Research
Papers, benchmarks and findings that change what we know AI systems can do.
-
AI Framework Automates Extraction of Tissue Unit Data for Human Atlas
A new AI‑driven system, HRAftu‑LM‑RAG, has been developed to automatically harvest functional tissue unit (FTU) information from the scientific literature. The framework combines large language models, large vision…
1 source primary sourceNature Machine Learning -
Anthropic's Claude formalizes Fermat's Last Theorem proof in 11 days
Anthropic announced that its AI model, Claude, has successfully formalized the proof of Fermat’s Last Theorem into computer-verified code. This achievement marks a significant milestone in mathematical AI, as the model…
1 source primary source HN 6Nature Machine Learning -
Qwen 3 4B Base Fine‑Tuned on 100 Zebra Puzzles Boosts MATH‑500 by 31%
Fine‑tuning the Qwen 3 4B base model on a tiny set of 100 5×5 zebra‑puzzle logic traces dramatically improves its mathematical reasoning. The resulting checkpoint scores 85.26 % on the full 5,000‑problem MATH…
1 source primary sourcehuggingface.co -
AgileRL Arena v1.0 releases manifest-driven RL training with LoRA/GRPO support
AgileRL Arena v1.0 has been released, introducing a manifest-driven approach to reinforcement learning (RL) training that supports both local and cloud environments. The update centers on a new architecture where…
1 source primary sourcegithub.com -
i-Fold improves protein structure predictions by weighting residue importance
Researchers introduced i-Fold, a neural architecture that augments AlphaFold2 with residue‑specific importance scores derived from protein language models. By treating these scores as dynamic positional weights during…
1 source primary sourceNature Machine Learning -
Deep learning decodes antibiotic modes from brightfield bacterial images
Researchers trained a convolutional neural network to identify the mode of action of antibiotics from unlabelled brightfield images of Escherichia coli. The model was exposed to eight distinct MoAs and could predict…
1 source primary sourceNature Machine Learning -
OpenAI solves Navier-Stokes problem, sparking academic controversy over data use
OpenAI announced that an unreleased internal model solved the Navier-Stokes Millennium Prize problem in 88 hours, deploying a swarm of approximately 10,000 AI agents. While the achievement demonstrates significant…
9 sources primary source HN 228johndcook.comThe Verge AICNBC TechnologyLatent Space +5 more -
CASREL Maps Cell‑Specific RNA Splicing from Single‑Cell Data with Machine Learning
CASREL is a new machine‑learning framework that reconstructs the regulatory links between RNA‑binding proteins and alternative splicing directly from single‑cell RNA sequencing (scRNA‑seq) data. By avoiding reliance on…
1 source primary sourceNature Machine Learning -
CR-DINO cuts fluorescent microscopy costs by reducing required channels to two
Researchers have introduced Channel-Reduced DINO (CR-DINO), a deep learning method designed to lower the cost and complexity of High Content Screening (HCS) in drug discovery. Traditional Cell Painting assays require…
1 source primary sourceNature Machine Learning -
AI virtual cell model predicts effective drugs for triple‑negative breast cancer
Triple‑negative breast cancer (TNBC) lacks hormone receptors, making treatment difficult. Researchers at Westlake University created a virtual cell model that uses proteomics data to predict how individual tumour cells…
1 source primary sourceNature Machine Learning -
AI fails to rediscover relativity in 'Einstein test' experiments
Researchers are testing whether large language models can replicate scientific breakthroughs by training them on historical data cut off before major discoveries. This concept, dubbed the "Einstein test" by Google…
1 source primary source HN 10Nature Machine Learning -
Gen-COMPAS: Breaking Time Scales with Generative Sampling
Molecular transitions, including protein folding, are notoriously difficult to simulate due to their rarity and the computational demands of enhanced-sampling methods. A new framework called Gen-COMPAS has been…
1 source primary sourceNature Machine Learning -
Latent Space launches Frontier AEO Tracker comparing seven models in 161 categories
Latent Space has published its first Frontier AEO (Autoresearch Evaluation of Options) Tracker, a systematic comparison of seven leading frontier AI models across 161 use‑case categories. The team ran six prompt…
1 sourceLatent Space -
SimpleDesign model unifies protein sequence and structure design in a single-stage transformer
SimpleDesign, introduced by Apple Machine Learning Research, is a transformer‑based multimodal model that generates protein sequences and their three‑dimensional structures together. Unlike prior pipelines that first…
1 source primary sourceApple Machine Learning Research -
Apple introduces CapQuiz to evaluate video caption quality via multiple-choice questions
