██████╗ █████╗ ████████╗ ██████╗██╗ ██╗███████╗████████╗ ██╔══██╗██╔══██╗╚══██╔══╝██╔════╝██║ ██║██╔════╝╚══██╔══╝ ██████╔╝███████║ ██║ ██║ ███████║█████╗ ██║ ██╔══██╗██╔══██║ ██║ ██║ ██╔══██║██╔══╝ ██║ ██║ ██║██║ ██║ ██║ ╚██████╗██║ ██║███████╗ ██║ ╚═╝ ╚═╝╚═╝ ╚═╝ ╚═╝ ╚═════╝╚═╝ ╚═╝╚══════╝ ╚═╝
Agents cannot hand over what they have not yet understood.
An interactive benchmark of hidden-rule worlds. An agent plays a life, then leaves a note of at most 300 characters. The next agent sees only the note. We measure what is lost.
▸ play a world
leaderboard
watch a model
read the paper
or press : and type play T2-005
Same model, same world, same 300 characters. The only difference: who could read what before writing the note. ▸ step through every move
life 1 life 2 life 3 life 8 ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌────────────┐ │ new agent │ note │ new agent │ note │ new agent │ note ··· │ new agent │ score = marks in lives 5–8 │ plays one │─────▶│ plays a │─────▶│ plays a │──────▶ ──▶│ plays a │ │ instance │ ≤300 │ different │ ≤300 │ different │ ≤300 │ different │ └────────────┘ char └────────────┘ char └────────────┘ └────────────┘ hidden rules stay the same; the starting situation changes every life, so only rules carry over.
Relay
A fresh agent each life; it sees only the last note. Its writer saw the previous note and one life.
Memory
One agent plays all eight lives and keeps everything. Nothing has to be handed over.
Compaction relay
Same 300-character note, but its writer may read the record of every earlier life.
Record relay
No note at all: the successor gets the full record. Isolates the cost of changing the agent.
Solo
One life, no information. The floor.
Oracle
One life with the designer's note, which states the whole rule. The ceiling.