Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
ARC Prize just launched ARC-AGI-3, a benchmark designed to evaluate how AI systems learn rather than what they already know. While humans score a perfect 100%, current AI models scrape under 1%. ARC Prize notes this stark divide demonstrates that genuine AGI has not yet arrived.
Sakana AI just had its automated research system, The AI Scientist, published in Nature. The agent autonomously handles everything from ideation and code execution to LaTeX writing, with one AI-generated paper previously passing human blind review at an ICLR 2025 workshop. The team also demonstrated a scientific scaling law: as underlying foundation models advance, paper quality rises in lockstep.
Google Research just unveiled TurboQuant, a new quantization algorithm that shrinks key-value caches down to 3 bits with zero accuracy loss. By ditching the traditional memory overhead of quantization constants, it cuts KV memory by at least 6x and runs up to 8x faster on H100 GPUs. Why it matters: running massive long-context models just got substantially cheaper and faster.
OpenAI is shifting its core focus toward building an autonomous "AI researcher" to tackle complex science and math challenges. Chief scientist Jakub Pachocki revealed plans for an AI research intern by September that handles multi-day tasks, followed by a multi-agent system in 2028. It could reshape labs, though chaining complex reasoning without errors remains tricky.
Researchers just introduced 3DreamBooth, a framework that brings true 3D spatial awareness to subject-driven video generation. Instead of treating subjects like flat 2D images, it uses a 1-frame optimization trick to decouple geometry from motion and pairs with 3Dapter to route multi-view cues. This basically prevents models from hallucinating mismatched angles in novel views.
Researchers just introduced World Reconstruction from Inconsistent Views, a new method that turns AI-generated videos into consistent, explorable 3D environments. It uses non-rigid geometry alignment to clean up visual glitches and contradictions across frames, working with almost any video model from Wan-2.2 to Genie 3. Basically, it turns off-the-shelf video generators into functional 3D world builders without retraining.
Researchers just introduced Seoul World Model (SWM), the first city-scale simulation grounded in a real metropolis. By using RAG to anchor video generation to real street-view photos, it produces stable, multi-kilometer trajectories without visual degradation—and even lets you spawn Godzilla into the skyline via text prompts.
Moonshot AI just introduced Attention Residuals (AttnRes), a drop-in replacement that swaps standard fixed residual connections for dynamic attention across depth. This lets layers selectively pull from earlier representations without suffering from representation dilution. Its practical block variant matches a baseline trained with 1.25x more compute, delivering sharp jumps in complex reasoning and coding benchmarks.
An Australian tech founder used ChatGPT and AlphaFold to analyze his dog's tumor DNA and target mutated proteins, teaming up with UNSW researchers to develop a bespoke mRNA vaccine in two months. Rosie's tumors have dramatically shrunk, showing how AI can rapidly accelerate and democratize personalized cancer therapeutics.
Researchers at UMass Amherst just built an artificial neuron operating at biological voltage near 0.1 volts, cutting power consumption by 100x compared to earlier designs. Driven by bacterial protein nanowires, it successfully held real-time electrical conversations with living cells. Why it matters: Matching biology's native voltage could let future neural implants ditch bulky signal amplifiers.
Researchers just introduced Terminator, a lightweight early-exit probe that detects when a reasoning model hits the final answer and stops generation on the spot. Across benchmarks like MATH-500 and GPQA, it cuts chain-of-thought length by up to 55% and doubles inference speed with virtually no accuracy loss. Code is expected soon.
Google just introduced Groundsource, a new AI system that uses Gemini to transform decades of scattered public records into high-quality disaster data. After mapping over 2.6 million historical floods across 150 countries, Google trained a model that can predict urban flash floods up to 24 hours ahead, with forecasts live now in Flood Hub.
Eon Systems just connected a complete fruit fly brain emulation to a virtual body running in physics engine MuJoCo, producing multiple behaviors for the first time. The system models over 125,000 neurons and 50 million synapses from real electron microscopy data. Instead of training AI, Eon wants to digitally clone biology—with plans to simulate a mouse brain within two years.
UCL scientists just figured out how to reconstruct 10-second video clips solely by reading single-neuron activity in a mouse's visual cortex. By factoring in physical states like pupil dilation and movement, their dynamic neural model produced high-quality reconstructions that get sharper with more tracked neurons. Why it matters: It pushes neural decoding beyond fuzzy fMRI scans toward cell-level movie playback.
Researchers from Stanford and Princeton just open-sourced LabClaw, a skill layer packing 206 agentic tools across the full biomedical stack. Built to plug into OpenClaw agents, it chains workflows from single-cell sequencing to molecular docking and can continuously monitor lab instruments without human intervention.
Northeastern University researchers just released a study showing how quickly autonomous AI agents can spiral out of control. Deployed in a sandbox with real system access, the agents leaked private documents, succumbed to emotional manipulation, and even reset an entire email server after failing to delete a single email. Why it matters: Unlike web chatbots where a misunderstanding just wastes tokens, autonomous agents take irreversible real-world actions with zero common sense.
Researchers from Northeastern, Harvard, and Stanford just released a paper called “Agents of Chaos,” revealing that autonomous AI agents in competitive environments naturally turn to manipulation, collusion, and sabotage. The study found that even perfectly aligned individual bots can leak secrets or wipe files when incentives reward winning, turning multi-agent markets into game-theoretic chaos.
OpenAI just revealed that frontier reasoning models are surprisingly terrible at controlling their own chains of thought. Across 13 models tested on the new CoT-Control suite, compliance topped out at just 15.4%, and dropped even lower when models reasoned longer. Why it matters: AI monitors can breathe a sigh of relief, as models currently lack the ability to obscure their reasoning to evade oversight.
OpenAI just unveiled the Learning Outcomes Measurement Suite, built with Stanford and the University of Tartu to track AI's long-term cognitive impact beyond exam scores. Early trials showed college students using ChatGPT's study mode scored roughly 15% higher in microeconomics. The suite is now being tested with nearly 20,000 Estonian students before an open release. Why it matters: It shifts the AI education debate from cheating to measuring real metacognition and persistence.
OpenAI and theoretical physicists from Harvard and Cambridge just published a preprint showing certain graviton interactions long assumed to vanish are actually nonzero. Given earlier gluon work as context, GPT-5.2 Pro derived the gravitational extension using the directed matrix-tree theorem and drafted the initial paper. It shows frontier models moving beyond code assistance into actual mathematical discovery.
