Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 35 · 2026
Aug 24 – Aug 30
20 papers, repos and datasets selected this week.
Papers
Picked Aug 28
- PaperVBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning→
It provides 300 procedural visual reasoning tasks with deterministic scorers aligned across image, video, and interleaved generation.
- PaperGated Recurrent Transformers: Expressive Depth through Recurrent Modulation→
At isoFLOPS, a 3-layer recurrent model matches 12-layer GPT-2 Small; at large scale it gets 62% fewer params and 59% less peak decode memory.
- PaperFrontierChallenge: Evaluating Scientific Workflow Completion→
Best systems passed only 20/97 tasks (20.6%) despite Avg. Score up to 87.9; electrochemistry hit 94.9 Avg. Score with 0% Pass Rate.
Picked Aug 27
- PaperOn-Policy Self-Distillation in Diffusion Models→
It converts image-level rewards into explicit denoising targets, winning 19/20 settings with up to 44.0% gains and 40-63% less GPU-hours.
- PaperAnnotations as Rollouts: Efficient and Scalable Reinforcement Learning for Video MLLMs→
Cuts step time to 2.2x SFT vs 4.9x for GRPO+CoT; Video-ORA-9B lifts temporal mIoU 62.5->66.0 and VSI-Bench 73.1 vs GPT-5's 55.0.
Picked Aug 26
- PaperReWorld: An Interactive World Model with Long-Horizon Memory→
Best among 6 recent interactive world models on control and generation quality; 11.95° rotation error, and recalls the start view after 64s rollouts with a 12-chunk cache.
- PaperEchoWM: Open and Enterable Omnimodal World Models→
Ranks first on WBench Navigation and achieves top-tier visual quality on SANA-WM-Bench, while supporting synchronized long-horizon audio and speech.
- PaperLet's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts→
Extrapolates optimal LR to 10T tokens with R^2=0.95 and validates it in stable pretraining of a 155B total / 17B active MoE.
Picked Aug 25
- PaperAgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale→
RL on 4,783 synthesized worlds lifts Qwen3.5-4B from 12.3 to 15.7 on EnterpriseOps-GYM and 45.9 to 56.0 on AIME26.
- Paper4DAnyone: Create Anyone in 4D from a Casual Monocular Video→
On DNA-Rendering and DyMVHumans, it outperforms prior methods in novel-view video quality and downstream 4DGS reconstruction, with robust in-the-wild generalization.
- PaperFreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution→
On RTX 5090, it reaches 77-83 tok/s on Qwen3.6-35B and 22-25 tok/s on DeepSeek-V4-Flash, with 1.5-2.3x higher decode throughput than edge baselines.
Picked Aug 24
- PaperEnvHarness: Awakening Static Worlds for Agent Learning→
Unlike papers generating new environments, it retrofits existing verifiable ones via rollout-driven modular wrappers.
- PaperCo-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL→
Improves LLMs by 3.0-8.6% on 7 text benchmarks and VLMs by 2.3-7.2% on 4 multimodal benchmarks; beats prior label-free methods by 0.8-2.0%.
Repos
Picked Aug 27
- RepoApodexAI/FrontierAgent→
Unlike most agent runtimes, it ships built-in evaluation and shows parallel Agent Team beating ReAct 38.5 vs 34.4 on APEX.
Picked Aug 25
- RepoLeonxlnx/unlazy→
Unlike orchestration runtimes, it enforces file-based completion gates and stop blocking, reporting 1.6-3.9× more effort and 4-10 self-found defects fixed.
Picked Aug 24
- RepozhengzihaoPKU/RoboECC→
Targets VLA deployment with offline split search plus online bandwidth-aware cut adjustment, unlike static edge-cloud partitioning.
Datasets
Picked Aug 28
- Datasetvaquill/open-india-law→
Unlike narrower Indian legal sets, it unifies 13.7M official cases, tribunals, and section-level legislation in one schema.
Picked Aug 27
- Datasetopen-gigaai/GigaBrain-0.7-SampleData→
Provides permissively licensed sample VLA data linked to large-scale multi-embodiment training for studying embodied foundation model pipelines.
Picked Aug 26
- Datasethamzabagirsakci/turkish-court-decisions→
It offers 11 million CC0 full-text Turkish court decisions spanning 1962–2026 across five court systems.
Picked Aug 24
- DatasetPatronusAI/figmatrace→
Unlike final-design datasets, it captures 3,469 expert Figma action-frame trajectories over 126 long-horizon tasks.