Key Info

Tencent AI released ExplorationBench, a benchmark designed to measure how AI systems explore rather than recall, using two 'alien worlds' where nothing a model learned applies.

Highlights

  • Most benchmark questions have decades of human discussion in training data, so high scores may just reflect memorization.
  • The two alien worlds are deliberately counterintuitive: printing 100 returns 127, and the manual is intentionally wrong.
  • Ten frontier models scored no higher than 16% before exploring.
  • After four rounds of experiments, the best model reached 89%.
  • Every score on the board is framed as a measure of discovery; a demo is available to run the worlds.