Search “Google DeepMind”

Imperial College London paper: search-augmented LLMs often cite web pages they never opened

Three authors from Imperial College London have published the paper "Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence," examining why search-augmented LLMs frequently cite web pages they never actually open. The study tested GPT-5.6 Luna and six open-weight configurations including GLM-5.3-Flash, gpt-oss-120b, gpt-oss-20b, gemma-4-31b, and qwen3-8b using 225 facts that changed after February 16, 2026, finding that models often answer based solely on search snippets without opening the source pages; under the same set of search results, the proportion of cases where no page was opened ranged from 2.6% to 76%, and when Luna used its built-in search, 53.6% of turns relied only on snippets. When answering from snippets alone, outdated answers accounted for 6.4% and misleading citations for 6.4%, whereas when pages were opened and cited, these fell to 1.3% and 2.3% respectively; placing the correct page directly into the prompt significantly reduced errors, but some models still favored old information even when shown the new facts.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Researchers extend LeCun's JEPA into a general-purpose world model spanning physics to biology

PhAI Labs, together with Stanford, Oxford, and Princeton, proposed JEPA-Anything, which splits JEPA's single prediction into multiple specialized prediction modules and lets each module differentiate into distinct roles during training, thereby covering seven domains—including physics, robotics, weather, single-cell, and clinical—with a single architecture. In a simplified Pong environment, prediction error dropped 35%, and on unseen intervention combinations it dropped 13%; on the Burgers equation benchmark the error was nearly halved, but after 50 steps the advantage shrank to about 3%; in simulated walking-robot planning, it won two of three environments. The team also screened biological data to identify a candidate liver cancer regimen combining IL-18 with CD73 blockade, and tested it in organoids, patient tissue, and mice; the paper also notes that this does not yet show it can become a therapy, and the code and models have been made public.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.