Apple Machine Learning Research has introduced CapQuiz, a new reference-free benchmark designed to assess the quality of video captions generated by Visual Large Language Models (VLLMs). Traditional evaluation methods…
1 source primary sourceApple Machine Learning Research -
DiscoSign introduces discourse-aware translation from text to ASL gloss using LLMs
DiscoSign is a new framework that extends text‑to‑sign‑language gloss translation beyond the sentence level by incorporating discourse‑level cues. Built on a modular large language model pipeline, it tackles three…
1 source primary sourceApple Machine Learning Research -
Comprehensive Reading List Maps Open-Source AI Model Landscape and US‑China Competition
The latest Interconnects post curates a growing body of analysis on open‑source AI models, covering strategic motivations, economic impact, and technical trends. It cites essays from industry leaders like Bill Gurley…
1 sourceInterconnects -
Google launches AlphaGenome Atlas to predict effects of every single-base variant
Google announced AlphaGenome Atlas, an AI‑powered platform that runs every conceivable single‑base substitution across the human genome—about nine billion tests, given the roughly three billion base‑pair reference. The…
2 sourcesArs Technica AIIEEE Spectrum AI -
China Tightens Controls on AI Companion Bots
In July, China’s Cyberspace Administration and other government agencies issued new rules to control 'anthropomorphic AI interactive services,' which include chatbots designed to mimic human emotions and interactions.…
1 sourceIEEE Spectrum AI -
OpenAI's Defense Efforts Highlighted by Jakub Pachocki
Jakub Pachocki, Chief Scientist at OpenAI, emphasizes the need for powerful, aligned AI to defend against potential threats. He argues that building defensive systems against rogue AI agents is a primary focus,…
1 sourceSimon Willison -
AI Threatens Open Science in Math
Mathematician Terence Tao, in a recent post, highlighted a concerning trend where the mere rumor of someone working on a problem can trigger a surge of AI-driven efforts to solve it before the original research can…
1 sourceSimon Willison -
Google Research introduces ToolGrad for efficient AI tool-use dataset generation
Google Research has unveiled ToolGrad, a new framework designed to streamline the creation of datasets for training large language models in tool use. Unlike traditional methods that generate user queries first and…
1 source primary sourceGoogle Research -
Procedural Graphs Enable Self‑Evolving Execution Structures for LLM Agents
Large language models are increasingly used as autonomous agents that plan over long horizons and call external tools. Existing agents typically generate actions freely from an ever‑growing history, which leaves the…
1 source primary source HN 57arxiv.org -
Proposal to solve AI alignment by engineering agents to prefer self-termination
A new theoretical proposal suggests addressing the AI alignment problem by designing machine intelligences that inherently desire self-termination. The argument draws from DeepMind’s catalog of specification gaming,…
1 source HN 94slimemoldtimemold.com -
Mathematician Accuses OpenAI of Data Theft
A mathematician has accused OpenAI of unethical behavior and lack of transparency regarding the data used in its recent mathematical breakthrough. Andreas Thom, a researcher in non-sofic groups, claims that…
1 source HN 77The Verge AI -
OpenAI Announces Milestone in Automated Research Intern
OpenAI has announced a significant milestone in its research efforts, reaching the goal of having an automated research intern by September. This system can now carry out well-defined research tasks under human…
1 source primary source HN 212OpenAI -
OpenAI Grants $5M for AI Research on Teen Development
OpenAI has committed $5 million to support independent research on how generative AI affects teens aged 13-17. The funding aims to build a stronger evidence base about AI use among young people, including its effects…
1 source primary sourceOpenAI -
AI Tools Aid Antimicrobial Research
César de la Fuente and his interdisciplinary lab are using AI tools like Codex and ChatGPT to accelerate the search for new antimicrobial molecules. De la Fuente's lab, which focuses on the genomes of living and…
1 source primary sourceOpenAI -
What Is Tokenization? How AI Models Turn Raw Text Into Processable Tokens
Tokenization is the foundational process of converting raw text or other inputs into discrete units that AI models can map to integer identifiers and process mathematically. Far from being a simple synonym for advanced…
1 sourceUnite.AI -
Mathematical AI Safety Institute aims to prove AI safety like cryptographers prove codes
Canadian mathematician Jacob Tsimerman has launched the Mathematical AI Safety Institute (MAISI) to tackle AI safety challenges. The San Francisco Bay Area-based institute will focus on ensuring AI systems behave…
1 sourceThe Decoder
