Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 30 · 2026
Jul 20 – Jul 26
20 papers, repos and datasets selected this week.
Papers
Picked Jul 24
- PaperSelf Gradient Forcing: Native Long Video Extrapolation→
With only a 5s training window, SGF matches Self Forcing at 5s and improves native extrapolation at 60s and 240s, reaching videos of several minutes.
- PaperMasked Visual Actions for Unified World Modeling→
Beats Ctrl-World on DROID and unseen BEHAVIOR: LPIPS 0.0945/0.123 vs 0.362/0.196, with higher SSIM and PSNR.
- PaperRynnBrain 1.1: Towards More Capable and Generalizable Embodied Foundation Model→
It adds native 3D grounding, contact-point prediction, and masked cross-embodiment actions, then tops three embodied benchmarks.
Picked Jul 23
- PaperABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPU→
Streams 720P at up to 16 FPS on one RTX 5090 with 1.2 s latency and ~19 GiB VRAM; authors report competitive controllability and coherent long-horizon rollouts.
- PaperText Template Tokens Are Implicit Semantic Registers in Diffusion Transformers→
Template-token heads can be pruned training-free for 20% attention FLOPs reduction with only a 1.4-point GenEval drop.
Picked Jul 22
- PaperTimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs→
It models temporal grounding as interval sets, using matching-free Wasserstein RL and multi-span labels.
- PaperLoop the Loopies!→
At equal compute, it beats a vanilla 30B-A3B baseline and reaches 92.1% AIME, 94.21% AMC, plus gold-medal IMO/IPhO performance.
Picked Jul 21
- PaperUnderstanding Reasoning from Pretraining to Post-Training→
Post-RL performance is predicted by pretraining loss, and RL reward-curve slope grows about linearly with pretraining tokens; the same pattern appears on a 1B math LM.
- PaperXiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories→
Uses 100K+ hours of auto-labeled real trajectories, replacing scarce teleoperated task labels.
Picked Jul 20
- PaperSEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning→
Best average performance across ALFWorld, Search-based QA, and WebShop: 91.8, 45.7, and 78.9, beating prior baselines shown in Fig. 1.
- PaperRoboTTT: Context Scaling for Robot Policies→
On real robots, it improves average task score by 87% over single-step GR00T, and is the only method to complete the 5-min 10-stage task.
Repos
Picked Jul 23
- Repopaxlabs-inc/machine-genome→
Provides append-only, cryptographically signed provenance with Merkle-log inclusion and consistency proofs, without requiring a blockchain or vendor-specific stack.
Picked Jul 22
- RepoRed-EAD/FlowBlock→
Unlike draft-model decoding work, it parallelizes diffusion LM blocks training-free, reaching 2.95× TPS.
Picked Jul 21
- RepoTencent-Hunyuan/Hy-Embodied-RxBrain-1.0→
Unlike single-task world models, it unifies VQA, 4-frame prediction, and action-image planning in one released 6.2B model.
Picked Jul 20
- RepoPengZhang64/circuit-framework→
Unlike trading-agent repos, it enforces deterministic risk approval and preserves NO_TRADE outcomes.
Datasets
Picked Jul 24
- DatasetGlint-Research/VulnSetSimple→
Per-row tool renaming makes tool-binding generalization measurable instead of letting models memorize fixed API names.
Picked Jul 23
- Datasetgeorgymamarin/geosteering-world-cup-2021-expert-interpretations→
Unlike most annotation datasets, it provides 176 expert labels on identical cases with known truth, enabling direct measurement of expert noise and aggregation accuracy.
Picked Jul 22
- DatasetFINAL-Bench/Aether-7B-5Attn-checkpoints→
Releases three raw 7B pretraining checkpoints with weights, config, and tokenizer for direct trajectory analysis.
Picked Jul 21
- Datasetmicrosoft/RESOURCE2SKILL→
Makes reusable agent skills a concrete, inspectable artifact with provenance and validation across real software domains.
Picked Jul 20
- Datasetopen-alchemy/code-alchemy→
Near-trillion-token scale unifies rewriting, QA, dev tasks, dialogue, and execution traces.