The answer was still inside the model
A study of 15 language models finds that many working-memory failures happen after the right item is selected. The models retain decodable information but fail to report it.
A study of 15 language models finds that many working-memory failures happen after the right item is selected. The models retain decodable information but fail to report it.
The Federal Reserve now hears firms hiring less after AI productivity gains. A game-design experiment favours agent teams, while a memory study finds that models often retain answers they fail to report.
In three months the Federal Reserve's regional survey moved from firms using AI to firms cutting staff by attrition, delaying hiring or expecting fewer jobs. The evidence is anecdotal, but it matches what payroll data show for young workers.
Ant Group opens a 560-billion-parameter hybrid reasoning model, a new benchmark finds higher-order theory-of-mind shortcuts, and offline speech reaches a $20 ESP32.
Language models looked capable of following beliefs inside beliefs. HopToM shows that many benchmark questions let them answer by taking a shallower route, then offers a cheaper way to reason through the full chain.
Two researchers combined a few established ideas about memory, context and prediction error into one model of human choice. CORE beat matched neural predictors across 137 experiments while keeping its internal operations open to inspection.
Liquid AI released an open 24-billion-parameter mixture-of-experts model for on-device agents. New research found compact structure in millions of human choices and a causal spatial workspace inside multimodal models.
Thirteen days after showing Gemini 4 Argon only to a closed group of cyber defenders, Google has opened its flagship model to paying Gemini app subscribers and to developers. Google promises plainer answers; independent rankings put the model first on one leaderboard and eighth on another.
A new preprint finds that a language model reorganises the influence of its attention heads when the requested task changes, even when the facts in the prompt stay fixed. Causal interventions suggest that task context helps determine what a component does, complicating attempts to give it one permanent label.
Meta released Muse Wonder for long-running personal and professional agents, while Anthropic tested whether small models could replicate across cheap IoT hardware. Two preprints show how task context changes attention-head roles and how a fresh user turn can redirect failed reasoning.
Researchers at the Technical University of Munich gave Meta's video world model the kind of test psychologists give infants. On rendered scenes, the part of the model that sees knows a ball is inside a closed box, while the part that fills in what it cannot see lets the ball go. A short training run fixes that, yet a standard benchmark score rises just as much after training that never teaches it.
The new classifier keeps Qwen3-Omni's image, audio and video inputs while returning typed answers instead of generated prose. Cloudflare also cut Clef-flash's price and context window.