Top 10 in AI, week of 2 November 2026

Ranked by upvotes on Hugging Face Daily Papers, for papers featured 2 to 8 November 2026 (counts as of 8 November 2026). Each item says in plain words what the paper reports. The results are the authors' own claims, not checked by us.

1. The authors review agentic computer graphics, where language-model agents write and refine graphics code. They introduce a unified taxonomy for describing agentic systems across graphics tasks, and an agentic literature system that helps researchers analyze papers at scale. The authors also discuss open challenges, including coordinating agents and tools and reliably evaluating interactive graphics systems. 144 upvotes. Paper: arXiv 2611.01860 · Discussion
2. Wave Forcing generates long videos continuously by letting the model use earlier chunks before their denoising is complete. It spreads staggered denoising and cache-refresh stages across full-model replicas to produce one output chunk per concurrent forward round after pipeline fill. The authors report that on Wan2.1-T2V-1.3B, 5-step WF keeps overall quality comparable to Rolling Forcing with higher Dynamic Degree on short and long videos, and that 4-step WF achieves 147.6 DiT FPS. 127 upvotes. Paper: arXiv 2611.00859 · Discussion · Code
3. The authors propose the cognitive symbiont, an AI companion built to augment one person's thinking. They describe a lifelong, embedded, real-time, private, low-friction and proactive companion that senses what the person senses through external devices and keeps their context across years. The authors call the paradigm AI-augmented human intelligence (AHI) and centre it on the cognitive gap between what a person could bring to a situation and what they can hold in working memory. 119 upvotes. Paper: arXiv 2611.05376 · Discussion
4. VFlash predicts a robot's whole visual plan, such as a box or a path, in one step instead of one coordinate at a time. It uses a lightweight drafter that generates an entire primitive in one forward pass. The authors report lower prediction errors than autoregressive decoding across the evaluated 2D and 3D robot trajectory tasks, and approximately 25x faster generation for 3D point motion forecasting in preliminary single-rollout measurements. A world action model conditioned on VFlash plans reaches 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the baseline. 107 upvotes. Paper: arXiv 2611.08449 · Discussion
5. CoworkOS is a harness for long-horizon work that separates the model's decisions from the runtime that keeps track of state. The model decides actions, delegation and acceptance, while the runtime maintains state and feedback, and humans and agents can edit the same documents, presentations, CAD models and Blender scenes. The authors report it is highly competitive with Codex and DeepSeek Harness across 12 benchmarks and 6 categories of real-world applications. 100 upvotes. Paper: arXiv 2611.03178 · Discussion · Code
6. SLA is a robot-control model that adds three-dimensional spatial alignment during continual pre-training. Starting from a pretrained pi05 policy, it jointly optimizes action prediction and alignment between intermediate visual embeddings and spatial representations from a frozen 3D perception model. The authors curate and unify heterogeneous robot datasets, add feature-dependent gates, and use all auxiliary components only during training. 93 upvotes. Paper: arXiv 2611.08373 · Discussion
7. Route Anything studies how an AI system can pick the right model or policy for each task while balancing performance and cost. The authors develop a theoretical foundation for optimal routing under quality-cost trade-offs and build an omni-modal trajectory matrix spanning 25 benchmarks, with 374,401 scored runs from 452 candidate configurations. They then train and test different routers on it. 80 upvotes. Paper: arXiv 2611.05275 · Discussion · Code
8. The authors survey how agent harnesses are used around video generation models. They define a harness as an executable runtime that constructs context, exposes tools, maintains state, routes feedback and manages budgets around video models. The report examines three directions: agentic video generation, agentic world modeling, and recursive self-improvement in video model development. 78 upvotes. Paper: arXiv 2610.40191 · Discussion · Code
9. Pandora simulates psychotherapy clients whose wording follows an individual's own language patterns. It updates psychological and relational states in response to therapist interventions, and a separate expression planner uses language patterns distilled from the client's utterances. The authors report the highest expression similarity and overall disclosure-alignment score among the compared systems on 115 cases, and in a four-evaluator study of 45 matched disclosure items, AI-like judgments in 7.2% of ratings against between 43.3% and 67.2% for the baselines. 77 upvotes. Paper: arXiv 2611.00617 · Discussion · Code
10. KlingAI Research reports what it learned about multimodal understanding while building successive generations of Kling. The authors argue that generation quality is the projection of multimodal understanding through the data, conditioning, alignment and evaluation chain, and examine the chain module by module. Kling-MLLM, a shared multimodal foundation model, unifies image, video, audio, synchronized audio-video and interleaved inputs. 75 upvotes. Paper: arXiv 2611.04916 · Discussion

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