The AI Solved It Once—Then Turned the Reasoning Into Code
Researchers turn language-model reasoning into inspectable code that reuses solutions across changed questions, with lower repeat costs.
language models
The Secret Word That Colored a Story Without Ever Appearing
A hidden word never appears in a story, yet models can detect its thematic shadow—and stronger secrecy instructions may make it easier to spot.
A Half-Second Brain Signal Carried More Than the Word Itself
A silent-reading EEG study found that one half-second brain signal carried traces of the passage, not just the word being read.
AI Answers Survived Scrambling—but Key Clues Stayed Put
When researchers scrambled AI reasoning traces, models still retrieved answers—until key numbers and answer clues disappeared.
Inside a Language Model’s Feedback Loop, Words Can Tip Into Chaos
A fixed-loop experiment reveals how a language model’s continuous feedback can shift from repeating phrases to chaos—and why tiny numerical changes matter.
The Hidden Price of Teaching AI to Think in Fewer Words
A study finds AI favors language linked to human reading difficulty, and shows how rewarding brevity can push that shift further.
AI Research Ideas Traveled Far—but Stayed on Familiar Roads
A study of 524,308 AI-generated hypotheses finds they can stretch beyond nearby papers yet still follow science’s established paths.
The Hidden Chemistry in How a Language Model Reads 3D Molecules
A molecular chemistry test finds that language models can rank 3D shapes, while their reliance on hydrogen bonds exposes a striking blind spot.
A New AI Method Tracks Probability Across DNA’s Many Paths
A sampling method recovers lost probability in DNA-to-protein language models, reaching estimates that pruning could not feasibly compute.