Papers.Data.Code @papersdatacode

📋 ML Weekly Recap · Oct 19 – Oct 25 ⚡ Trends ▸ Synthetic executable environments are scaling agent training for GUI, data, and computer-use tasks ▸ Embodied world models increasingly use shared 3D/egocentric representations across humans and robots ▸ Efficiency-focused generative systems emphasize few-step or one-step rendering and video enhancement 🧭 TL;DR 📄 Learning from Fleet Experience: A Collective Learning Paradigm for Robots Chao Yu, Shu'ang Yu, Hao Lin et al. Fleet-log retrieval learning boosts real-robot success without new expert corrections. 📄 arxiv.org/abs/2610.25736 ⭐ Hy-Memory Practical persistent agent memory with hierarchical recall and zero-cost operations. ⭐ github.com/Tencent-Hunyua… 💡 Agents are shifting from static benchmarks toward executable, self-improving environments. → t.me/papersdatacode…

Links: ⭐ github.com/OpenImagingLab…

💻 OpenImagingLab/FlashVSR_Plus Runs one-step diffusion video super-resolution with untiled 4K inference using hybrid sparse attention and streaming decode options. Key points: • Untiled 4K inference on a single 24GB RTX 4090 • Hybrid attention gives up to 2× DiT acceleration • Cuts peak memory by 45% vs standard DiT attention Shows that high-quality diffusion VSR can reach practical untiled 4K inference on a single consumer GPU. ⭐ 142 stars

@zhaoywang_CS Links: 📄 arxiv.org/abs/2610.26828 ⭐ github.com/MSR-Orchard/Wi…

📄 Agent World Model in the Wild: Computer-Use Agent Environment Synthesis by Zhaoyang Wang (@zhaoywang_CS), Qianhui Wu et al. Builds open-source synthetic PC worlds with GUI, CLI, and MCP access, then generates and validates cross-interface tasks via multi-agent execution gates. Key points: • 341 long-horizon environment-task pairs synthesized • Tasks can require 100+ steps; GPT-6-Astra solves only 60% • Qwen models trained only in WildAWM improve on 8/8 OOD benchmarks Open-source synthetic computer worlds can train agents that generalize broadly beyond the generated environments.

@Enderfga Links: 📄 arxiv.org/abs/2610.28630 ⭐ github.com/Enderfga/Solar…

📄 SolarRenderer: Few-Step Generative Reference Rendering by Guian Fang (@Enderfga), Shengju Qian et al. Uses native reference tokens as aligned structural controls, then adds a separate 4-step LoRA so one frozen backbone renders many structure types fast. Key points: • 4-step rendering is 12.3x faster than 50-step sampling • SOTA on both in-distribution and 16-type generalization benchmarks • Generalizes to 11 unseen structure types, including skeletons and point clouds A single reference-pathway control interface can generalize across heterogeneous and even unseen structural video representations.

Links: ⭐ github.com/MemTensor/MemT… 📄 arxiv.org/abs/2610.26825

💻 MemTensor/MemT-1.0 Adds persistent native memory to chat models so they can save, clear, and later reuse past interactions without replaying the original context. Key points: • Persistent state is ~66× smaller than full KV cache at 128K history • MemT-1.0-Flash and Nano rank 1st and 2nd on 11 memory benchmarks • Supports memory snapshots and restore via saved server-side state files Shows that long-horizon conversational memory can be carried in fixed-size latent state instead of replaying full text history. ⭐ 291 stars

Links: 📄 arxiv.org/abs/2610.26855 ⭐ github.com/Ropedia/ropedi…

📄 Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience by Tianqi Liu, Hanfei Zhao et al. Replaces robot-specific action conditioning with shared camera-pose and skeleton controls, so one video world model can predict human and robot loco-manipulation. Key points: • SOTA across DreamDojo, Ψ0, ω-HOME, Xp-Teleop, and AgiBot • Human scaling: camera ATE drops 2.21→1.79 cm from 1k→10k h pretraining • Robot scaling on Ψ0: LPIPS improves 0.427→0.365 from 1k→10k h Large-scale human egocentric pretraining transfers to robot world modeling when locomotion and manipulation are expressed in a shared kinematic format.

Links: 📊 huggingface.co/datasets/hongz… 📄 arxiv.org/abs/2602.00122 ⭐ github.com/zion-zion-zion…

📊 VDEBench by hongzhuyi Contains 3,506 dense document-page edit examples with image input, text instruction, target image, and optional masks for English and Chinese text/table edits. Key points: • 3,506 test examples total • 7 edit types: 3 text and 4 table tasks • Bilingual English and Chinese document images Useful for benchmarking models that must edit document images locally while preserving layout, typography, and nearby content. 📥 1.6k downloads

Links: 📄 arxiv.org/abs/2610.25736

📄 Learning from Fleet Experience: A Collective Learning Paradigm for Robots by Chao Yu, Shu'ang Yu et al. Failed rollout contexts retrieve similar successful segments from fleet logs, and those actions are prioritized to fine-tune a shared policy without new expert actions. Key points: • Phone packaging: 79.2% to 98.3% success with 96 robots • Across 5 tasks: 93.3-99.2% success over 120 eval trials per task • Removing failure-guided retrieval drops phone packaging from 86.7% to 64.2% Shared autonomous experience can materially improve deployed robot reliability without scaling expert corrective supervision.

Links: 📄 arxiv.org/abs/2610.26625 ⭐ github.com/Ropedia/EgoXpe…

📄 EgoXperience: Learning World-Grounded Egocentric Representations from Human Experience by Xinhao Ji, Hui En Pang et al. A joint feed-forward model for egocentric video that predicts camera, depth, hands, and object poses together in one metric world frame. Key points: • HOI4D camera nATE reduced by 68.9% vs strongest baseline • EgoX-Bench hand W-MPJPE and depth AbsRel cut by 65.6% and 42.9% • HOT3D object AUC-S50 improves by 28.0% vs strongest baseline Joint world-space supervision across camera, scene, hands, and objects yields much more coherent egocentric reconstruction than separate pipelines.

Links: 📄 arxiv.org/abs/2610.21656 ⭐ github.com/Zhanli-Li/Data…

📄 Data World: Scaling Data Environments from Scratch by Zhanli Li, Zeyuan Xu et al. A mechanism-first pipeline that creates synthetic file environments, audited tasks, and rubric rewards from explicit world models rather than existing datasets. Key points: • +20.90 on DABstep-Hard over Qwen3.6-35B-A3B • +14.28 on DAComp over the base model • +7.06 on DataWorld-Bench over the base model Explicit generative worlds can supply scalable, verifiable supervision for training data-analysis agents on open-ended insight tasks.

Links: 📊 huggingface.co/datasets/prome… 📄 arxiv.org/abs/2610.19159 ⭐ github.com/prometheus-eva…

📊 into-the-unknown by prometheus-eval Contains 15 paper-based research tasks and 135 agent submissions with PDFs, event logs, and working files for open-ended follow-up paper writing. Key points: • 15 tasks from published or submitted papers • 135 submissions from 9 agents • Includes PDFs, JSON trajectories, logs, and file-system snapshots Useful for benchmarking AI agents on end-to-end scientific research and follow-up paper generation. 📥 1.3k downloads