Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 36 · 2026
Aug 31 – Sep 06
20 papers, repos and datasets selected this week.
Papers
Picked Sep 4
- PaperRepo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills→
It adds verified, reusable skill graphs distilled from repositories, boosting fixed-budget research agents 134.3% on MLE-bench.
- PaperSolarWM: Open Data and Scalable Training for Long-Horizon Video World Models→
Standardizes 10 video sources and trains 5B-33B backbones for hours-long rollouts from 5-second clips.
Picked Sep 3
- PaperStudentSim: Training LLM-based Student Simulators→
Across chess, L2 writing, and math, it beats GPT-5.4 on both fidelity and responsiveness; in chess: F=0.51/R=0.91 vs 0.23/0.72.
- PaperZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training→
Zero-shot real-robot success rises from 36.1% to 77.8% when scaling from target-robot data to 120k hours of video.
Picked Sep 2
- PaperDoes On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement→
It replaces noisy teacher-guided on-policy distillation with teacher-free entropy-scaled token penalties, gaining +35.41 Avg@32 on AIME24.
- PaperGenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling→
On ImageNet-256, SiT gets gFID 0.97 with CFG and 1.45 without; on text-to-image, MMDiT reaches 0.90 GenEval.
Picked Sep 1
- PaperBeyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models→
Gets 82.6% on LIBERO-Plus and 92.5% on RoboTwin 2.0; with 20% RoboCasa-GR1 data, beats full-data GR00T-N1.6 49.5% vs 47.6%.
- PaperCode as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning→
Code-as-World-VL reaches SOTA on QuantiPhy and beats leading proprietary models, including Gemini-3.1-Flash.
Picked Aug 31
- PaperSelf-OPD: On-Policy Distillation for Flow Matching Models without Teacher→
Authors report Self-OPD beats prior RL and teacher-based OPD on single- and mixed-reward benchmarks, without any task-specific teacher.
- PaperZero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization→
On 7 unseen RoboTwin tasks, it reaches 47.0% success, +29.5 points over the strongest video-action baseline.
Repos
Picked Sep 4
- RepoAMAP-ML/DreamX-Creator→
Open-sources a 7B joint generator and 1-step 2K refiner, with single-GPU inference and optional multi-GPU sequence parallelism.
Picked Sep 3
- Repoamap-cvlab/ABot-Recon→
Streams monocular 3D reconstruction causally over arbitrary-length videos with fixed 12-frame memory.
Picked Sep 2
- RepoNanako0129/sepia→
Targets structural AI tells that StoryScope says remain highly detectable: 93.2% macro-F1 from narrative features alone; surface edits only shift 95.5% to 93.9%.
Picked Sep 1
- RepoMLO-lab/LeVJEPA→
Eliminates EMA targets, predictors, and masking yet beats V-JEPA 2 with 5.6-20.8× less pretraining compute.
Picked Aug 31
- RepoTencent/WeMM-Embedding→
Unlike single-modality embedders, it unifies five input types in one space and retains 98.7% image/video performance at 256 dimensions.
Datasets
Picked Sep 4
- DatasetOpenDataArena/Spark-234K→
Unlike raw-text scientific corpora, it builds 234K QA pairs from paper-level reasoning skeletons with claims, evidence, quantities, assumptions, and limits.
Picked Sep 3
- Datasetmicrosoft/XL-DocBench→
Unlike broad corpora, it tests evidence-grounded QA over documents up to 2,935 pages with page-level evidence, cross-document questions, abstention, and only 38.36% top accuracy.
Picked Sep 2
- Datasetrazanihababdellatif/smart-city-digital-twin-ecosystem-dataset→
Built-in causal cascades and stress-time missingness make cross-system forecasting, outage modeling, and robust joins directly testable.
Picked Sep 1
- Datasethimalaya-ai/hermes-function-calling-nepali→
Makes Nepali tool-use evaluation measurable with exact call matching and script-transfer comparisons against fixed ground truth.
Picked Aug 31
- DatasetDISLab/Q-CARE→
Unlike larger QA datasets, it pairs 800 queries with 16 RAG outputs each for reference-free evaluation across both retrieval and generation.