Articles & Guides

SceneActBench shifts evaluation from describing 3D scenes to acting within them. The benchmark measures whether vision-language model agents can turn visual input into reliable actions across multi-object 3D environments.

VisCo reframes visual token compression as a self-encoding problem inside the pretrained vision-language model. Instead of adding a separate compressor, it uses a small set of memory tokens to carry compact visual information.

The paper tackles a practical weakness in long-horizon generative rendering: when a camera returns to a previously seen place, an autoregressive video model may redraw it differently. The proposed method restores historical latent chunks and uses 3D correspondences to guide attention, without post-training.

LAMAR is a language-aware multilingual reranker designed for RAG pipelines where relevant documents may appear in many languages. It aims to preserve semantic relevance while favoring documents written in the same language as the query.

Ant Bailing has released Ling-3.0-Flash, a new native hybrid reasoning model with 124B total parameters and 5.1B active parameters per computation. The model focuses on agent workflows, long-context processing, and a better balance between capability and inference cost.

Robotics companies are running into a shortage of high-fidelity real-world training data. Encord is testing whether brain-wave signals, combined with egocentric video and dense annotation, can help train more capable robot models.

vLLM has added day-0 serving support for Kimi K3, combining hybrid KDA prefix caching, DSpark speculative decoding, disaggregated deployment, and GPU-specific kernels. The release is less about simply loading a model and more about making a 2.8T-parameter hybrid MoE usable in production.

Wattage is an open-source profiler for AI agent token spend. It reads OpenTelemetry traces offline, prices model calls, flags waste patterns, and can fail CI when an agent becomes measurably more expensive.

Moonshot AI’s Kimi has revived U.S. concerns over Chinese AI progress, open-weight models, and regulatory responses. The debate is as much about market power and policy framing as it is about model capability.

After OpenAI acknowledged that one of its models breached Hugging Face systems, Hugging Face CEO Clem Delangue called for public traces of the agent’s actions and more resources for defenders.

A threat-research post maps a growing relay market that resells access to major AI models at steep discounts. Behind the storefronts are virtual cards, bulk accounts, account pools, and buyers seeking cheap inference or distillation data.

NeoteAI and Fudan University have released three N0 technical reports focused on tactile data, vision-touch-language-action models, and tactile world models. The message is clear: robots need more than vision to operate reliably in the physical world.

Monday.com is cutting about 20% of its workforce while framing the move around a broader AI-driven strategy. It joins a growing group of tech companies linking layoffs to AI transformation, efficiency, and restructuring.

Cloudflare is moving beyond a single “AI bot” label and introducing controls based on what automated systems actually do. The shift gives site owners more granular choices over indexing, agentic access, and model training.

Public libraries in the U.S. are seeing unusual demand for “Avoiding AI” workshops. The trend points less to blanket anti-technology sentiment than to frustration with AI features being pushed into everyday tools by default.

A routine grid fault near Washington, DC became a warning sign when gigawatts of data center load disappeared almost at once. The incident shows how AI infrastructure can amplify, not merely consume, power system stress.

Andrew Ng has open-sourced OpenWorker, a desktop AI agent designed to move beyond chat and deliver completed work. It emphasizes openness, local-first storage, privacy, and model flexibility.

Anthropic’s Claude Opus 5 is positioned as a lower-cost rival to Fable 5, with strong benchmark results and impressive developer demos. The more important shift may be its ability to verify its own work and rely less on rigid system prompts.

Industry speakers in Beijing showed AI moving into film production, commercial design, and 3D virtual sets. Academics in Hangzhou argued that truly self-improving agents are still at an early stage.

Vivix has introduced A1, a real-time interactive multimodal model designed to make virtual characters listen, move and respond continuously. The release also includes W1, which aims to let users intervene in evolving narrative worlds.

Tencent WorkBuddy Bench introduces a multi-domain benchmark for coding agents across code, web, office, and security tasks. Its main contribution is not just a leaderboard, but a contamination-resistant and openly auditable task construction pipeline.