Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
MoVerse just introduced a real-time video world model that turns a single regular image into an interactively navigable 360° world. By decoupling 3D Gaussian world memory from video rendering, it pulls off 8 FPS interactive roaming on a single consumer RTX 4090 GPU. The code and weights are slated for release following compliance review.
Revolution Medicines just showed that a drug targeting the slippery RAS protein nearly doubled survival time in advanced pancreatic cancer patients, jumping from 6.6 to 13.2 months. Now, researchers are deploying AI to screen millions of molecules against notorious, smooth-surfaced culprits like p53 and MYC that fuel roughly 70 percent of cancers. What used to take decades of chemistry trial-and-error might soon become a solved computation problem.
Researchers just introduced Arbor, an autonomous agent framework that turns messy trial-and-error into a structured hypothesis tree. A long-lived coordinator oversees strategy while isolated executors run tests, achieving over 2.5x the relative gains of Claude Code and Codex across long-horizon research tasks.
UC Berkeley's RDI just launched Agents' Last Exam (ALE), a demanding benchmark testing AI on real desktop workflows across 55 industries. OpenAI's GPT-5.5 topped the board with a 24.0% pass rate, narrowly beating Anthropic's new Claude Fable 5 at 22.0%. Why it matters: even the best models scored 0.0% on the hardest tasks, proving AI agents still struggle with genuine professional workloads.
Researchers from KAIST AI and AIPARK just unveiled Lip Forcing, the first autoregressive diffusion model for video-to-video lip sync. By distilling a 14B teacher into a 2-step causal student, the 1.3B model hits 31 FPS on a single H100, easily clearing the real-time threshold.
Google DeepMind just released RCT results showing its Gemini tutor accelerated math learning by 1.2 to 2.5 years in just eight weeks across Sierra Leone schools. Instead of handing out solutions, Gemini acted as a Socratic guide, asking scaffolding questions 76% of the time and giving direct answers only 2%. Why it matters: It proves AI can foster critical thinking rather than student laziness, though teacher involvement remains critical.
Researchers just unveiled Anchor World, an embodied simulator controlled by actual 3D human motion instead of text prompts or key presses. By feeding in body movements alongside scene anchor views, it renders egocentric video of people navigating spaces and handling tools like sinks or knives. It could give humanoid robots a massive new source of first-person training data.
Anthropic just revealed research testing Claude against dedicated chemistry tools ChemDraw and MestReNova on NMR spectral analysis. Without domain-specific fine-tuning, Claude Opus 4.7 matched or beat both tools on hydrogen shift prediction with an average ±0.079 ppm error, while successfully performing inverse structure elucidation from raw 1D spectra. Why it matters: Frontier models are beginning to assist with complex analytical lab data, though benchmarks remain small.
Anthropic just shared unreleased internal data revealing that Claude now authors over 80% of the code merged into its repository, boosting engineer output eightfold since 2024. While systems are rapidly heading toward recursive self-improvement, the company warns the risk of losing control could arrive sooner than institutions are prepared for.
OpenAI just updated GPT-Rosalind, its specialized model series for enterprise life sciences research, baking in GPT-5.5's agentic coding and tool-calling skills. The new model beats GPT-5.5 on wet lab troubleshooting, medicinal chemistry, and genomics while burning up to 31% fewer tokens, aiming to slash the time scientists spend connecting dry-lab data to wet-lab experiments.
ByteDance just unveiled WavTTS, a zero-shot speech generation model built on Diffusion Transformers and flow matching. Instead of relying on compressed latent representations, it directly models raw audio waveforms to bypass information loss. Why it matters: it clones voices from just a few seconds of audio, proving that end-to-end waveform modeling can rival mainstream latent-space models.
Microsoft just unveiled its Majorana 2 scalable quantum processor, using an AI-designed material stack to make topological qubits over 1,000 times more reliable than before. Qubit lifetimes jumped from milliseconds to a mean of 20 seconds. Why it matters: The breakthrough cut Microsoft's roadmap in half, targeting a fault-tolerant quantum computer by 2029.
Stanford researchers just published a blind study showing law professors preferred AI-generated answers over responses written by fellow professors 75% of the time across nearly 3,000 contract law questions. AI models were flagged as pedagogically harmful only 3.5% of the time, compared to 12% for human peers. It proves AI is handling nuanced, subjective judgment far better than skeptics assumed.
Researchers just introduced AutoScientists, a decentralized multi-agent system designed for long-running scientific experimentation. Instead of relying on a rigid central planner, agents self-organize into teams, vet ideas before burning compute, and share failures to avoid redundant runs. It optimized GPT training 1.9x faster than Autoresearch while topping previous agents across biomedical benchmarks.
Biohub just unveiled ESM, a protein "world model" featuring the ESMFold2 design engine, a 2.8B-sequence model called ESMC, and an atlas of 1.1 billion predicted structures. Instead of spending months physically screening candidate binders in a lab, researchers can now design and rank functional antibodies computationally in just a few days. Lab tests against five cancer and immune targets, confirmed via cryo-EM, prove the designs actually work.
University of Rochester researchers just developed a solar-thermal desalination technique using laser-etched black metal that purifies seawater without producing toxic brine waste. The superwicking surface leverages the coffee-ring effect to push salts aside as reusable solids rather than dumping them back into the sea. It opens up an energy-efficient path to fresh drinking water while harvesting battery-grade lithium directly from ocean salts.
Nvidia just introduced LocateAnything, a visual grounding framework that lets you pinpoint and crop any object in an image using plain text descriptions. By decoding bounding boxes in parallel rather than token-by-token and training on a massive 138-million-sample dataset, it boosts dense object detection speeds by up to 10x while maintaining high localization accuracy.
Datacurve just released DeepSWE, a new benchmark tailored for agentic coding models. While frontier models often post nearly identical scores on existing leaderboards, DeepSWE simulates day-to-day developer workflows to expose where their actual capabilities diverge.
Google DeepMind just introduced AlphaProof Nexus, an AI agent that solved nine open Erdős math problems and 44 OEIS conjectures. The system pairs an LLM with the Lean proof assistant to automatically generate and verify proofs, running at a cost of just a few hundred dollars per problem. Why it matters: AI is moving past standard contest benchmarks into genuine mathematical discovery, though problems demanding brand-new mathematical constructions remain out of reach.
Researchers just introduced TriSplat, a feed-forward 3D reconstruction model that builds scenes with oriented triangle primitives rather than Gaussian splats. By generating surfaces directly in a single pass in about 1.2 seconds, it exports simulation-ready meshes straight into physics engines and Unreal without post-processing headaches.
