Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 43 · 2026
Oct 19 – Oct 25
12 papers, repos and datasets selected this week.
Papers
Picked Oct 21
- PaperEgoXperience: Learning World-Grounded Egocentric Representations from Human Experience→
SOTA on scene and hand-object recovery; vs strongest baselines, errors drop 68.9% camera, 65.6% hand, 42.9% depth, with +28.0% object pose.
- PaperLearning from Fleet Experience: A Collective Learning Paradigm for Robots→
Across 5 real tasks, policies reach 93.3-99.2% success; on phone packaging, 96 robots lift success from 79.2% to 98.3%.
Picked Oct 20
- PaperLearning to Think and Explore in 3D World→
53.2% avg accuracy on 6 benchmarks: +22.0% over base Qwen, +1.0% over Gemini 3.6 Flash, with 50-70% fewer frames than open 3D methods.
- PaperData World: Scaling Data Environments from Scratch→
RL-zero on 2,135 synthetic tasks improves Qwen3.6-35B-A3B by +7.06 on DataWorld-Bench and +14.28/+20.90/+12.13 on 3 external benchmarks.
Picked Oct 19
- PaperFDE-Gym: The Last Mile of AI Delivery→
Best model scores 0.56 on 95 L3 tasks and resolves 14/95; even with all key facts on 23 tasks, it reaches >=0.8 uplift on only 10.
- PaperMLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering→
On MLE-Bench Lite, medal rate rises from 37.9% to 72.7%, matching GLM-5.3 and Kimi-K3 and beating same-scale baselines.
- PaperMAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments→
MAI-Gym-9B sets open <10B SOTA: 78.4% on MobileWorld, 76.2% on OSWorld-Verified; real-phone train app success rises 28.8%→48.1%.
Repos
Picked Oct 21
- RepoTencent-Hunyuan/Hy-Memory→
Supports zero-LLM-cost chat writes/search in lite mode and coding diff-chain capture with zero LLM calls.
Picked Oct 20
- RepoOpenMOSS/MOSS-Image→
Combines layout control, text rendering, 4K-token prompts, and public fine-tuning support in a compact 5B model.
Datasets
Picked Oct 21
- Datasethongzhuyi/VDEBench→
Makes precise document-image editing measurable by checking both local edits and full-page preservation on dense layouts.
Picked Oct 20
- Datasetprometheus-eval/into-the-unknown→
Makes open scientific research measurable end-to-end by pairing real follow-up tasks with complete agent outputs and evaluation artifacts.
Picked Oct 19
- Datasetdaimonrobotics/Daimon-Infinity→
Makes tactile-centric fine manipulation measurable at scale with synchronized touch, vision, language, and action streams.