Identifies a mismatch between generative reward models and scalar scoring in existing RL, proposing Ranking-based Reward Construction to let generative reward models provide effective rewards for LLM policy optimization.
Argues chunk-embed-top-k retrieval is structurally unsound for table-dense documents — on a 780-page financial report, 86.8% of content lines were table rows — and replaces it with interpretable agentic retrieval operations.
Introduces trace-grounded parametric profiling for event counting across 2,190 controlled video tasks, decoupling event count, rate, and duration to show video LMs systematically fail at basic temporal reasoning.
A real-world-grounded diagnostic benchmark that jointly evaluates numerical physics simulators and generative video world models on physical fidelity, identifying which physical principles or parameters are violated.
Shows stale tool traces hijack agent policies — on Qwen3-1.7B, history pollution flips 32.1% of decisions that were correct under the original trajectory — and introduces a paired benchmark with synchronized Original, Polluted, and Oracle State views.
Black Forest Labs' FLUX 3 generates video with synchronized audio up to 20 seconds, supports multiple scenes, camera angles, multilingual dialogue with lipsync, keyframe control, and intelligent video continuation. Available via Together Serverless Inference.
Actual is optimized for personal compute with low CPU overhead during inference and multi-platform portability, positioned as a companion for local agent setups like Hermes.
New research introduces Skill Entropy, a metric for benchmarking and training long-horizon reasoning in LLMs, published on Hugging Face papers.
Research paper exploring knowledge flow, modality synergy, early unification, and training recipes for multimodal pretraining.
AMD agreed to acquire Taalas, a Toronto startup that builds chips with AI models hardwired into silicon, pushing deeper into the AI inference market against NVIDIA's dominance.
Chinese humanoid robot maker Unitree priced its Shanghai STAR Market IPO at 150.8 yuan per share, valuing the company at approximately $9 billion — the only humanoid robot manufacturer turning a profit while rivals remain pre-revenue.
NVIDIA published Cosmos 3 open world models under the Linux Foundation's OpenMDW 1.1 license, designed for generating training data and testing policies for robotics and autonomous vehicles.
The US Navy's largest shipbuilder signed performance-based deals worth up to $900M over seven years for robotic welding and sanding, deploying physical AI across carrier, submarine, and destroyer programs.
Google Maps launched AI-powered agentic features that let users complete real-world tasks directly from the navigation app, transforming it from a navigation tool into a task-completion assistant.
A shared memory layer that works across AI coding tools, letting agents retain context between sessions and reducing repeated instructions.
A free tool that enables side-by-side comparison of frontier speech-to-speech AI models, helping developers choose the right voice model for their use case.
Seedance 2.5 video generation model is now available as an API on Atlas Cloud, supporting 30-second single-take clips, 50 reference images, and 4K output resolution.
A security audit tool scanned 30 production MCP servers for OAuth implementation quality, finding widespread security gaps with only one server achieving an A-grade.
Omilia's agentic CX platform grew ARR 10x to $60M since Series A. Clients include Capital One, RBC, Discover, and Taco Bell. Led by Expedition Growth Capital.
Naive builds autonomous company infrastructure — AI agents that handle end-to-end business setup. Led by Nexus Venture Partners with 30,000 registered developers on the platform.
Applies Spotify-style session-based recommendation to online retail using a two-headed Vector AI. Co-led by Bessemer Venture Partners and Gradient Ventures.
Source:
