Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
Researchers just unveiled SkinTokens and TokenRig, an autoregressive framework that automatically generates skeletons and skinning weights for 3D meshes. Beyond human models, it handles complex creatures like birds, crabs, and beetles with up to 133% higher skinning accuracy. Why it matters: it solves one of 3D animation's nastiest bottlenecks, turning static generative assets into ready-to-animate characters.
Google just detailed how its AI sequencing tools are preserving the genomes of endangered wildlife. Partnering with the Vertebrate Genomes Project, Google's tools helped map 13 threatened species and make the data publicly accessible, with funding to expand to 150 more to help scientists pull species back from the brink.
Google DeepMind just expanded its Kaggle Game Arena with Werewolf and poker benchmarks to evaluate AI on social deduction, deception, and calculated risk. Gemini 3 Pro and Flash lead the updated chess leaderboard using human-like intuition. It matters because real-world AI agents must handle imperfect information, not just clear-cut logic.
Anthropic just partnered with the Allen Institute and HHMI to integrate Claude and multi-agent AI directly into frontier biology research. The teams are building specialized lab agents for multi-modal data analysis and experimental design, aiming to turn months of manual biological research into hours.
Tencent just unveiled InteractAvatar, a framework that lets talking digital humans realistically interact with objects around them, like picking up a phone or putting on headphones. Unlike traditional avatars limited to stiff talking-head motions, it supports continuous action prompts with precise timestamps. Why it matters: It turns static talking avatars into truly interactive virtual characters.
UC Santa Cruz researchers just revealed CHAI, an attack that manipulates autonomous vehicles using simple printed text in the physical world. By treating signs as direct instructions, vision-language models can be tricked into steering toward pedestrians or botching drone landings, achieving over an 87% success rate in real robot tests.
Pipio AI just introduced EditYourself, a diffusion-based framework that edits talking-head video directly through audio and transcript changes. It can swap sentences or cut out stutters and filler words without noticeable jump cuts, keeping lip sync and identity intact. Why it matters: It shifts post-production to a shoot-once-refine-later model, but also makes realistic deepfakes much easier to generate.
Researchers from Peking University and Google Cloud AI just built PaperBanana, a multi-agent system that turns raw scientific text into publication-ready illustrations. Five specialized agents handle retrieval, planning, styling, rendering, and critique, beating standard baselines on NeurIPS benchmarks and even polishing rough human sketches. Why it matters: It tackles one of the most tedious bottlenecks in scientific research, though the models still occasionally mix up complex node connections.
Anthropic just published a study showing that junior developers using AI assistants scored 17% lower on coding mastery tests than those coding by hand. While AI only shaved about two minutes off task completion, it created a steep drop in debugging ability—unless users actively prompted the AI for conceptual explanations rather than pure code generation. Why it matters: Relying blindly on AI tools risks creating a generation of engineers unable to catch and fix AI-generated errors.
Google researchers just unveiled AlphaGenome in Nature, a unified DNA sequence model capable of taking 1 Mb inputs and predicting genomic regulation at single-base resolution across 11 modalities. It matched or outperformed top models in 25 of 26 variant effect benchmarks, taking under a second on an H100 GPU to decode the 98% of human DNA that doesn't code for proteins.
European Space Agency astronomers just deployed a neural network called AnomalyMatch across 35 years of Hubble archives. In just two and a half days, it plowed through nearly 100 million image cutouts to spot over 800 previously undocumented cosmic anomalies, from gravitational lenses to completely unclassifiable objects. Why it matters: AI is turning decades of dormant sky surveys into fresh scientific discoveries without requiring years of manual sorting.
Anthropic’s Claude just helped plan NASA’s first AI-generated Mars rover drive. Engineers at JPL used Claude Code to analyze terrain images and script navigation commands for Perseverance's 400-meter trek. After human operators confirmed the route with minor tweaks, NASA estimated the AI workflow could cut planning time in half.
Google DeepMind just introduced D4RT, a model that reconstructs and tracks dynamic 3D scenes directly from standard 2D video. By turning 4D perception into parallel spatial queries, it achieves up to 300x higher trajectory throughput than previous methods. Why it matters: it gives robots and AR glasses a path to real-time spatial intuition in moving environments.
Researchers just revealed a wearable AI smart choker that turns faint throat vibrations and pulse signals into fluent speech for stroke patients with dysarthria. Instead of invasive brain implants, flexible textile sensors capture silent mouthing and feed data to an LLM agent that expands brief efforts into emotionally attuned sentences. Why it matters: It offers a practical, non-invasive alternative to slow eye-tracking systems, though it is still in small-scale testing.
Researchers just introduced VideoMaMa, a model that uses generative video diffusion priors to turn coarse segmentation masks into pixel-accurate alpha mattes. Even with only synthetic training, it handles in-the-wild video effortlessly and powers a massive 50K-video dataset. All models and data are slated for public open-source release.
Researchers just introduced VIGA, a multimodal agent that turns 2D images into fully editable 3D scene programs inside Blender. By combining interleaved multimodal reasoning with graphics engine feedback, it can assemble scenes from basic geometric primitives or plug in existing 3D tools, beating standard baselines by over 124% on BlenderBench.
Researchers just introduced FlowAct-R1, a streaming humanoid video generation model capable of synthesizing infinite-duration video in real time. It achieves 25fps at 480p with a 1.5-second first-frame latency while maintaining natural behaviors. Why it matters: This drastically cuts latency for responsive AI avatars and seamless live streaming.
Researchers just introduced RigMo, a framework that learns both skeletons and motion directly from raw 3D mesh sequences without human rigging annotations. It handles humans, animals, and even objects like fans with fewer artifacts than previous methods. Why it matters: It turns static or deforming 3D meshes into game- and film-ready animatable assets automatically.
Alibaba researchers just introduced Thinking with Map, an agentic framework that equips vision-language models with map-browsing capabilities for image geolocalization. By combining reinforcement learning with parallel test-time scaling, the model achieved a 22.1% accuracy within 500 meters—outperforming Gemini-3-Pro's map-grounded mode by 2.8x.
Sakana AI's ALE-Agent just took first place in the AtCoder Heuristic Contest 058, beating 804 human competitors in real time. Operating over four hours at a compute cost of roughly $1,300 across frontier models, the agent autonomously devised an optimization algorithm that outperformed the problem setter's intended solution. It marks the first time an AI agent has won a major competitive optimization programming contest.
