TL;DR: Enantiomers share atoms, bonds, and pairwise distances yet behave differently in chiral environments, so molecular encoders must respect rotations without going blind to reflection. GSF-χ lets every stereogenic unit modulate all pairwise attention through a reflection-even phase field, while a handedness pseudoscalar χ reverses a Chiral-RoPE on latent query–key blocks; a C₂ projection keeps ECD peak counts and positions mirror-even and peak signs mirror-odd. GSF-χ leads every central-ECD output and improves axial Rotation and Symbol by 12.6% and 7.9% over the strongest baseline.

TL;DR: Rotary position encodings make attention sensitive to relative position, but prior graph versions assign each node a rotation from spectral coordinates, so the rotary factor between two nodes cannot distinguish the routes connecting them. AW-RoPE places the rotations on edges and sums the transported features over all walks, so contributions along different routes can reinforce or cancel: an exact variant evaluates the complete sum by a differentiable linear solve, and a sparse variant truncates it at finite depth. Across five synthetic tasks both variants reduce nRMSE by 15–58% relative to the strongest baseline, and on real superpixel, peptide and OGB benchmarks the sparse recurrence attains the best mean on every dataset with Performer kernels and on twelve of thirteen datasets with GIN.

TL;DR: Recent NMR structure-elucidation models rely heavily on large corpora of computed spectra and drop sharply on experimental measurements. We build NMRSpec, a large-scale corpus of experimental ¹H and ¹³C spectra mined from chemical literature, and propose NMRTrans, which models spectra as unordered peak sets and aligns the model's inductive bias with the physical nature of NMR. To our knowledge NMRTrans is the first NMR Transformer trained solely on large-scale experimental spectra, improving Top-10 Accuracy over the strongest baseline by +17.82 points (61.15% vs. 43.33%).

TL;DR: Foundation models for medical image segmentation struggle under out-of-distribution shifts, often producing fragmented false positives on OOD tumors. R²-Seg is a training-free Reason-and-Reject framework: an LLM-guided anatomical reasoning planner localizes organ anchors and generates multi-scale ROIs, then two-sample statistical testing rejects candidates from a frozen foundation model (BiomedParse) that are not significantly different from normal tissue. On multi-center, multi-modal tumor segmentation benchmarks, R²-Seg substantially improves Dice, specificity, and sensitivity over strong baselines with no parameter updates.

TL;DR: MatVerse is a materials foundation model built upon Intern-S2-Preview (35B-A3B) and specialized through large-scale materials-domain post-training on ~150K multimodal examples covering materials knowledge, scientific reasoning, quantitative problem solving, structural understanding, and multimodal scientific interpretation, constructed with teacher-guided reasoning distillation, scientific problem synthesis, consistency-based filtering, and retrieval-based multimodal data construction. Across OmniMatBench, MatSciBench, and MaCBench, MatVerse consistently improves over its base model, outperforming GPT-5.2 on MatSciBench (63.59% vs. 59.6%) while also improving inference efficiency on calculation-intensive problems.

TL;DR: Training-free feature forecasting accelerates diffusion sampling by predicting features at skipped denoising steps, but open-loop caches trust the forecast in full at every skipped step. We show reliability can be observed from the cache itself: two forecasts agree where the feature trajectory is smooth and diverge where prediction turns hard, and their disagreement is a free runtime signal. RACER uses it as a training-free closed-loop controller that shrinks uncertain forecasts toward the last computed feature and, at the riskiest steps, refreshes the feature while repaying the cost by skipping a later one. At equal denoiser evaluations RACER improves the strongest open-loop baseline across SD3.5-Large, FLUX.1-dev, Wan2.1-14B, and HunyuanVideo on DrawBench, VBench, and COCO.

TL;DR: Positional Encodings (PEs) are essential for injecting structural information into GNNs, particularly Graph Transformers, yet their empirical impact remains insufficiently understood. We introduce a unified benchmarking framework that decouples PEs from architectural choices, comparing 8 GNN and Transformer models, 9 PEs, and 10 synthetic and real-world datasets — over 500 model-PE-dataset configurations. Weisfeiler-Lehman distinguishability and other expressiveness proxies do not reliably predict downstream performance; highly expressive PEs can even degrade real-world tasks, while several simple, previously overlooked model-PE combinations match or outperform recent state-of-the-art methods.

TL;DR: Deep learning for NMR structure elucidation is impeded by reliance on synthetic datasets, which introduces domain shifts on real-world experimental spectra, and by the lack of standardized evaluation protocols. NMRGym is the largest and most comprehensive standardized benchmark from high-quality experimental NMR data to date: 269,999 unique molecules paired with high-fidelity ¹H and ¹³C spectra, strict quality control, a scaffold-based split to prevent data leakage, and fine-grained peak-atom annotations, with an evaluation suite covering structure elucidation, functional group prediction, toxicity prediction, and spectral simulation.

TL;DR: Deep learning holds immense promise for spectroscopy, yet research and evaluation in this emerging field often lack standardized formulations. SpectrumLab is a unified platform that systematizes the field: a comprehensive Python library with data processing and evaluation tools and leaderboards, a SpectrumAnnotator module that generates high-quality benchmarks from limited seed data, and SpectrumBench, a multi-layered benchmark suite covering 14 spectroscopic tasks and over 10 spectrum types with spectra from over 1.2 million substances. Empirical studies with 18 cutting-edge multimodal LLMs reveal critical limitations of current approaches.

TL;DR: QCBench is a benchmark of 350 calculation-oriented chemistry questions spanning eight subfields — analytical, inorganic, organic, physical, polymer, and technical chemistry, biochemistry, and materials science — designed to test the domain-specific quantitative reasoning of large language models. Answers are scored with high-precision numerical tolerance and an answer-verification model, enabling systematic comparison across LLMs and chemistry categories.

TL;DR: A comprehensive and user-friendly deep learning library for protein-related tasks, with a benchmark evaluating CNNs, RNNs, transformers, and GNNs across protein function prediction, subcellular localization, protein-protein interaction, and structure prediction. We also introduce DeepProt-T5, a series of fine-tuned Prot-T5-based models that achieve state-of-the-art performance on four benchmark tasks while remaining competitive on six others.