Top 10 in AI, week of 12 October 2026

Ranked by upvotes on Hugging Face Daily Papers, for papers featured 12 to 18 October 2026 (counts as of 18 October). Each item says in plain words what the paper reports. The results are the authors' own claims, not checked by us.

1. Waypoint-OPD passes a teacher model's reinforcement-learning gains to a student model, weighted by whether the teacher's early and late shifts agree. It suppresses the shift where they point in opposite directions. The authors tested Qwen3-4B teachers with Qwen3-8B, Qwen3-14B and Qwen3-30B-A3B students on math and code reasoning, and report the highest domain-average score among the compared methods in all six domain and student settings. 667 upvotes. Paper: arXiv 2610.21601 · Discussion · Code
2. OmniCompass is a new benchmark for how well AI models find and understand things in video. It has two tracks: vision-language (820 videos, 2,460 instances) and audio-visual (240 videos, 720 instances), testing when a target appears, where it is and what happens. Across 50 multimodal models, the authors report that open-source models still lag proprietary ones, especially on audio-visual understanding, and that precise timing and location remain hard even for leading models. 578 upvotes. Paper: arXiv 2610.18829 · Discussion · Code
3. Jeva is a model design for fast decisions among options. The paper says current designs miss four needs: telling a missing answer apart from uncertainty, giving the same answer when options are reordered, handling growing sets of options, and reusing work across requests. The authors report near-perfect detection of missing answers, no changed decisions under any tested reordering, scaling to thousands of candidates, and lower latency from reuse. 418 upvotes. Paper: arXiv 2610.18773 · Discussion · Code
4. VideoJev answers questions about video by scoring the options instead of writing out an answer. Built on pretrained multimodal language models, its 4B and 35B-A3B versions improve mean mIoU over their backbones by 6.7 and 10.8 points on three temporal grounding benchmarks, with measured speedups of 2.8 to 3.1 times. On four video question-answering benchmarks the authors report performance comparable to the backbones in Thinking mode, with measured speedups above 100 times. 367 upvotes. Paper: arXiv 2610.18615 · Discussion · Code
5. MJev-Spatial answers spatial questions, such as how far away an object is, in one scoring pass instead of generating text. Built on Qwen3.5-4B and trained on 12.3M question-answer pairs, it scores 65.64% overall among eight tested models, ranks first on seven of twelve benchmarks and beats its unmodified base model on all twelve. The authors say it runs 2.5 times faster than the fastest autoregressive baseline. 354 upvotes. Paper: arXiv 2610.22276 · Discussion · Code
6. Any Time, Any Space (ATAS) keeps long generated films consistent from shot to shot. It turns a screenplay into shot requests with explicit location, camera and event-time references, and checks a memory of scenes and events before and after each clip is generated. Across four video generators, the authors report better scores on all cross-shot consistency measures and the highest human ratings on all six criteria. 343 upvotes. Paper: arXiv 2610.20571 · Discussion · Code
7. Player-1 trains AI agents to play games using synthetic practice environments. It fine-tunes computer-use models on 6,044 gameplay trajectories, then trains with reinforcement learning in sixteen synthetic 2D and 3D environments. The authors report that Player-1-9B reaches 38.33% success on MineDojo-300, against 12.63% for the strongest baseline they evaluated. 334 upvotes. Paper: arXiv 2610.20584 · Discussion · Code
8. DiffusionART speeds up reinforcement learning for image and video generation models. On HunyuanVideo-13B it reaches the late-stage reward levels of DiffusionNFT and Flow-GRPO-Fast with 12.02 and 11.72 times fewer GPU hours, the authors report, and it needs no per-task tuning. 320 upvotes. Paper: arXiv 2610.12572 · Discussion · Code
9. GroundingPI is a 4B model that points to objects in images, for robots and driving. The authors report an average of 73.68% across 34 grounding benchmarks against 44 baselines, above the larger GPT-6 Astra at 71.54%. Used as a visual backbone, it gets an average open-loop L2 error of 0.296 m on nuScenes. 320 upvotes. Paper: arXiv 2609.39601 · Discussion · Code
10. OpenGCUA is an open stack for building agents that use computers. Its first release raises the mean partial score by 19.6 points across all eight models tested on OSWorld-v2, including GPT-6 Astra and Claude Opus 5, and cuts reported steps per task by an average of 53.0%. The authors release their implementation, prompts and evaluation trajectories. 316 upvotes. Paper: arXiv 2610.22200 · Discussion · Code

Follow Top 10 in AI by email on its page. Spotted a mistake? Anyone can propose a change to this page.