World models stumble on human belief as DeepMind agents enter EVE
MENTIS improved action prediction by adding belief and goals to a world model. DeepMind will test long-horizon EVE agents offline, while Honor's humanoid-robot run rests on CCTV's account without independent timing.
Artificial Intelligence··Evening
MENTIS adds mental variables to action prediction
Mental World Modeling adds variables such as belief, attention, goals, and social relationships to world models that otherwise simulate only physical state. The training-free MENTIS pipeline raised direct-answer F1 on the 448-scene Menti-Bench from 63.3 to 87.9; the human baseline is 98.5. Within the pipeline, GPT-4.1 scored 84.9, above the 83.6 of a direct GPT-5.6-Sol setup using self-consistency. Each action is separated into a physical carrier and a mental payload.[1]
DeepMind will test long-horizon action in an offline EVE copy
The research partnership between Google DeepMind and Fenris Creations targets agents that keep learning, remember, plan over long horizons, and cope with many players at once. Its first stage will run on an offline copy of EVE Online before anything reaches live players. The teams previously released Aura Guidance, which uses Gemini to pass player-written knowledge to new pilots. No performance result has been reported for the new partnership; the announcement defines the research environment and intended capabilities.[2]
Honor's running time awaits independent measurement
China's state broadcaster CCTV said Honor's bipedal Lightning robot ran 100 metres in 9.32 seconds at the World Humanoid Robot Games. That is below Usain Bolt's 9.58-second time from 2009. Al Jazeera published the report as CCTV's account, and no independent timing is available. The reported time therefore presents a striking company demonstration of physical movement, although it cannot be treated as a verified sporting record.[3]
