Can a model hide a skill from safety tests?
One preprint built a model with an encrypted skill that the usual tests could not draw out.
Ask your own in the app → TopicEvaluation · AI & Machine Learning
How do researchers test machine learning models?
Evaluation is how researchers check whether a model actually works. The papers we have covered probe hidden skills, unsafe chains of harmless requests, medical image reading, made-up search terms and the tests models write themselves. A recurring lesson: how you measure can change the answer as much as the model does.
Paper by paper
The papers we've covered
- AI & ML Evaluation
ICLR 202010 min listen
Can Machines Invent Tools Nobody Taught Them?
Two teams of learning agents played hide-and-seek hundreds of millions of times with no instructions. They invented shelters, ramps, and a way to surf on boxes.
Peer-reviewed Simulation experiments1 environmentPeer-reviewed conference paper Moderate evidence
- AI & ML Evaluation
arXiv8 min listen
Does Reward Training Make Chatbots Power-Hungry?
A classic worry says trained models learn to seek power. Tracked through training, one model's rate rose to 46.1%, then fell to 11.8%.
Preprint Checkpoint experiments1 model family Preliminary evidence
- AI & ML Evaluation
arXiv8 min listen
Can AI Give Doctors a Useful Second Opinion?
Eight AI models played a panel of specialists on 190 hard published cases. They reasoned well, but the final call was the weak spot.
Preprint Benchmark on published cases Preliminary evidence
- AI & ML Evaluation
arXiv8 min listen
Does How You Feed AI the Calendar Change Its Forecasts?
Forecasting models can now take the hour, weekday and holidays as inputs. How you encode that calendar can swing holiday forecast error by up to 83%.
Preprint Benchmark study Preliminary evidence
- AI & ML Evaluation
arXiv8 min listen
Can AI Write Tests That Catch Bugs in Quantum Code?
A test that waves broken code through is worse than no test. This study checks whether AI-written tests for quantum programs can tell right from wrong.
Preprint Benchmark study Preliminary evidence
- AI & ML Evaluation
arXiv8 min listen
Do Fancy Training Goals Beat Plain Error for Table Data?
This preprint fairly retested 18 clever training goals for spreadsheet prediction on 41 datasets. Most did not consistently beat the plain, well-tuned standard.
Preprint Large benchmark Preliminary evidence
- AI & ML Evaluation
arXiv8 min listen
Can We Stop AI From Adding Made-Up Words to Searches?
Search systems ask language models to add related words to short queries, and some are invented. This preprint cuts made-up terms by 85% without hurting search.
Preprint Benchmark experiments Preliminary evidence
- AI & ML Evaluation
arXiv8 min listen
Can a General AI Model Read Breast Ultrasound Scans?
Two general AI models read 2818 breast ultrasound images with no extra training. On one dataset, calling every image benign would have beaten them.
Preprint Benchmark on 2818 images Preliminary evidence
- AI & ML Evaluation
arXiv8 min listen
Can Harmless Requests Make an AI Agent Cause Harm?
A harmless-looking request can make an AI assistant email an internal report outside the company. This study builds a system that hunts for such chains.
Preprint Simulated red-teaming experiments Preliminary evidence
- AI & ML Evaluation
arXiv8 min listen
Can an AI Learn When to Say 'I Don't Know'?
Chatbots guess confidently even when they don't know. This study compares two ways to teach a model when to hold back: retrain it, or add a separate checker.
Preprint Controlled experiments1 model Preliminary evidence
- AI & ML Evaluation
arXiv9 min listen
Can an AI Model Hide a Skill That No Test Can Find?
Safety teams test AI models for dangerous skills. This preprint hides an encrypted skill inside a model's weights, and the usual tests could not draw it out.
Preprint Cryptographic construction + small experiments Preliminary evidence
- AI & ML Evaluation
Artificial Intelligence Review8 min listen
Do LLMs Beat Classic Models at Catching Network Attacks?
Small language models learned to flag network attacks and scored above 95 percent. But a plain classic method, XGBoost, still won, faster and far cheaper.
Peer-reviewed Benchmark1 dataset4 traffic classes Limited evidence
Questions people ask
Short answers, from the papers.
Does the test setup change the ranking?
Yes. In one forecasting study, simply encoding the calendar differently changed which model looked best.
Ask your own in the app →Is a test that runs a good test?
No. Model-written tests usually ran but were better at passing correct code than catching wrong code.
Ask your own in the app →Keep up with AI & Machine Learning research, ten minutes a paper.
The papers that matter in this hub, explained, and you can ask each one questions.
