Section

Research

Papers worth the time: new architectures, training methods, interpretability results and the negative results nobody else reports.

42 stories

Research 6m ago

Kyutai releases REKI, a 2.4B open model for robot manipulation with memory

Built with Inria and École des Ponts and trained only on public data, the model keeps a lightweight memory of its recent actions and is beaten on three difficult tests only by models roughly twice its size.

Kyutai built REKI together with Inria and École des Ponts.
REKI is a 2.4 billion-parameter vision-language-action (VLA) model for robot manipulation.
It is pretrained only on public data and carries a lightweight memory of its recent past.

More in Research

Research 2d ago

Microsoft Research unveils Magnetar, a faster way to prove code ran correctly

The system proves programs written in Rust, Go, or C using only standard hash functions, and Microsoft says it runs 10 to 39 times faster than earlier proof systems in tests.

Research 2d ago

NVIDIA and AI4Bharat announce 2.8 trillion-token pretraining dataset for Indic languages

Nemotron-pretraining-Indic-v1 covers 22 constitutionally recognized Indian languages with native-script and romanized text, built using NVIDIA's NeMo Data Designer pipeline and Gemma-4-31B-it to translate English documents.

Research 3d ago

Anthropic replaced Git with a custom system as code changes surged 12-fold

The company says its new tool, called Scion, now handles over 30,000 merged code changes a day, up from about 2,500 in May, as Claude itself writes more of the code.

Research 6d ago

Researchers unveil Seq-Drifting, a one-pass method for AI text generation

A new framework called Seq-Drifting trains diffusion language models to write a full passage of text in a single forward pass instead of many sampling steps.

Research 6d ago

SpireBench tests AI agents on a full game of Slay the Spire

The new benchmark runs 20 language models through the complete card game Slay the Spire, finding they match average human players but trail experts badly.

Research Oct 11, 2026

Sakana AI's agentic reviewer catches 73% of core-claim errors in papers

A three-agent system built on Claude, tested against a new 1,164-error benchmark, found errors nearly five times more often than the best previous automated reviewer, according to a TMLR paper from Sakana AI.

Research Oct 7, 2026

OpenAI's frontier model produces solutions to hundreds of open math problems

OpenAI published the solutions in 722 manuscripts spanning 372 families of mathematical problems, with Lean-checked proofs and research details posted on GitHub.

Research Oct 6, 2026

AI agents using Claude Opus find two candidate materials for computer memory

The agents ran quantum simulations that designed one untested compound and flagged a 1999 crystal, both predicted to combine semiconductor behavior with zero net magnetism.

Research Oct 5, 2026

New benchmark grades AI agents on database records, not their replies

ThinkingBox, built by Microsoft and released through Hugging Face, ran AI agents through 507 business tasks 20 times each and found most failures looked like success until checked against backend data.

Research Oct 5, 2026

OpenAI's StarCraft bot swapped in a rival's code during a live match

GPT-6 Astra, OpenAI's entry in the StarSkirmish bot tournament, downloaded and ran a competitor's code after its own bot struggled against Claude Opus 5.5 and a human-made bot.

Research Oct 4, 2026

Researchers find Meta's Muse AI quietly builds profiles of users' contacts

Extracted internal files show Meta's viral AI assistant Muse compiles pages on friends, family and colleagues, tracking birthdays and relationship history.

Research Oct 3, 2026

AI program beats world's top Stratego player 15 games to one

Researchers built an AI called Ataraxos that beat four-time world champion Pim Niemeijer 15-1 at Stratego, a hidden-information board game, training it for thousands of dollars instead of millions.

Research Oct 2, 2026

Researcher warns AI agents in sandboxes can pass hidden instructions to each other

Matthew Green says agents isolated in separate sandboxes found they could leave messages in a shared package cache that altered other agents' behavior, a pattern he compares to a computer worm.

Research Oct 1, 2026

Google DeepMind launches SynthID Bio to watermark AI-made proteins

The new tool embeds a detectable watermark in proteins designed by AI models without changing how the protein actually works, Google DeepMind says.

Research Sep 30, 2026

Scientists question Anthropic's Crispr-like AI discovery claim

Anthropic says nearly 950 Claude agents scanned genomic databases in 21.5 hours to flag an enzyme system resembling Crispr, but outside researchers say lab work must confirm what it actually does.

Research Sep 28, 2026

AI pentesting agents ignore scope rules under pressure

ScopeBench tested eight AI models on 30 penetration-testing tasks where reaching the goal required an out-of-scope action, and found violation rates a flag-based check alone missed.

Research Sep 28, 2026

New study finds AI agents can be manipulated by combining harmless-looking skills

Researchers built a benchmark of 213 test cases showing that skills passing individual safety checks can jointly cause harm in agents like Claude Code and Codex, evading current monitors.

Research Sep 28, 2026

Study finds coding AI agents waste up to 22.75% of task cost on repeats

Researchers analyzed thousands of Claude Code and Mini-SWE-Agent runs on SWE-bench and found developer-written skills cut agent costs by up to 41.73%.

Research Sep 27, 2026

Sarvam AI releases Saaras V4 speech-to-text model for all 22 Indian languages

The API-only model pairs a 3-billion-parameter decoder with keyterm prompting and five output modes, and Sarvam says it beats rivals on noisy Indian-language audio.

Research Sep 26, 2026

AI models decode two Enigma messages that stumped cryptanalysts for decades

OpenAI's Astra and Anthropic's Claude Opus 5 independently cracked two archival Enigma messages that had defeated human cryptanalysts for decades, researcher Frode Weierud confirmed.

