Only meaningful ML signals. Papers, repos & datasets. Selected, not collected. Highlights, trends & TL;DRs. 📄💻📊 → https://xitter.cf/t.co/nGRf9ZTJks
Joined April 2026
- Tweets556
- Following22
- Followers66
- Likes133
📊 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
ALT Data statistics charts showing fund counts, scale trends, and top regions distribution
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.
ALT Full pipeline diagram showing collection, encoding, retrieval, and prioritized fine-tuning method stages
📄 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.
ALT Full pipeline diagram showing tokenization, scene-entity blocks, and output heads
📄 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.
ALT Full pipeline diagram showing world spec, canonical construction, observation and evaluation stages
📊 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
ALT Pipeline diagram showing evaluation framework with direction, bar, and rubrics
📄 Learning to Think and Explore in 3D World
by Yijiang Li, Bingyang Wang et al.
A single-view interactive VLM that learns to navigate for evidence instead of consuming full 3D or video inputs. It is SFT-initialized, then RL-trained from answer outcomes.
Key points:
• 53.2% avg accuracy on 6 benchmarks, +22.0% over base Qwen
• Beats Gemini 3.6 Flash by 1.0% average accuracy
• Uses 50-70% fewer input frames than prior open-source 3D methods
Outcome-based RL can teach VLMs not just to reason from observations, but to acquire the right observations for 3D tasks.
ALT Compares paradigms and shows active interaction pipeline with rollout example illustrating core contribution
📄 MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Engineering
by Shangheng Du, Jinxin Shi et al.
Trains an agentic LLM to autonomously organize ML experiments, using outcome-filtered SFT and windowed RL over decisions between experiment results.
Key points:
• 72.7% medal rate on MLE-Bench Lite vs 37.9% for its base model
• Matches GLM-5.3 and Kimi-K3 at 72.7% medal rate on MLE-Bench Lite
• 84.8% Med+ and 50.0% gold on MLE-Bench Lite
Long-horizon post-training on real experiment feedback can turn a mid-size coding model into a competitive autonomous ML engineering agent.
ALT Full pipeline diagram showing MLE tasks, interaction harness, and agentic post-training method
📄 MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments
by Panrong Tong, Quyu Kong et al.
A unified GUI RL platform that shares the exact rollout path for eval and training, and scales learning with synthesized resettable apps and verified tasks.
Key points:
• 78.4% on MobileWorld, beating UI-Venus-2-9B's 75.2%
• 76.2% on OSWorld-Verified vs UI-Venus-2-9B's 70.8%
• Real-phone train app success improved 28.8% to 48.1% after RL
Shared infrastructure plus synthesized, verifiable GUI tasks can make online RL transfer beyond mocked environments to real app use.
ALT Comprehensive system architecture showing RL training, environment management, and GUI instances pipeline
📄 FDE-Gym: The Last Mile of AI Delivery
by Dongxu Yang, Hongda Yan et al.
A benchmark for ambiguous enterprise delivery tasks that scores deployed solutions by business KPI uplift minus interaction cost, not binary task completion.
Key points:
• Best L3 score is 0.56 mean; only 14 of 95 tasks resolved
• Oracle per-task model choice still reaches only 0.70 and 30/95 resolved
• Top models drop from 0.73-0.82 to 0.39-0.62 when constraints must be found
Shows the main gap in enterprise agent deployment is uncovering unstated constraints and turning them into reliable KPI gains.
ALT Pipeline diagram showing data sources, environments, tasks, and benchmark construction stages
Weekly Top · Oct 12 – Oct 18
Follow for next week’s top
📄 Paper: S-Space: Exploring Spatial Workspace in Multimodal Language Models
It identifies a fixed linear 3D spatial subspace in MLLM hidden states and causally edits it.
💻 Repo: NandhaKishorM/vegaml
Unlike recent typed-decision repos, it adds local multimodal decisions with conformal sets and abstention.
📊 Dataset: 2077AIDataFoundation/TouchScale
Unlike recent egocentric manipulation datasets, it adds synchronized 880-taxel full-hand tactile gloves across 100 released human-demonstration hours.
📋 ML Weekly Recap · Oct 12 – Oct 18
⚡ Trends
▸ Frozen models increasingly improve via external memory, tools, harnesses, and rulebooks
▸ Decision-style scoring replaces autoregressive generation for faster, cacheable structured inference
▸ Interactive agent training shifts toward synthetic worlds, recursive environment repair, and verifiable rewards
🧭 TL;DR
📄 Cortex: Consolidating Context into Parametric Memory
Jintian Zhang, Feng He, Xinnian Liang et al.
Converts context into persistent, mergeable memory with strong retention-recall tradeoffs.
📄 arxiv.org/abs/2610.20399
⭐ vegaml
Practical local decision API adds calibration, abstention, and long-context typed outputs.
⭐ github.com/NandhaKishorM/veg…
💡 Systems increasingly externalize memory and decision structure for efficient agent improvement.
→ t.me/papersdatacode_digests
ALT ML Weekly Recap · Oct 12 – Oct 18. External memory augmentation; Decision-style scoring; Synthetic agent worlds.