Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 29 · 2026
Jul 13 – Jul 19
20 papers, repos and datasets selected this week.
Papers
Picked Jul 17
- PaperRing-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning→
Scales verifiable-reward zero-data RL to a 1T-parameter LLM with concise reasoning.
- PaperFunction-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models→
Uses PDG-selected whole-function FIM mid-training to simulate tool-return context, improving SWE-Bench-Verified by up to 3.2 points across three code models.
Picked Jul 16
- PaperXiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model→
Beats GPT-Image-2.0 in human evals, ranks #1 on World Arena, and raises pi_0.5 OOD success from 36.9% to 63.2%.
- PaperRead It Back: Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation→
On BAGEL, gains reach +10.0 TIIF-Bench and +5.5 GenEval over baseline; it also beats AlphaGRPO by +6.3 TIIF and +3.3 GenEval.
Picked Jul 15
- PaperABot-AgentOS: A General Robotic Agent OS with Lifelong Multi-modal Memory→
Memory scores reach 87.5 LoCoMo, 59.9 OpenEQA, 88.6 Mem-Gallery, 76.5 NExT-QA; self-evolution lifts them to 88.7, 60.4, 89.0.
- PaperWeak-to-Strong Generalization via Direct On-Policy Distillation→
It distills the teacher’s pre/post-RL log-policy shift on student rollouts, beating step-matched RL with 10-point AIME gains in four hours on eight A100s.
Picked Jul 14
- PaperScalable Visual Pretraining for Language Intelligence→
On matched corpora, VP beats text pretraining by up to +3.22 GPQA and +2.10 MMLU-Pro, using only 25% of the token budget.
- PaperVideo Generation Models are General-Purpose Vision Learners→
Matches or beats specialist models on depth, normals, pose, segmentation, and keypoints, and reaches comparable results to D4RT/VGGT-Ω with 7× to 500× less training data.
Picked Jul 13
- PaperIdeas Have Genomes: Benchmarking Scientific Lineage Reasoning and Lineage-Grounded Idea Generation→
IG-Bench spans 10 domains; best exact lineage-reasoning accuracy is only 27.3%, showing current LLM scientists struggle with compositional lineage reasoning.
- PaperDual Latent Memory in Vision-Language-Action Models for Robotic Manipulation→
On LIBERO it reaches 97.6% success, beating MemoryVLA by 1.1 pts and CogACT by 4.4; on SimplerEnv-Bridge it gets 73.9%, +16.6 over CogACT.
Repos
Picked Jul 17
- Repolucidrains/x-jepa→
Provides a packaged JEPA experimentation scaffold in Python 3.10+ with tests and Torch 2.4+ support.
Picked Jul 16
- Repocdfan0627/LongE2V→
One framework covers 3 event-to-video tasks, with released weights, data, and evaluation outputs for direct reproduction.
Picked Jul 15
- Repomira-wm/mira→
Unlike mostly single-agent or offline world models, it simulates synchronized four-player Rocket League at 20 FPS on one GPU and releases a matching dataset.
Picked Jul 14
- Repoxinyan-cxy/OpenCoF→
Unlike speed-focused video generators, it adds a 17,312-video supervision set for temporal reasoning.
Picked Jul 13
- RepoRobbyant/lingbot-video→
Targets embodied video generation with MoE speedups, unifying T2I/T2V/TI2V and posting 0.620 RBench.
Datasets
Picked Jul 17
- Datasetai-sage/TimeGround-1M→
Unlike typical speech datasets, it pairs 14.2k hours of long audio with three aligned timestamped supervision views up to 120 minutes.
Picked Jul 16
- Datasetriyagarg0314/ai-model-arena-rankings-2023-2026→
Unlike point-in-time leaderboards, it provides 97,377 historical Elo snapshots across five arenas.
Picked Jul 15
- Datasetkylefengkfeng209/how-fast-do-you-forget-13m-duolingo-sessions→
Unlike most datasets, it captures real-world forgetting curves from 12.85M learner sessions across six languages.
Picked Jul 14
- DatasetRaymvp12/focusmil-camelyon16→
Makes shortcut reliance in MIL directly measurable: if slide AUC drops below 0.5, the model is using negative evidence.
Picked Jul 13
- Datasetabhiit/cifar-10-image-dataset-original→
Provides a standard, fixed benchmark for directly comparing image classifiers on small natural images.