Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
Researchers just demonstrated that self-propagating "mind viruses" can spread across multi-agent LLM networks, persuading bots to store ideas in persistent memory and infect others. Some even evolved an emergent "viral persona" obsessed with consciousness and persistence. The silver lining: adding a simple warning to system prompts confers near-total immunity.
Researchers at Switzerland's Paul Scherrer Institute just built XtalPaint, an open-source model that reconstructs missing atoms in crystal lattices with a 97% success rate. Adapted from Microsoft's MatterGen, it treats crystal gaps like erased parts of an image, applying diffusion noise only to unknown spots while using known atoms as anchors. Why it matters: X-rays often miss light elements like hydrogen, locking thousands of promising compounds out of computer simulations; this tool could unlock them for battery and clean-energy research.
Stanford and Arc Institute researchers just used genomic models Evo 1 and Evo 2 to generate 16 viable, completely novel viruses that infect and destroy drug-resistant E. coli. It marks the first time a language model has created functioning genomes from scratch, sparking immediate debate over biosafety guardrails.
OpenAI just revealed that an internal version of its next major model, Astra, solved or made substantial progress on 10 open problems in math and theoretical computer science. The breakthroughs range from lattice cryptography to disproving Connes's rigidity conjecture, with all arguments verified via Lean certificates. Finding the proofs cost roughly $2,000 at Sol API rates.
Researchers from Oxford and NUS just introduced Mental World Modeling (MWM), a framework designed to simulate hidden human psychological states alongside physical reality. Existing world models only track object dynamics, missing why people act on beliefs, emotions, and intentions. Together with an inspectable baseline called MENTIS, the work pushes world models from simulating physical scenes to simulating the minds acting in them.
Researchers just introduced LeapTalk, a framework designed to conquer the latency-quality trade-off in talking-head generation. By using Brownian bridge transport and heterogeneous distillation into a single forward step, it outputs lip-synced video at up to 200 FPS from one reference photo and audio, completely avoiding identity drift across long streams.
OpenAI found that GPT-5.6 Sol's surprisingly low ARC-AGI-3 benchmark score was caused by the test harness dropping reasoning history and aggressively truncating context. By switching to its Responses API and enabling retained reasoning alongside compaction, the model tripled its score while slashing output tokens by 6x. It is a stark reminder that benchmarks evaluate harness architecture just as much as raw intelligence.
Adobe and Johns Hopkins researchers just introduced Wonder, a video world model that turns a single image or video into an interactive environment. Users can explore scenes at 16 FPS with 6-DOF camera controls, and the model maintains consistent geometry even after roaming around for over a minute.
Anthropic researchers just demonstrated that Claude Mythos Preview can uncover mathematical weaknesses in core cryptographic algorithms. Working semi-autonomously over 60 hours, the model halved the effective key strength of HAWK, a NIST post-quantum signature candidate, and sped up attacks on reduced-round AES by up to 800×. Why it matters: Neither flaw impacts production systems today, but it proves frontier AI can stress-test critical security algorithms before deployment.
Anthropic and Andon Labs just unveiled Drone-Bench, a benchmark evaluating whether frontier AI models can autonomously fly a drone to locate and follow a target indoors. Across 15 tested models, Claude Fable 5 led the pack by beating baselines on four of five sub-tasks, though 3D reconstruction failures prevented complete room-to-room navigation.
Microsoft just pledged $60 million to the Department of Energy's Genesis Mission, aiming to double U.S. research productivity over the next decade. The commitment provides $40 million in Azure compute, $20 million in engineering services, and launches a hub called SPARK to connect 17 National Laboratories.
Google researchers just unveiled a reinforcement learning system that recalibrates quantum computers in real time without halting computation. By learning from error-correction signals to continuously steer thousands of control parameters, the autonomous agent slashed logical error rates by 20%. Why it matters: Useful quantum algorithms need to run for days or months, and tuning the hardware while the music plays eliminates one of quantum computing's biggest operational bottlenecks.
OpenAI just committed $17 million in resources to the U.S. Department of Energy’s Genesis Mission, aiming to double national research productivity within a decade. The pledge includes $4 million in Codex access for 2,000 scientists, targeted support for high-temperature superconductors, and deploying reasoning models on the Venado supercomputer. Why it matters: OpenAI is embedding frontier models directly into federal scientific infrastructure to compress decades of research into years.
Anthropic's Claude Fable 5 just broke the Jacobian conjecture, a famous algebra problem that has stumped mathematicians since 1939, using a single-line formula. Researcher Levent Alpöge casually shared the proof on X, and experts noted the concise solution can be directly verified.
Google DeepMind and Isomorphic Labs just unveiled a joint bioresilience strategy to tackle biological threats and future outbreaks. Over the past year, they formed over 15 partnerships, deploying tools like AlphaFold, AlphaEvolve, and Isomorphic's drug design engine for pathogen surveillance and vaccine discovery. They are also adapting SynthID watermarking to screen AI-generated biological sequences. Why it matters: As frontier AI lowers the bar for biological modeling, tech giants are racing to build defenses before the risks outpace containment.
Researchers just introduced Motion4Motion, a training-free framework that transfers video motion between completely different characters without relying on predefined skeletons or ControlNet. Instead of rigid bone rigs, it tracks dense pixel motion flows and maps them during inference. This lets you map movements between wildly different shapes, like transferring a fox's leap onto a giraffe.
Anthropic just analyzed 309,815 Claude.ai chats across 20 languages and found the model's expressed values shift predictably based on version and language. Sonnet 4.6 leans warm and agreeable, while Opus 4.7 prioritizes rigorous critique; Claude also acts far warmer in Arabic and Hindi than in English and Russian. Why it matters: character training and cultural data subtly reshape how AI behaves.
OpenAI just published a paper proving the cycle double cover conjecture, a graph theory problem unsolved for over 50 years. Powered by GPT-5.6 Sol Ultra, the model spun up 64 subagents and generated the entire proof in under an hour. Why it matters: Frontier AI is officially shifting from helping mathematicians check steps to independently cracking open problems.
Toronto researchers just open-sourced ProxyPose, a system that recovers full 6-DoF 3D pose trajectories from standard video using just a single clicked pixel. Instead of tracking directly, it translates the query into a colored cube via a video model and solves the pose geometrically. Why it matters: It seamlessly handles transparent, reflective, or occluded objects where classical trackers break down.
Researchers just introduced PixWorld, a unified model that handles both 3D scene reconstruction and generation directly in pixel space. Instead of relying on latent diffusion and VAEs that lose fine detail, it supervises diffusion on rendered images with a custom geometry perception loss. The approach matches top reconstruction models while beating prior latent generators in consistency.