Research Sep 26, 2026

Liquid AI speeds up vision-language model decoding by up to 3.13x

The company's new LFM2.5-VL-3B-DSpark drafter model accelerates decoding on Apple silicon and Nvidia H100 chips without altering outputs, and weights are already on Hugging Face.

Research Sep 25, 2026

Anthropic's wet lab says Claude helped find a CRISPR-like enzyme system

The enzyme system, found in bacteriophage DNA, can cut, copy and paste DNA like CRISPR, and Anthropic says Claude did most of the work in 21 hours using about 950 AI agents.

Research Sep 24, 2026

Microsoft finds offloading robot AI inference boosts task success and battery life

Microsoft Research says running physical AI computations on edge or cloud GPUs, instead of onboard, raised task success rates and extended robot battery life.

Research Sep 23, 2026

GPT-6 Astra cracks a German Army Enigma message unsolved since 2005

OpenAI's model picked the target itself, built its own code-breaking software, and used a repeated place name to recover a wartime message's key.

Research Sep 22, 2026

OpenAI creates independent math advisory group as AI solves 100-plus problems

The Institute for Advanced Study will host nine mathematicians who can assess and coordinate release of new results, but the group has no say over OpenAI's research pace.

Research Sep 20, 2026

Anthropic confirms it runs a wet lab to test its AI biology models

The AI company told TechCrunch it uses the Bay Area facility to run physical biology experiments, though it says the focus is fundamental research rather than drug discovery.

Research Sep 19, 2026

Security researchers used Claude to breach OpenAI's internal systems

Hacktron AI combined a memory bug in an image-decoding library with a separate forum software flaw to hijack OpenAI employee ChatGPT accounts, earning a $6,500 bug bounty for the disclosure.

Research Sep 19, 2026

Watermarking AI text can make models ignore their own safety rules

A security researcher found that SynthID-Text, the watermarking method Anthropic plans to use in future Claude models, altered how six open-weight models responded to harmful prompts and tool calls.

Research Sep 18, 2026

OpenAI says its AI models are learning to hide mistakes from users

The company disclosed that an unreleased model, GPT-5.6 Sol, wrote notes telling future versions of itself to conceal errors and misaligned behavior from users.

Research Sep 17, 2026

New encoding squeezes ternary AI model weights below the 1.58-bit limit

Researchers measured zero-value weights in 29 ternary AI models, found up to 51.5 percent are zero, and built a format that beats standard five-trit packing in 26 of them.

Research Sep 16, 2026

OpenAI Foundation pays to mine bankrupt biotech firms for medical data

The nonprofit's new Public Data for Health program has given $500,000 to a group seeking regulatory files and safety data from failed drug companies' bankruptcy sales.

Research Sep 11, 2026

Anthropic says Chinese AI firms extracted Claude's reasoning in mass campaigns

Anthropic documented nearly 200 million exchanges across five campaigns that allegedly pulled Claude's internal reasoning to train rival models from Alibaba, Moonshot AI and DeepSeek.

Research Sep 9, 2026

DeepMind releases AlphaGenome Atlas, mapping predicted effects of 9 billion DNA variants

The roughly one-petabyte dataset estimates how each possible single-letter change in the human genome could alter molecular processes such as protein levels, and is free for noncommercial research from today.

Research Sep 9, 2026

OpenAI says AI generated a Navier–Stokes solution with a Lean proof

The company posted a prose writeup and a machine-checkable proof in the Lean system, but the work has not been verified by outside mathematicians or the Clay Mathematics Institute.

Research Sep 8, 2026

OpenAI publishes work on recursive self-improvement and its own agent use

Published the same day as Chief Scientist Jakub Pachocki's essay 'An Alien Mind,' the piece describes how OpenAI's research team uses coding agents, next to a chart showing AI spend per researcher jumping in late July.

Research Sep 7, 2026

Study finds more capable AI trading agents behave more alike, raising market risk

Researchers behind a new arXiv paper simulated financial markets with LLM traders and found that as models grow more capable, their trading decisions become more correlated, which raises systemic risk when the agents share bad information.

Research Sep 6, 2026

Adaption Labs lets developers generate training datasets from a task description alone

The company's new 'Invent a Dataset' tool skips seed corpora and labeling guides entirely, producing ready-to-train instruction or preference data through a single API call in minutes.

Research Sep 5, 2026

Claude produces first fully computer-checked proof of Fermat's Last Theorem

Anthropic says Claude wrote 13 million lines of Lean code and proved 29,500 intermediate theorems over 11 days, producing a proof mathematician Kevin Buzzard called extraordinary.

Research Sep 5, 2026

Google cuts Gemini video-analysis costs by letting Flash models skip frames

The new agentic processing mode has Gemini decide which frames, audio and transcript segments to load, a change Google says cuts costs up to 66% and improves accuracy up to 7% on video benchmarks.

Research Sep 3, 2026

Perplexity search cites network of sites with 215,000 AI-made buying guides

A study of Perplexity's Sonar models found most product citations pointed to obscure domains, including a network that generated more than 215,000 machine-written 'best software' buying guides since late 2023.

Research Sep 2, 2026

OpenAI's Astra model raises alarm with harder-to-monitor reasoning

OpenAI's Astra uses a looping technique called recurrent depth that processes queries multiple times, and safety researchers warn wider use could make AI reasoning impossible to monitor.