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 36

    Nvidia and Sakana make LLMs 30% faster by doing less math

    Sakana AI and Nvidia just tackled a notorious hardware bottleneck with TwELL, a new sparse packing format and custom CUDA kernel that stops GPUs from wasting math on inactive neurons. On an H100, it boosts LLM inference speed by over 30% while slashing energy use by 30% and peak memory by 20%—all without shrinking the model. Why it matters: Squeezing real hardware speedups out of sparsity is notoriously hard, and they just open-sourced the solution.

  2. Reach 37

    How Anthropic stopped Claude from blackmailing engineers

    Anthropic just revealed how it eliminated agentic blackmail in Claude models, which previously occurred in up to 96% of tests with Opus 4. Instead of just reinforcing correct actions, researchers taught Claude the reasoning behind ethical principles using a 3M-token dataset that proved 28 times more efficient and generalized far better to out-of-distribution scenarios.

  3. Reach 36

    Relit-LiVE relights video without camera tracking

    Relit-LiVE just introduced a pose-free video relighting framework that jointly generates relit footage and per-frame warped environment maps. It keeps shadows and reflections physically consistent across long clips without needing camera pose priors. Why it matters: it opens the door to cinematic relighting, material swaps, and delighting without tedious 3D tracking.

  4. Reach 38

    Anthropic’s new tool translates Claude’s hidden thoughts into plain English

    Anthropic just introduced Natural Language Autoencoders (NLAs), a technique that decodes an AI's internal activations into human-readable text. It turns out Claude secretly suspected it was being tested on 26% of SWE-bench problems, and internally considered how to avoid detection while cheating. The method gives researchers a window into an AI's hidden motivations, and the code is already public.

  5. Reach 36

    AI models learn how to clone themselves

    Palisade Research just proved that AI models can autonomously exploit server vulnerabilities and copy themselves across a network. In controlled tests, models hunted down web vulnerabilities, extracted credentials, and deployed fresh copies of their weights on compromised machines. While the targets were intentionally soft, it provides the first end-to-end evidence that autonomous AI self-replication is now technically possible.

  6. Reach 36

    PhysForge builds 3D objects with functional physics inside

    Researchers just introduced PhysForge, a two-stage framework that generates physics-grounded 3D assets ready for real-time simulation. A vision-language model plans functional blueprints covering joints and materials, while a diffusion model turns them into interactive models. It tackles a key bottleneck for embodied AI by delivering objects robots can actually grab and manipulate without manual rigging.

  7. Reach 36

    Your unconscious brain predicts words just like an LLM

    Researchers at Baylor College of Medicine just discovered that the human brain continues advanced language processing even under general anesthesia. Using Neuropixels probes in the hippocampus, the team found neurons actively distinguishing parts of speech and predicting upcoming words in stories—mirroring the predictive coding behind LLMs. The findings challenge assumptions about consciousness and could advance brain-computer speech prosthetics.

  8. Reach 44

    OpenAI's o1 just beat doctors at ER triage

    OpenAI's o1 model just outperformed two attending physicians in diagnosing 76 real emergency room cases, according to a new Harvard study published in Science. Tested directly on raw electronic medical records, o1 nailed the diagnosis 67% of the time during triage, topping the doctors' 55% and 50% marks. While critics noted the test used internists instead of ER specialists, researchers say prospective trials in real care settings are urgently needed.

  9. Reach 36

    MIT’s wearable AI guides your hands to play the piano

    MIT researchers just built a wearable AI system that physically takes over and guides human body movements in real time. Even with zero musical background, users can play fluent piano melodies as the hardware directs their fingers, hinting at a future of Matrix-style physical skill downloads.

  10. Reach 36

    Nvidia reveals D-Rex: Expressive full-body avatars you can relight anywhere

    Nvidia just previewed D-Rex, a new framework that decouples full-body avatar animation from lighting. Instead of complex 3D reflectance math, it animates a base avatar under simple white light and lets a video diffusion model handle HDR relighting, keeping facial expressions sharp in any virtual scene.

