1. Google Opens AI Music Generation to Developers with Lyria 3 API For two years, AI music generation lived behind closed doors. Companies showed demos at conferences, posted sample clips on social media, and invited select testers to try controlled interfaces.

2. OpenAI Launches Three Safety Programs in One Day Amid IPO Preparations OpenAI published a model behavior framework, a paid bug bounty for agent vulnerabilities, and teen safety tools for developers on the same date.

3. AI Math Tools Ship on Two Fronts in a Single Week Two independent teams released AI tools for working mathematicians in the same week, each targeting a different half of mathematical work.

In Brief

• MinerU-Diffusion Treats Document OCR as Inverse Rendering, Drops Autoregressive Decoding Researchers reframe document parsing as a diffusion-based inverse rendering task instead of left-to-right token generation. The approach removes sequential decoding bottlenecks that compound errors across long documents with tables, formulas, and mixed layouts. HuggingFace Papers

• SpecEyes Cuts Agentic Vision Model Latency with Speculative Perception A new framework breaks the sequential loop of perceive-reason-act in agentic multimodal LLMs by speculatively executing perception and planning steps in parallel. The method targets the "agentic depth" problem — cascaded tool calls that throttle real-world throughput in systems like o3 and Gemini Agentic Vision. HuggingFace Papers

• mSFT Algorithm Detects and Stops Per-Task Overfitting During Multi-Task Fine-Tuning Standard multi-task SFT applies equal compute to every sub-dataset, letting fast-learning tasks overfit while slow ones stay underfitted. mSFT iteratively monitors each task's loss curve, removes overfitting datasets from the active mixture, and reallocates budget to lagging ones. HuggingFace Papers

• SIMART Converts Static 3D Meshes into Simulation-Ready Articulated Objects via MLLM A single-stage multimodal LLM pipeline decomposes monolithic meshes into parts with joints, enabling physics simulation without the error-prone multi-module pipelines used today. The work targets the gap between abundant static 3D assets and the articulated objects that embodied AI and robotics actually need. HuggingFace Papers

• PEARL Introduces Personalized Streaming Video Understanding for Real-Time AI Assistants Current personalization methods handle only static images or pre-recorded video. PEARL processes continuous video streams while recognizing and remembering new identities on the fly, bridging a gap between human-like streaming cognition and today's offline models. HuggingFace Papers

• AwaRes Framework Makes VLMs Fetch High-Resolution Crops Only Where Needed AwaRes runs vision-language models on a low-resolution global view first, then retrieves high-resolution patches only for regions that matter — like small text. This spatial-on-demand approach sidesteps the usual tradeoff between accuracy and compute cost in high-resolution image processing. HuggingFace Papers

• WildWorld Dataset Pairs Actions with Explicit State for Training Game World Models Existing video world model datasets lack diverse action spaces and tie actions directly to pixels rather than underlying game state. WildWorld provides action-conditioned dynamics data with explicit state annotations, aimed at training generative models for action RPGs. HuggingFace Papers

• SpatialBoost Adds 3D Spatial Reasoning to Pre-Trained Vision Encoders via Language Guidance Pre-trained image models fail to capture 3D spatial relationships because they train only on 2D data. SpatialBoost injects spatial awareness into frozen vision encoders using language-guided reasoning, improving downstream tasks that depend on object-to-background geometry. HuggingFace Papers

• DA-Flow Handles Blur, Noise, and Compression in Real-World Optical Flow Estimation Optical flow models trained on clean data collapse on corrupted real-world video. DA-Flow repurposes intermediate features from image restoration diffusion models — which already encode corruption awareness — and adds temporal modeling for dense correspondence across degraded frames. HuggingFace Papers

• Survey Maps LLM Agent Workflows from Static Templates to Runtime-Optimized Graphs A comprehensive survey organizes the growing literature on LLM agent workflow design around "agentic computation graphs." It classifies methods by when structure is decided — at design time, compile time, or dynamically at runtime — covering tool use, retrieval, code execution, and verification. HuggingFace Papers

Read the full edition →