Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 28 · 2026
Jul 06 – Jul 12
20 papers, repos and datasets selected this week.
Papers
Picked Jul 10
- PaperAccurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning→
Unlike textized or task-specific models, it natively reasons across proteins, molecules, and crystals with inspectable structure-linked traces, reaching SOTA on 67 of 86 benchmarks.
- PaperRynnWorld-Teleop: An Action-Conditioned World Model for Digital Teleoperation→
Runs at 40+ FPS on one H100; policies trained only on generated data show zero-shot Sim2Real transfer, and adding synthetic data consistently improves success rates.
Picked Jul 9
- PaperRynnWorld-4D: 4D Embodied World Models for Robotic Manipulation→
Authors report temporally and spatially coherent 4D predictions and state-of-the-art real-world dexterous bimanual manipulation performance.
- PaperHierarchical Sparse Attention Done Right: Toward Infinite Context Modeling→
It learns chunk retrieval inside forward attention, preserving accuracy while extrapolating 64x longer contexts.
Picked Jul 8
- PaperPixWorld: Unifying 3D Scene Generation and Reconstruction in Pixel Space→
Beats prior generation methods and matches SOTA reconstruction: e.g., 18.88 PSNR on Re10K 1-view generation and 28.58 PSNR on Re10K 8-view reconstruction.
- PaperThe Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning→
On Qwen3-4B and 1.7B under FP8 rollout mismatch, MIPU gives the best average reasoning performance and more stable training than prior mismatch-handling baselines.
Picked Jul 7
- PaperABot-M0.5: Unified Mobility-and-Manipulation World Action Model→
Claims state-of-the-art long-horizon task success and fine-grained control on mobile/manipulation benchmarks, with real-robot success shown on tasks like Find Toaster and Arrange Flower.
- PaperEvolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks→
Across 22 held-out tasks, EFT models beat their base models by 10.22% on average; scaling training tasks from 15 to 355 adds 14.1% average gain.
Picked Jul 6
- PaperAgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents→
On fixed A0, no-store wins 3/10 while skill-enabled setups reach 6/10; the release includes 298 tagged trajectories and prompts for reproducible memory studies.
- PaperProgram-as-Weights: A Programming Paradigm for Fuzzy Functions→
Compiles natural-language fuzzy specs into reusable LoRA programs, beating prompted 32B models with 50x less inference memory.
Repos
Picked Jul 10
- Repofrenzymath/Danus→
Unlike agent systems that revise outputs, it stores only verifier-approved claims in a shared fact graph, yielding 3,157 verified facts and 8,616 dependency edges in one run.
Picked Jul 9
- RepoRobbyant/lingbot-vision→
Releases self-supervised ViT backbones specialized for dense patch features up to 1.1B parameters.
Picked Jul 8
- Repoopen-gigaai/giga-world-1→
Covers end-to-end workflows and ships Stage-2 distillation that cuts diffusion from 20 steps to 4-6 steps.
Picked Jul 7
- RepoXingyu-Zheng/MrFlow→
Unlike retrained fast samplers, it accelerates pretrained flow-matching models 8.25x-21x training-free.
Picked Jul 6
- Reporolekkona/bghira-bark→
Cuts long jobs on 3x 4090s from 2-5 min to 30-60 s via true batching, torch.compile, and GPU-side sampling.
Datasets
Picked Jul 10
Picked Jul 9
- Dataseturadkr/biorxiv-life-sciences-preprints-corpus→
Provides a clean labeled biology corpus with complete category targets, making domain classification and embedding evaluation directly measurable.
Picked Jul 8
- Datasetosunlp/early-experience→
Makes early-experience training directly reproducible and comparable; IWM/SR beat IL across reported agent benchmarks.
Picked Jul 7
- DatasethandsomeYun/TIC-VLA→
Unlike navigation datasets, it aligns frame-level chain-of-thought with synchronized actions across three benchmarks.
Picked Jul 6
- DatasetSyn4D/Syn4D→
Unlike real-scene datasets, it provides synthetic multiview 4D videos with aligned depth, masks, tracking geometry, and object meshes.