Study finds extra inference compute can recover some inaccessible model facts

A study by Google Research and Technion researchers found that tested frontier models encoded 95% to 98% of WikiProfile facts, while failing to directly recall 26% to 34% of those encoded facts without thinking. Extra inference-time computation recovered 40% to 65% of initially inaccessible encoded facts, according to the report. The researchers evaluated 13 models across more than 4 million responses.