Skip to content
IFAI

Research

Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.

13 stories in the last 7 days

  1. Reach 31

    Sakana AI builds an AI agent to write and optimize CUDA kernels

    Sakana AI just unveiled an agentic framework and a new benchmark called robust-kbench to automate low-level CUDA kernel engineering. The system translates PyTorch code into raw CUDA kernels and uses an evolutionary meta-generation loop with LLM verifiers to iteratively boost runtime speed. Why it matters: Squeezing maximum performance out of GPUs is notoriously tedious, and handing kernel optimization over to AI agents could save developers months of manual tuning.

  2. Reach 35

    AI still can't reliably read an analog clock

    Alek Safar just introduced ClockBench to test multimodal models on analog clocks, and they struggled hard. While humans achieved 89.1% accuracy, the best model, Google's Gemini 2.5 Pro, only hit 13.3% with a median error of an hour. Interestingly, models solve time math easily once they get the time right, proving visual spatial reasoning remains the real bottleneck.

  3. Reach 32

    Harvard Medical School's new AI finds drug combos to cure sick cells

    Harvard Medical School just unveiled PDGrapher, a free graph neural network that spots gene targets and drug combos capable of reverting diseased cells back to a healthy state. Tested across 11 cancer types, it ranked therapeutic targets up to 35% higher and ran 25 times faster than existing models.

  4. Reach 35

    OpenAI reveals why language models can't stop hallucinating

    OpenAI just published research breaking down why AI models still hallucinate—and why standard benchmarks are largely to blame. Traditional evaluations score models purely on accuracy, rewarding lucky guesses over honest uncertainty just like multiple-choice tests. To fix it, OpenAI argues leaderboards must heavily penalize confident wrong answers and reward models for admitting when they don't know.

  5. Reach 31

    DeepMind exposes a hard mathematical ceiling in vector search

    Google DeepMind researchers just proved that single-vector embeddings have a fundamental geometric limit that breaks advanced RAG setups. On their new LIMIT stress test, top embedding models from Google and Snowflake saw recall drop below 20%, while old-school BM25 excelled. Why it matters: simply scaling models won't fix retrieval.

  6. Reach 31

    Tencent's R-Zero lets LLMs train themselves from scratch

    Tencent AI Lab and Washington University in St. Louis just unveiled R-Zero, a reinforcement learning framework that allows LLMs to self-evolve without any human-labeled data. By splitting a base model into competing Challenger and Solver roles, it autonomously generates progressively harder curriculum tasks. In benchmarks, it boosted Qwen3-4B's math reasoning score by 6.49 points. Why it matters: It could eliminate the massive data annotation bottleneck for reasoning models.

  7. Reach 31

    Smart packaging ditches pricey sensors for color-changing ink and AI

    University of Vaasa researchers just developed a smart packaging method that pairs color-changing inks with artificial intelligence. The system catches subtle environmental shifts with near human-eye accuracy, offering an affordable, easily recyclable way to track freshness in foods and medicines without electronic sensors.

  8. Reach 44

    Stanford study finds AI is gutting entry-level jobs

    A Stanford study using ADP payroll data revealed that employment for 22–25-year-olds in AI-exposed roles like software engineering and customer support dropped 16% since late 2022. While veteran staff stay protected by tacit on-the-job knowledge, recent grads rely on book smarts that AI can easily mirror, threatening to break the entry-level career ladder entirely.

  9. Reach 37

    OpenAI's bio model supercharges stem cell reprogramming

    OpenAI just teamed up with Retro Biosciences to create GPT-4b micro, a specialized model for protein engineering that redesigned Yamanaka factors. In lab tests, the redesigned proteins boosted stem cell reprogramming marker expression by over 50-fold compared to wild-type controls while improving DNA repair. The breakthrough could drastically speed up regenerative therapies and anti-aging research.

  10. Reach 31

    Surrey's AI navigates dense cities without GPS

    Researchers at the University of Surrey just developed PEnG, an AI navigation system that bypasses GPS satellites entirely. By pairing satellite maps with feeds from a simple monocular camera, it slashed urban positioning errors from 734 meters down to just 22 meters. Why it matters: it offers a reliable, low-cost fix for autonomous cars and delivery robots navigating tunnels and high-rise concrete jungles.

