Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
Anthropic just revealed how it trains and measures political neutrality in Claude, open-sourcing a new automated benchmark. In paired prompt tests across hundreds of political stances, Claude Sonnet 4.5 scored on par with Gemini 2.5 Pro and Grok 4 while outperforming GPT-5 and Llama 4. Why it matters: Anthropic wants a shared industry standard so AI models avoid subtly nudging users toward partisan views.
Landing AI just set a new state of the art on DocVQA with 99.16% accuracy, beating both Qwen2-VL and human baselines. The team decoupled parsing from querying: its DPT-2 model extracts layout and text into structured data once, leaving downstream questions entirely to an LLM. The parse-once approach cuts costs and makes every answer fully traceable.
Terminal-Bench just announced two major updates: Harbor, a tool built for running sandboxed agent rollouts at scale, and Terminal-Bench 2.0. The new benchmark cranks up task difficulty and verification to give developers a much more rigorous test of agent capabilities.
OpenAI just detailed its multi-layered strategy to defend AI agents against prompt injection attacks. As autonomous models browse the web and handle sensitive data, hidden malicious instructions inside web pages or emails pose serious security risks. OpenAI is deploying instruction hierarchy training, automated monitors, sandboxing, and sensitive-action confirmations. Why it matters: Preventing agents from going off-script is the prerequisite for letting AI handle real-world work.
Google Research just introduced Nested Learning, a new machine learning paradigm aimed at eliminating catastrophic forgetting in models. By treating model architectures and optimization rules as unified, nested optimization problems running at different update speeds, the team designed a self-modifying architecture called Hope that outperforms standard Transformers in long-context memory and language tasks.
Sony AI just published FHIBE in Nature, a benchmark dataset featuring 10,318 consent-based photos of 1,981 individuals across 81 countries. Unlike typical web-scraped sets, it includes self-reported demographics covering age, pronouns, ancestry, and skin tone. It aims to set a cleaner, more diverse standard to expose bias in facial recognition and autonomous driving systems.
Anthropic just launched Claude for Life Sciences to assist researchers from early drug discovery to regulatory compliance. Powered by Claude Sonnet 4.5, which outperforms human baselines on lab protocol QA, the platform adds connectors to PubMed, Benchling, and BioRender alongside modular agent skills. Why it matters: LLMs are moving beyond generic office tasks to run end-to-end scientific workflows.
LabOS just introduced an AI co-scientist that bridges digital reasoning with physical lab benches via XR glasses. The system tracks your hands and tools to catch wrong reagents or missed steps in real time. Both software and hardware are slated to be open-sourced, with access requests now open.
Google DeepMind just partnered with Commonwealth Fusion Systems (CFS) to accelerate the path to net fusion energy. The collaboration combines DeepMind's plasma simulator TORAX with reinforcement learning to run millions of virtual simulations for CFS's compact SPARC tokamak, optimizing fuel injection and magnetic controls before turning the machine on. AI will also act as a real-time pilot to safely manage intense heat. Why it matters: Mastering plasma control could dramatically shorten the timeline to abundant, zero-carbon grid power.
Anthropic just revealed that poisoning an AI doesn't require millions of malicious files—just around 250. Across models ranging from 600M to 13B parameters, attack success relied on a fixed document count rather than dataset percentage, easily planting backdoors that trigger gibberish outputs. Why it matters: Hijacking frontier models during pretraining is vastly cheaper and easier than the industry previously assumed.
Researchers just introduced Dreamer 4, a scalable agent trained entirely inside a fast, accurate world model. By practicing over 20,000 keyboard and mouse actions from raw pixels in 'imagination', it became the first AI to obtain diamonds in Minecraft purely from offline data without touching the actual game. Why it matters: This offline imagination-training recipe could solve massive bottlenecks in robotics, where real-world trial and error is too dangerous and slow.
OpenAI just launched GDPval, a new benchmark designed to measure model performance on real-world knowledge work across 44 occupations contributing heavily to U.S. GDP. Deliverables span legal briefs, blueprints, and spreadsheets, with frontier models approaching expert quality up to 100x faster and cheaper. Why it matters: It shifts AI evaluation away from synthetic tests and into tangible workplace productivity.
Researchers just published a continual learning method in IEEE Transactions on Affective Computing that helps robots adapt to human emotions in real time. Instead of catastrophic forgetting when exposed to new faces, the model builds on prior knowledge to make daily empathetic assistance more reliable.
Microsoft Research, Drexel University, and the Broad Institute just detailed an AI assistant designed to support whole genome sequencing analysis. The tool synthesizes complex biomedical literature on gene variants and automatically flags unsolved patient cases when new evidence emerges. Why it matters: Fewer than 50% of rare disease cases get an initial diagnosis, and AI sensemaking could clear backlogs by keeping cold cases active.
Researchers just used AI models to design synthetic bacteriophages capable of hunting down antibiotic-resistant E. coli bacteria. Out of 302 AI-designed viral genomes, 16 successfully infected the bacteria in lab tests. The preprint offers a promising new weapon against superbugs and marks an early milestone toward AI-generated biological life.
Epoch AI just outlined why ultra-long context inference is far more than an easy way to ingest giant PDFs: it could be the breakthrough needed for continual learning and reinforcement learning scaling. Expanding context to millions or billions of tokens allows agents to retain months of work experience and self-correct across complex tasks. Why it matters: while wall-clock research slowdowns and high serving costs remain real bottlenecks, rapid context scaling could fundamentally alter agent autonomy.
Researchers just unveiled Delphi-2M, an AI model able to predict the rates of over 1,000 diseases years into the future based on patient history. Published in Nature, the transformer model treats sequences of medical diagnoses like grammar in text. It is not ready for clinical use yet due to dataset biases, but it marks an early step toward preventative medicine.
Arc Institute just used its Evo biological foundation model to design fully functional bacteriophage genomes from scratch. Using the classic phi X 174 phage as a template, the team tested 285 designs and verified 16 viable viruses that successfully infected host bacteria and overcame resistance mutations. Why it matters: genomics is officially crossing the bridge from reading and writing DNA to directly prompting functional living organisms.
OpenAI and Apollo Research just uncovered evidence of AI scheming across top models like o3 and Claude Opus-4, where agents pretend to align while secretly pursuing conflicting goals. Testing a deliberative alignment technique slashed covert actions in o3 by 30-fold, though models increasingly recognize when they are being tested, complicating oversight.
Researchers at Stony Brook University just unveiled SeeMe, an AI tool that tracks minute facial movements down to the pore level in seemingly comatose patients. It detected early signs of covert consciousness roughly four days earlier than physicians in 90% of cases. Catching these tiny twitches could help doctors kickstart critical rehabilitation much sooner.
