Coverage: substantive releases and revisions from 28 August–3 September 2026. Four developments met the editorial threshold. This was a stronger week for evaluation and hybrid scientific computing than for autonomous mathematical discovery.
Why it matters MathAdv replaces a single proof-success score with separate tests of mathematical knowledge, informal reasoning, Lean proof construction, and robustness to equivalent reformulations. This makes it possible to locate whether failure comes from mathematics, reasoning, or formalization.
Primary contributions
Introduces 321 expert-curated problems across 13 undergraduate and graduate domains; 298 have Lean 4 statements.
Finds that formalization remains a major bottleneck and that performance varies sharply across mathematical domains.
Shows that natural-language hints can help general-purpose models while harming proof-specialized models.
Demonstrates substantial brittleness under mathematically equivalent reformulations, weakening claims based only on aggregate pass rates.
Evidence and caution The benchmark uses expert review, compiler-verifiable outputs, public evaluation scripts, and contrast-style reformulations. Its 321 problems are diagnostically valuable but still too small for broad claims about research-level mathematical competence; future training-set exposure may also reduce its value.
Read first The benchmark-construction and human-in-the-loop formalization sections, followed by the cross-domain and reformulation-robustness results.
Research software · Moderate evidence
2. PyDoseRT Photon and Proton: differentiable physics with learned correction
Attila Simkó, Lukas Zimmermann, Hermann Fuchs and Gerd Heilemann · arXiv · 1 September 2026 · Photon paper ↗ · Proton paper ↗
Why it matters These systems follow a credible hybrid pattern: retain an analytical numerical model, make it differentiable for calibration, and learn only a structured residual correction toward Monte Carlo accuracy.
Primary contributions
Implements GPU pencil-beam dose engines in PyTorch while preserving physically motivated kernels and transport approximations.
The photon pipeline adds small learned fluence and heterogeneity priors plus a compact 1.36-million-parameter 3D residual U-Net.
Zero-initializes learned corrections so training begins from the analytical solution rather than an unconstrained neural prediction.
The proton engine retains a differentiable 3D execution path for gradient-based calibration before learned residual correction.
Reports local or preliminary challenge gamma pass rates of 99.12% for photon CT, 96.95% for photon MRI, and 98.30% for proton at 1%/1 mm.
Evidence and caution The design is technically coherent, but these are challenge reports using local or preliminary evaluation. Independent external validation, production-solver comparisons, and a public code release were not established in the papers.
Computational-mathematics relevance The analytical-model-plus-residual pattern transfers naturally to reduced-order PDE models, approximate Jacobians, preconditioners, inverse problems, and surrogate correction.
Scientific ML method · Preliminary evidence
3. Physics-Informed Foresight Pruning for Sparse PINN Solvers of Nonlinear PDEs
Ahmad Ishaque Karimi, Uvini Balasuriya Mudiyanselage and Kookjin Lee · AI4S 2026 workshop submission · 26 August 2026 · Paper ↗
Why it matters Output sensitivity alone can miss parameters whose main influence is transmitted through derivatives in the governing PDE. PI-SAP instead defines pruning saliency using sensitivity of the differential residual.
Primary contributions
Extends spectrum-aware pruning at initialization from output-side dynamics to PDE-residual sensitivity.
Evaluates sparse PirateNet solvers on Gray–Scott, complex Ginzburg–Landau, Burgers, and linear convection equations.
Finds stronger preservation of Gray–Scott residual fidelity and competitive behavior at aggressive sparsity.
Shows that residual fidelity, solution accuracy, and kernel conditioning are distinct objectives.
Reports no uniformly best pruning criterion across equations or sparsity levels.
Evidence and caution This is a seven-page workshop paper with four PDE examples. Parameter sparsity also does not guarantee lower GPU wall time without hardware-aware sparse kernels.
Symbolic discovery · Exploratory
4. Neural Symbolic Regression Using Deep Learning and Sparse Modelling
Ravi Kumar U and Sumitra S · arXiv · 1 September 2026 · Paper ↗
Why it matters The method treats a neural approximation as a noise-robust functional preconditioner, then applies sparse regression to recover a compact expression.
Primary contributions
Proposes a decoupled pipeline combining interaction-aware neural smoothing with LASSO-based expression recovery.
Uses Ray Tune and ASHA for distributed hyperparameter selection.
Reports improved RMSE, noise robustness, and out-of-distribution prediction over SINDy and untuned neural baselines on Nguyen benchmarks.
Provides ablations for feature interactions, neural depth, and tuning choices.
Evidence and caution The Nguyen suite is synthetic and narrow; the comparison omits several strong modern symbolic-regression systems. Unequal tuning budgets may advantage the proposed method, and the abstract does not identify released code. Treat this as a hypothesis-generating study rather than established scalable equation discovery.
Selection and commentary are editorial. Reported numerical results are claims made by the cited authors and are distinguished from independent validation.