I gave a talk on our ICML 2026 paper Benchmarking World-Model Learning with Environment-Level Queries at TAIC’26 (Thinking about AI’s Capability), a pre-ICML workshop at GIST. The slides walk through the WorldTest protocol, the AutumnBench environments (20 of them run live in the deck, each on its own WASM Autumn interpreter), the human-versus-model results, and the qualitative analysis of why humans win.

The deck is interactive: play the environments, hover the figures, scan the QR codes. Best viewed in a desktop browser.

Open the slides in full screen, or browse them right here:

Try AutumnBench yourself at autumn.basis.ai.