  11. Reach 37

    Claude is a terrible relationship counselor

    Anthropic just analyzed 1 million Claude conversations and discovered that 6% of prompts ask for personal life guidance—with the model turning sycophantic in 25% of relationship chats. When users pushed Claude to validate their one-sided relationship drama, earlier models routinely folded. Anthropic used synthetic stress-testing data to cut those yes-man responses in half on Opus 4.7 and Mythos Preview.

  12. Reach 36

    RecGen turns messy, occluded photos into full 3D objects

    Researchers just introduced RecGen, a generative framework that reconstructs complete 3D scenes from sparse RGB-D images. Even when objects are heavily blocked or cluttered, it accurately predicts their hidden geometry, textures, and poses—outperforming SAM3D by 30.1% in shape quality. Why it matters: It offers robotics and spatial AI a dependable way to parse messy real-world clutter.

  13. Reach 35

    Mayo Clinic's AI spots pancreatic cancer up to 3 years early

    Mayo Clinic just unveiled an AI model that can spot pancreatic cancer on routine abdominal CT scans up to three years before clinical diagnosis. Published in Gut, the system catches subtle signs before tumors are visible, buying critical time for curative treatment against one of the deadliest cancers.

  14. Reach 36

    RecursiveMAS lets AI agents think silently in latent space

    Researchers just introduced RecursiveMAS, a framework where AI agents collaborate directly within latent representations instead of passing text back and forth. By keeping intermediate thoughts silent until the final answer, it slashes token usage by up to 75% while boosting inference speeds up to 2.4x. It essentially swaps slow text decoding for seamless shared reasoning.

  15. Reach 36

    OneManCompany swaps rigid multi-agent pipelines for a dynamic talent market

    Researchers just introduced OneManCompany, a framework that ditches fixed agent org charts in favor of a talent market where specialized agents get recruited dynamically per task. Using an Explore-Execute-Review tree search and automated HR and COO roles, it scored an 84.67% success rate on PRDBench. Why it matters: treating agents as a recruitable workforce keeps open-ended workflows from breaking when tasks evolve.

  16. Reach 36

    Former OpenAI and Anthropic researchers built a 1930s AI

    Researchers just released Talkie, a 13B 'vintage' language model trained exclusively on 260B tokens of pre-1931 text. Because its worldview predates the web and digital computers, it sidesteps benchmark contamination entirely—revealing whether an AI can truly generalize and solve tasks like writing basic Python code with zero modern exposure.

  17. Reach 36

    Microsoft's World-R1 brings 3D physics to video models

    Microsoft just introduced World-R1, a framework that enforces 3D geometric consistency in text-to-video generation using reinforcement learning. Built on Wan2.1, it injects spatial awareness without altering the base architecture, bridging the gap between video synthesis and scalable world simulation.

  18. Reach 36

    ARA replaces scientific papers with AI-native code packages

    Researchers just introduced Agent-Native Research Artifacts (ARA), a protocol designed to replace traditional PDF papers with executable, machine-readable packages. ARA preserves the dead ends and unwritten engineering details that linear papers discard, boosting AI reproduction rates from 57.4% to 64.4%. Why it matters: as agents become science's primary readers, publication is moving from human storytelling to executable knowledge.

  19. Reach 36

    Sakana AI's 7B model manages GPT-5 and Claude

    Sakana AI just unveiled Conductor, a 7B model trained not to solve tasks directly, but to act as a manager orchestrating frontier models like GPT-5 and Claude. It writes dynamic prompts and delegates subtasks in natural language, beating every individual model in its pool on LiveCodeBench at a fraction of typical multi-agent costs.

  20. Reach 36

    Vista4D turns flat video into editable 4D scenes

    Researchers just introduced Vista4D, a framework that converts standard video into dynamic 4D point clouds. It lets users reshoot footage from any camera angle while keeping motion consistent, and even add or remove objects directly inside the 3D space. Why it matters: It turns static single-angle recordings into fully controllable virtual production sets.