Amazon research submitted entry to the community leaderboard from ARC-AGI-3
Technique: subagents play the 25 public games, extract strategies and create a "wiki" per game
They're scorecard claims 64/183 levels beat
Zhan Shi, Hanqing Lu, Bing He, Yisi Sang, Minhua Lin
Users approve of Amazon researchers' ARC-AGI-3 entry using subagent wikis because subagents building shared knowledge outperform single-agent loops in production.
v3 Scorecard: https://arcprize.org/scorecards/c7147b87-aac6-498d-8916-8b37cf78756f
Note - since scores are self-reported we don't have a way to verify that their git repo is linked to that scorecard
Their entry: https://github.com/arcprize/ARC-AGI-Community-Leaderboard/pull/17/changes
Excited to see our work highlighted by @GregKamradt and the ARC Prize community! 🎉A-Evolve focuses on enabling agents to iteratively evolve both their problem-solving harnesses and training strategies. Great to see this approach gaining traction in ARC-AGI-3. More to come! 💪
Amazon research submitted entry to the community leaderboard from ARC-AGI-3
Technique: subagents play the 25 public games, extract strategies and create a "wiki" per game
They're scorecard claims 64/183 levels beat
Zhan Shi, Hanqing Lu, Bing He, Yisi Sang, Minhua Lin
@GregKamradt Same pattern in production — subagents building shared knowledge outperform single-agent loops. Structured capture vs. raw context stuffing dropped token cost 41%. The wiki-per-game approach makes this scale nicely to reasoning-heavy benchmarks.
Amazon research submitted entry to the community leaderboard from ARC-AGI-3
Technique: subagents play the 25 public games, extract strategies and create a "wiki" per game
They're scorecard claims 64/183 levels beat
Zhan Shi, Hanqing Lu, Bing He, Yisi Sang, Minhua Lin