Visual perception
Hover over the screen for exact cell coordinates and colors.
Benchmark 06/ARC-AGI-3/25 public games
Beam-1.0
on ARC-AGI-3.
Reflection's open 501B model drops into 25 games with no rules, no goal and no examples. It has to probe the screen, write a world model of each game in Python, and clear every level in about as few moves as a person.
The experiment
Public RHAE0 – 100
–
Levels clearedall 25 games
–
Working now
–
UNOFFICIALSelf-run on the public games with the official engine and scorecard. Published rows below are the authors' own reports with frontier models and much larger budgets. Nothing here is an ARC Prize verified score.
01 / Console
Watch Beam think, move by move.
Language reasoning
What Beam was thinking
Memory & tools
Visual memory
02 / Progress
Moves per model call.
Each line is one game. A dot marks a cleared level.
03 / Spend
Where the tokens go.
Input and output tokens per game. Cached input is free on Reflection's quota.
04 / The whole set
All 25 games, at a glance.
Latest screen of every game. Click one to open it in the console.
Level clearedLevel not clearedPlaying nowNot started
05 / Leaderboard
Next to the harnesses that cracked it.
Public-set RHAE. Open harness rows link to their sources.
06 / Look closer
Explore each game.
Moves, model calls, tokens and time for every game Beam has played.
Protocol
One open model.
No rules given.
Every game runs with the same prompt, tools and limits. The agent only ever sees the screen; scores come from the engine's own scorecard, which the agent never sees.
Beam-1.0 model card ↗RHAE (Relative Human Action Efficiency) compares the moves an agent needs per level with first-time human players; later levels weigh more, and the final score averages all 25 games. A game Beam hasn't finished counts as 0 in the headline score.