid =z t + 1 2∆tk 1.(60) The nominal field is then evaluated at this midpoint and capped again: ˜k2 =F(z mid,xt), k 2 = Π∥·∥≤Vmax(˜k2).(61) The Euclidean predictor passed to the geometry-aware residual correction is: ∆z Euc = ∆tk 2.(62) NaN-safe fallback.All midpoint computations are guarded. If any component ofz mid is non-finite, the algorithm setsz mid←zt before evaluating the second slope."},{"citing_arxiv_id":"2605.09707","ref_index":17,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Adaptive Data Harvesting for Efficient Neural Network Learning with Universal Constraints","primary_cat":"cs.LG","submitted_at":"2026-05-10T19:09:49+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":4.0,"formal_verification":"none","one_line_summary":"A reinforcement learning policy learns to adaptively harvest data samples, improving empirical constraint satisfaction and training efficiency for Lyapunov NNs and PINNs.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"tion entity, and A is the action space encompassing all possible sampled batches of training examples (or their distribution parameterization). The transition function P : S×A×S→ [0, 1] models how the inner optimization entity evolves in response to selected actions, while the reward function R : S×A→I R quantifies the effec- tiveness of the selected examples. The discount factor γ∈ [0, 1) controls the trade-off between immediate and future rewards, and µ0 : S→I R specifies the initial state distribution. We next detail how to apply this RL-based framework to train Lyapunov NNs and PINNs. 3.1 Application to Lyapunov function learning We next show how adaptive data harvesting can be applied to learning a Lyapunov function vθ(x) with parameter θ, certifying a large subset of Sψ as safe. To Adaptive Data Harvesting for Efficient Neural Network Learning with Universal Constraints Algorithm 2RL-Guided Adaptive Expansion ROA Require: Testing dynamics f ψtest, simulation horizon T, initialization methodD Ensure:Lyapunov NNv θ for testing dynamics 1: Initialize RL policy π. // Training 2:fort= 1 to RL episodedo 3: Initialize random training dynamics f ψ, Lya- punov NNv θ, safe levelc 0. 4:fork = 1 to ROA resampledo 5:Generate multiplier based on training state: αk =π(s k). 6: Sample a batch of points from the expanded regionV θ(αkck), and denote it asX. 7:Forward simulate the batch forTsteps: S← {x∈X:f ψ T (x)∈V θ(ck)}. 8: Train vθ on X using labels yi = +1 for xi ∈S , yi =−1 otherwise. 9:c k+1 ←max x∈S vθ(x), computer k, updateπ. 10:end for 11:end for 12: Initialize Lyapunov NN vθ, safe level c0.// Testing 13:fork= 1 to ROA resampledo 14: Generate multiplier based on training state αk = π(sk). 15: Sample a batch of points from the expanded regionV θ(αkck), and denote it asX. 16: Forward simulate the batch for T steps: S← {x∈X:f ψtest T (x)∈V θ(ck)}. 17: Train vθ on X using labels yi = +1 for xi ∈S , yi =−1 otherwise. 18:c k+1 ←max x∈S vθ(x). 19:end for this end, "},{"citing_arxiv_id":"2605.09299","ref_index":115,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"LagrangianSplats: Divergence-Free Transport of Gaussian Primitives for Fluid Reconstruction","primary_cat":"cs.GR","submitted_at":"2026-05-10T03:45:50+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"A framework that structurally enforces divergence-free velocity and long-range transport coherence in 3D fluid reconstruction from 2D videos via divergence-free kernels advecting Lagrangian Gaussian splats.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.08915","ref_index":35,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Physics-Informed Neural PDE Solvers via Spatio-Temporal MeanFlow","primary_cat":"cs.LG","submitted_at":"2026-05-09T12:29:10+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Spatio-Temporal MeanFlow adapts MeanFlow to PDEs by replacing the generative velocity field with the physical operator and extending the integral constraint to the spatio-temporal domain, yielding a unified solver for time-dependent and stationary equations with improved accuracy and generalization.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"i=1 ∥ai∥2 2) and the operator norm bound∥v·g∥ 2 ≤ ∥g∥ op∥v∥2, we obtain: ∥rMeanFlow∥2 2 ≤3l 2 t ∥rTemp∥2 2 + 3 \u0012 l−l t ls ∥g∥op \u00132 ∥rSpac∥2 2 + 3∥r∥2 2 (34) where r= lt −l ls \u0014 m(x−ξ) ⊤ −l 2 s ∂m ∂ξ \u0015 ·g | {z } Boundary or Higher-Order Discretization Errors + (l t −l)Kh| {z } Source Term Error + (l 2 t −γl 2) ∂m ∂τ| {z } Temporal Scaling Mismatch ,(35) collects the remaining structural terms. Therefore, as ∥K∥op ≤ρ and the state variables are sufficiently smooth such that their norms are bounded, we have, applying generalized triangle inequality for squaredL 2 norms again, ε= 3∥r∥ 2 2 ≤9(lt −l) 2 · \" 1 l2s · ∥rg∥2 2 +ρ 2∥h∥2 2 + \u0012 l2 t −γl 2 lt −l \u00132 ∂m ∂τ 2 2 # , = ( 9∥r g∥2 2 + 9l2 s ·ρ 2∥h∥2"},{"citing_arxiv_id":"2605.07444","ref_index":7,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Accelerated and data-efficient flow prediction in stirred tanks via physics-informed learning","primary_cat":"cs.CE","submitted_at":"2026-05-08T08:49:40+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Physics-informed constraints on implicit neural representations yield more accurate and stable predictions of stirred-tank flows than purely data-driven models when training data is scarce, with diminishing returns at larger dataset sizes.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.03399","ref_index":28,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"PODiff: Latent Diffusion in Proper Orthogonal Decomposition Space for Scientific Super-Resolution","primary_cat":"cs.LG","submitted_at":"2026-05-05T06:21:04+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"PODiff performs conditional diffusion in a fixed, variance-ordered POD latent space to enable efficient probabilistic super-resolution of high-dimensional scientific fields with lower memory and better-calibrated uncertainty than pixel-space or dropout baselines.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.01702","ref_index":64,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Floating-Point Networks with Automatic Differentiation Can Represent Almost All Floating-Point Functions and Their Gradients","primary_cat":"cs.LG","submitted_at":"2026-05-03T04:06:41+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":8.0,"formal_verification":"none","one_line_summary":"Floating-point neural networks with automatic differentiation can represent arbitrary floating-point functions and their gradients under mild conditions.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.18491","ref_index":37,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Faster by Design: Interactive Aerodynamics via Neural Surrogates Trained on Expert-Validated CFD","primary_cat":"cs.LG","submitted_at":"2026-04-20T16:42:35+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"A graph-based neural operator trained on expert-validated race-car CFD data reaches accuracy levels usable for early-stage interactive aerodynamic design exploration.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.18277","ref_index":13,"ref_count":1,"confidence":0.55,"is_internal_anchor":false,"paper_title":"Dissipative Latent Residual Physics-Informed Neural Networks for Modeling and Identification of Electromechanical Systems","primary_cat":"cs.LG","submitted_at":"2026-04-20T13:52:12+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"DiLaR-PINN learns dissipative effects in electromechanical systems via a skew-dissipative latent residual PINN that guarantees non-increasing energy and uses recurrent curriculum training for partial observations.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null}],"limit":50,"offset":0}