  11. Reach 34

    Chan Zuckerberg Initiative builds an AI that reasons like a cell

    CZI just released rBio, the first AI reasoning model trained on virtual cell simulations rather than raw experimental data alone. Powered by the foundation model TranscriptFormer, it lets researchers use plain language to deduce complex gene perturbations and cellular responses before running costly wet-lab experiments. It effectively brings conversational logic to cellular biology, though current capabilities focus strictly on gene perturbation.

  12. Reach 31

    Microsoft's MindJourney lets AI mentally roam 3D worlds

    Microsoft just introduced MindJourney, a research framework that helps vision-language models solve tricky 3D spatial tasks. Instead of guessing from a flat 2D image, it uses a video-based world model to simulate moving through virtual rooms and check unseen angles, boosting spatial benchmark accuracy by 8% without any retraining.

  13. Reach 32

    NASA and IBM just built an AI model for the Sun

    NASA and IBM just launched Surya, an open-source heliophysics foundation model trained on 9 years of solar observation data. It can visually predict solar flares two hours in advance—beating existing benchmarks by 16%—to help protect satellites and power grids from space weather chaos.

  14. Reach 31

    Inclusion AI pits LLMs against each other in real-world apps

    Inclusion AI just introduced Inclusion Arena, an LLM leaderboard that tests models inside real-world production apps rather than synthetic benchmarks. By quietly routing prompts from apps like Joyland and T-Box into blind battles, it uses Bradley-Terry modeling to rank outputs based on genuine user choices. Claude 3.7 Sonnet and DeepSeek v3 led early testing across 500,000 comparisons.

  15. Reach 31

    Berkeley and Stanford researchers build GEPA to optimize LLMs without costly RL

    Researchers from UC Berkeley, Stanford, and Databricks just introduced GEPA, a prompt optimization method that ditches reinforcement learning. Instead of thousands of trial-and-error runs guided by simple scores, GEPA has models reflect on detailed execution logs and error traces, beating traditional methods with up to 35 times fewer trials.

  16. Reach 35

    MIT's generative AI designs brand-new superbug killers

    MIT researchers just used generative AI algorithms to dream up over 36 million theoretical molecules, successfully discovering novel antibiotic candidates that kill drug-resistant gonorrhea and MRSA in mice. Rather than modifying existing drugs, these candidates feature completely unseen chemical structures that attack bacterial cell membranes.

  17. Reach 49

    OpenAI's reasoning system snags gold at IOI

    OpenAI just announced that its reasoning system scored at a gold medal level in the 2025 International Olympiad in Informatics (IOI) online contest. The model finished #1 among all AI participants and ranked 6th overall against top human coders, continuing a run that recently saw IMO gold-level scores and an AtCoder podium.

  18. Reach 31

    Meta FAIR's TRIBE model wins Algonauts 2025

    Meta FAIR just won first place at the Algonauts 2025 brain modeling competition with TRIBE, a 1B-parameter multimodal model. By fusing Llama 3.2, V-JEPA 2, and audio features, it accurately predicts complex fMRI brain activity while people watch movies. The full code and research paper are already available.

  19. Reach 31

    TU Graz's AI-powered blueprint for custom proteins

    TU Graz researchers just unveiled a computational framework from their HelixMold project that combines parametric design and AI to build custom proteins from scratch. The system generates thousands of backbone variations and deploys dedicated machine learning models to vet flexible loop regions. Why it matters: It pushes protein engineering away from tweaking natural molecules toward on-demand catalyst design for drugs and bioplastics.

  20. Reach 32

    DeepMind's Perch AI decodes wild soundscapes

    DeepMind just updated Perch, its open bioacoustics AI model now available on Kaggle. Trained on nearly twice as much data from public sources like Xeno-Canto, the model now handles challenging underwater audio from coral reefs alongside terrestrial birds. Conservationists can search complex audio and build custom wildlife classifiers in under an hour.