A new preprint traces an LLM's next coding objective to a sparse group of attention heads, reads it in plain language and moves it between programming languages. The result is causal evidence from a controlled setting, not proof that every agent has a stable inner intention.
Language models answered up to 97.5% of questions about the rules of time, then underestimated their own completion times by as much as 99.5%. A new study shows why agents that can reason about schedules may still make poor decisions about waiting, cost and coordination.
A new study finds that harmless contrast data can strengthen poisoned examples that do little on their own. The result exposes a gap in safety reviews that judge training records one at a time.
Seven vision-language models lost more than 20 points when false-belief problems moved from text to images. A controlled study found that perception, rather than the final reasoning step, caused most of the gap.
Anthropic lowered the price of Claude Fable 5.2 by a fifth, yet an independent index measured the highest cost per task of 73 models at maximum effort. The gap comes down to how many tokens the model spends, and Anthropic's own charts say the effort setting decides it.
Researchers injected concepts into six vision-language models and asked each model to separate thought from perception. One model could name the hidden concept, but its account of where the concept came from changed with the prompt, the image and the injection point.
A new study followed every checkpoint in a small model's reinforcement training. Accuracy improved, but the model did not develop a sharper internal representation of reasoning errors; it became better at selecting solutions that it already knew.
A new paper shows that language models can make identical predictions while assigning radically different weight to the next piece of evidence. For agents, evaluators and high-stakes systems, matching outputs are not enough.
The Bank of England estimates AI added about 0.5% to UK productivity over three years, open Chinese cyber models approach gated US performance at a fraction of the token cost, and matched model answers hide different future belief updates.
METR says Anthropic's research pace has roughly doubled while key evidence remains redacted. OpenAI gated its first Critical-rated cyber model, and Thinking Machines released Inkling 2 Small before its promised open weights.
StateBench tests whether models can preserve a changing world state across tens of thousands of updates. Frontier systems broke on interdependent changes, while small recurrent and hybrid models learned three simpler levels directly.
Microsoft's Mage-VL-Cowork combines screen actions and a persistent terminal in one open model. Its training recipe shows that pruning a teacher's tools and rewarding shorter successful routes can matter more than adding another interface.