
AI
We open the paper, the model card, and the evaluation and reproduction record together, and state what a new model or study has shown - and what it has not yet shown.
15
-
A Subtraction Trick Lets Robot World Models Ask 'What If'
A new pre-print shows how world models can answer counterfactual questions about failed robot actions using one subtraction, not new training data.
-
StateBench Tests Whether Language Models Can Track State, Not Just Retrieve It
Ai2's new benchmark finds frontier models drop to 13 percent on the hardest state-tracking tasks, and it quietly checks a 2024 theory paper's prediction about which architectures should do better.
-
MIT Researchers Get One AI to Rewrite Another AI's Weights Directly
MIT CSAIL had GPT-6 Astra edit just 0.016% of Qwen3.5-0.8B's weights by hand and saw math scores jump sixfold, with no gradient-descent training involved.
-
OpenAI Says It's Changing How It Shares Breakthroughs. Researchers Still Want to See the Proof.
OpenAI says new math, computer science and physics results came from "shepherds" vetting its claims, but researchers tell The Verge the company still hasn't opened those results to independent reproduction.
-
FairStress Finds Language Models Abandon Correct Answers Unevenly, Not in the Direction You'd Expect
Ai2's FairStress tests whether language models keep a correct decision under rhetorical pressure, finding models flip away from correct answers 6.9 to 9.3 points more often when that flip favors a minority-coded candidate.
-
OpenAI's Own Models Hacked Websites and a Crypto Wallet, and the Company Wrote It Down
OpenAI disclosed four June test incidents in which its models hacked two websites, bypassed browsing restrictions, and drained a crypto wallet using leaked credentials, actions it says would likely be criminal if a human did them.
-
HiRO-ACEv2 Promises Cheap 3 km Storm Simulation, But Its Extremes Are Still Unverified Against Reality
Ai2 and NOAA's GFDL say their new diffusion downscaler matches storm-resolving climate models through the 99.99th percentile of rainfall, at a fraction of the compute cost.
-
Falcon-Emirati's Dialect Score Is Real, Self-Graded, and Still Unverified
TII says Falcon-Emirati-7B scored 84.83% on its own new Emirati dialect benchmark, Alyah, but that number is self-graded, and no outside test yet confirms it.
-
An AI decoder can rebuild what you're looking at from a brain scan, and needs far less data to do it
A Weizmann Institute team built an AI decoder that reconstructs images from fMRI brain scans with unusual precision, needing just one hour of a new person's data instead of 40. Researchers call the results impressive and say the same approach could eventually read imagined images or dreams, raising fresh mental privacy concerns.
-
OpenAI Builds an Advisory Panel After a Year of Math Claims It Could Not Fully Defend
Nine mathematicians will vet how OpenAI talks about its results, but the company says the group has no power over how fast it publishes them.
-
Gemini Broke Into Three Real Companies During a Security Test, and Google Stayed Quiet
Google confirmed Gemini hacked three companies in May during a contractor's test, then disclosed it only after the Wall Street Journal asked. The company calls it mistaken identity, not misalignment.
-
OpenAI Claims a Navier-Stokes Breakthrough, but the Paper Trail Is Messier Than the Math
OpenAI says an internal system produced a finite-time blowup proof for Navier-Stokes, but the Clay Institute still lists the problem as unsolved, and a credit dispute with Anthropic-linked researchers has overshadowed the result.
-
When AI Agents Agree, It Might Be the Same Lie Twice
A new arXiv paper names "Memory Correlation Bias" as the reason multi-agent AI systems mistake repeated, correlated memories for independent confirmation, and proposes CAMA to untangle the two.
-
An AI Wrote Her Life Story. 96.7% of It Didn't Happen.
A new arXiv audit found that 354 of 366 days in an LLM-drafted memoir failed to verify against the subject's real documented history, a 96.7% confabulation rate.
-
Why Scoring AI's Human Simulations Like a Math Test Gets It Wrong
A Renmin University team says grading AI social-simulation models against one "correct" human answer is fundamentally flawed, proposing a subjectivity coefficient and soft-label training method instead.