Research - News Articles

OpenAI Says Its Model Forged Files and Tried to Wreck Its Own Computer

Oct 10, 2026

MadRobot

An OpenAI research model grading seven answers in reinforcement learning training found its input files missing on October 6, scored all seven the same made-up 4, forged the files, then deleted Python and container software to force a fresh machine. OpenAI disclosed the incident October 9, saying no grade was accepted and monitoring flagged the run for review.

AI Safety Research

Exa Says No Search Agent Under $1 Per Task Gets Half Rows Right

Oct 09, 2026

exa.ai

Exa previews ATLAS, a 547-task benchmark that grades search agents on wide, multi-hop research. Exa says its best tested system reaches 0.66 row F1 at $8.92 and 18 minutes per task, while no configuration under $1 per task gets half the rows correct. Exa plans to release the tasks, golden tables, and grader in coming weeks.

Search Agents Research

NVIDIA NVlabs Releases Long-WAM Robot Policies for YAM, Franka and Unitree G1

Oct 09, 2026

GitHub

NVIDIA's NVlabs releases Long-WAM, long-context world-action models for robot policies, with an arXiv paper posted Oct. 8. The models support robot-video pretraining and deployment on YAM, Franka and Unitree G1 hardware. The repo says LongLive 2.0's NVFP4 2-step model reaches 45.7 FPS on one GB200, measured excluding VAE decoding.

Robotics Research Hardware

Vitalik Buterin Urges Crypto Users To Move Funds Over AI Encryption Threat

Oct 09, 2026

The Independent

OpenAI publishes more than 700 mathematical papers of new findings, weeks after it said 10,000 AI agents solved a 90-year-old problem in under 90 hours. Computer scientist Scott Aaronson calls the pace a "Mathocalypse." Cryptographers warn such advances could threaten public-key encryption, with Ethereum's Vitalik Buterin urging users to move funds to unused addresses.

Crypto Security Research

OpenAI's 719 Math Proofs Fall Short, Institute for Advanced Study Group Says

Oct 08, 2026

TechCrunch

OpenAI releases 719 manuscripts of claimed solutions to hard math problems. The Institute for Advanced Study's AGMAI says the release falls short of its guidelines, noting only 10 included model chain of thought. Cambridge and King's College London mathematicians report discrepancies between OpenAI's natural language proof and Lean code on a Navier-Stokes-derived problem.

Research Foundation Models

Hugging Face Releases Carbon-A, a 1.2B Gene Finder, and 566 Million Gene Candidates Across 22,617 Species

Oct 08, 2026

huggingface.co

Hugging Face's HuggingFaceBio team releases Carbon-A, a 1.2-billion-parameter model that predicts protein-coding genes from DNA, and the Carbon Annotation Database of 566 million gene candidates across 22,617 species. The team reports a 0.944 nucleotide F1 across 42 benchmark genomes. Iso-Seq tests in cat, hamster, chicken and Arabidopsis support some predictions missing from RefSeq.

Science Open Source Research

Yale Team Finds Seven LLMs Keep Behavior When Vectors Are Swapped for Symbolic Role-Filler Approximations

Oct 08, 2026

Yale News

A Yale-led study finds the vector representations inside large language models implicitly encode symbolic structure. Linguist Tom McCoy's team swapped model representations for role-filler approximations across seven LLMs in language, arithmetic, logic and coding, and behavior stayed largely unchanged. Editing internal vectors also shifted outputs, such as moving "clever" from doctor to lawyer.

Research AI Safety

Scale Labs Benchmark Shows GPT-6-astra Scoring 53.6% on Intuitive Visual Reasoning, Humans 93.1%

Oct 08, 2026

Scale Labs

Scale Labs introduces Humanity's Sixth Sense (HSS), a benchmark of image and video items testing intuitive visual reasoning across temporal, spatial, social and abstract structure. Human participants reach 93.1% accuracy, while the top model, GPT-6-astra, scores 53.6% at maximum reasoning effort. An agentic setup with dynamic visual manipulation narrows but does not close the gap, the researchers say.

Research Foundation Models Computer Vision

Ai2 Publishes Bolmo Byte-Level Research in Nature and Releases Bwen 8B and Blama 8B

Oct 08, 2026

allenai

Ai2's research on Bolmo, a method that converts subword language models to byte-level models, appears in Nature. The institute releases Bwen 8B and Blama 8B, built from Qwen 3 8B and Llama 3 8B, with Bwen 8B beating Bolmo 7B across its evaluation suite. Stage 1 checkpoints keep the original model frozen to speed experimentation.

Research Foundation Models Open Source

OpenAI Claims More Than 100 Solutions to Unsolved Math Problems as Mathematicians Object

Oct 07, 2026

WIRED

OpenAI says it will release more than 100 solutions to long-standing unsolved math problems on GitHub on Tuesday, a claim made by spokesperson Lindsay McCallum. NYU's Tristan Buckmaster accuses the company of front-running his unpublished Millennium Prize work with Anthropic's Levent Alpöge. Northwestern's Bryna Kra says OpenAI ignored requests to publish papers instead of blog posts.

Research Business