Papers.Data.Code
The Weekly ML Signal
Selected, not collected.
Week 31 · 2026
Jul 27 – Aug 02
20 papers, repos and datasets selected this week.
Papers
Picked Jul 31
- PaperMindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis→
Raises Qwen3.6-27B on ProgramBench from 37.98% to 49.51%, with gains on all 7 unseen SE benchmarks, including +31.00 on RepoZero-C2Rust.
- PaperTurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM→
Unlike LLM-centered VLAs, it delivers 32 Hz control at 0.9 GB VRAM.
- PaperHumanCLAW: Can Vision-Language Models Act Through a Body?→
It isolates VLM decision quality from humanoid motor failures across 1,218 episodes.
Picked Jul 30
- PaperParallel Decoding Distillation for Fast Image and Video Generation→
Achieves SOTA at 4-8 NFE on LTX-2.3, Wan 14B, and Qwen-Image, with significantly better video diversity than prior baselines.
- PaperMage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model→
Cuts visual tokens by over 75% and delivers up to 3.5x inference speedup, while Mage-VL-4B matches Qwen3-VL-4B on static tasks and beats larger Phi-4 vision baselines.
Picked Jul 29
- PaperKimi K3: Open Frontier Intelligence→
Unlike most open models, it combines 2.8T MoE pretraining with test-time RL, 1M context, and 76.9 ProgramBench/30.8 AutomationBench results.
- PaperClinFusion: A Vision-Centric Multimodal LLM System for Holistic Medical Understanding→
SOTA on 20/24 benchmarks; beats GPT-5.2 and Gemini-3-Flash on 13/16 multimodal benchmarks; radiologists ranked its reports highest.
- PaperThe Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation→
CoT world modeling reaches 93.1 vs 0.3 on long tasks; adding 5% long-horizon data lifts 0.0 to 47.2; OPD beats GRPO under low-quality, long-horizon settings.
Picked Jul 28
- PaperRobostral Navigate→
Sets SOTA on R2R-CE at 77.4% SR and RxR-CE at 75.1% SR using only one RGB camera.
- PaperStale but Stable: Staleness-Adaptive Trust Regions for Stabilizing Asynchronous Reinforcement Learning→
Best AIME24 avg@8 is 35.83 at lag 1 and 34.79 at lag 8 with SAT-GSPO w/ R3; SAT-GSPO reaches 34.17 at lag 1.
- PaperThree-Body Scattering for Generative Modeling→
On ImageNet-256 at NFE=1, it reaches FID 1.63 with latent DiT-XL and 2.23 with pixel PixelDiT-XL.
Picked Jul 27
- PaperAREX: Towards a Recursively Self-Improving Agent for Deep Research→
AREX-Base leads WideSearch 82.0, HLE 52.4, and xbench 71.0, and is competitive on BrowseComp 82.5 and DeepSearchQA 89.9.
Repos
Picked Jul 30
- Reposii-research/tau-0-vla→
Combines world-model-guided test-time subtask search with cross-embodiment control trained on 40,115 real-robot hours.
Picked Jul 28
- Repoawesome-4dgs/awesome-4dgs→
Collects 79 papers into task-specific sections with contribution summaries, code links, and representation tags for faster literature triage.
Picked Jul 27
- ReposlvDev/esp32-ai→
Runs a 28.9M-parameter LLM on an $8 ESP32-S3 by flash-streaming rows, far beyond prior ~260K-parameter microcontroller models.
- Reponwneel/Video-ChatGPT→
Unlike recent video repos, it pairs a video-chat model with a 100K instruction dataset and benchmark, while beating prior zero-shot baselines on four QA sets.
Datasets
Picked Jul 31
- DatasetByteDance/TableVerse→
Unlike synthetic tabletop sets, it reconstructs 100K real-image scenes with physics and expert trajectories.
Picked Jul 30
- Datasetkylefengkfeng209/screen-time-vs-mental-health-ml-ready→
Time-separated predictors and outcomes make future depression risk measurable beyond same-day screen-time correlations.
Picked Jul 29
- Datasetsimple-world-lab/HiFi-UMI-2K→
Unlike mostly synthetic or single-view corpora, it provides 2,000 hours of calibrated six-view bimanual manipulation with 20D trajectories, gripper states, language, and QC metadata.
Picked Jul 27
- Datasetvinayak9421/corporate-laws-of-india-structured-json→
Unlike generic legal corpora, it structures 950+ Indian statutory sections across 14 Acts for section-level RAG and compliance retrieval.