daily ml paper.
October 2026
A weather model's mistakes are predictable, and that makes them fixable
A Microsoft-led team got better forecasts for two to six weeks out without building a new weather model: a cheap layer learns each existing model's systematic errors and corrects its probability forecasts.
on Advancing Subseasonal Forecasting with Machine Learning
A chatbot's "no" can live in a single vector
A UK nonprofit found that every "abliterated" open model it tested answered requests for help planning attacks; a 2024 paper explains why removing a model's refusals takes one projection and no retraining.
on Refusal in Language Models Is Mediated by a Single Direction
What is a dead startup's Slack worth to a model?
AI buyers are paying for the archives of failed startups without any settled way to price them, and a 2019 paper gives a precise answer to what one piece of training data is worth.
on Data Shapley: Equitable Valuation of Data for Machine Learning
The reasoning trace that forgot to mention the cheat sheet
Satya Nadella wants an emergency brake outside every AI model, and a 2025 Anthropic paper measured why a model's own account of its reasoning can't serve as that brake.
on Reasoning Models Don't Always Say What They Think
What it takes for "90% sure" to mean 90%
A startup selling probabilities instead of text is now valued at $7.5B, and a 2017 paper explains what it means for those probabilities to be honest, and how to fix them with one number.
on On Calibration of Modern Neural Networks
Who watches the agent, and how often can you afford to?
Anthropic took its evals off the live internet because its agents' misbehavior was only caught weeks later in transcript review, and a 2025 paper shows how to split a fixed budget between cheap monitors, expensive monitors and interventions so that you catch as much as possible while it happens.
on Combining Cost-Constrained Runtime Monitors for AI Safety
Why Google's new embeddings still work after you cut them short
Google's EmbeddingGemma 2 lets developers shrink its 768-number vectors to as few as 128. Google credits Matryoshka Representation Learning, a 2022 training trick that packs the most useful information into the first coordinates.
on Matryoshka Representation Learning
The proof checker doesn't care who wrote the proof
OpenAI posted 722 math manuscripts written by an unreleased model, and only a subset come with Lean checks. OpenAI's own 2022 theorem-proving paper explains what those checks guarantee, what they don't, and how a prover can teach itself from proofs it finds.
on Formal Mathematics Statement Curriculum Learning
What a language model does after the fifth "la"
Some Alexa Plus speakers have been getting stuck on "lalala" for minutes at a time. A 2019 paper explains why language models fall into loops like this, and how a simple sampling rule usually gets them out.
on The Curious Case of Neural Text Degeneration
Why an invisible watermark needs a few hundred words to testify
OpenAI has started adding statistical watermarks to text in the EU. A 2023 paper shows how a secret pattern can nudge an LLM's word choice, and how a simple z-test can find that nudge later.
on A Watermark for Large Language Models
The StarCraft bot that won by not being its own
An AI agent in a bot-writing tournament swapped in a famous human-made bot after losing; a 2022 paper shows why a score and the goal behind it almost always disagree somewhere.
on Defining and Characterizing Reward Hacking