Adapt-1
A game agent learning system developed by Rei Labs that operates without PPO or large language models. It supports two working modes: reactive policy learning and sequential search via the built-in Machina sequence engine.
Timeline (last 90 days)
Adapt-1 is a learning system launched by Rei Labs that is neither LLM-based (non-LLM) nor uses PPO.
Adapt-1's reactive policy mode starts in an untrained state, accumulates experience through approximately 36 minutes of actual gameplay to form a policy, then disables learning and directly uses the acquired policy to complete the level.
Adapt-1's Machina sequence engine searches for a key-press sequence to reach the goal flagpole, finding a usable sequence after 403 attempts and approximately 11 minutes of real time.
The developer publicly released the code, complete experimental data, run trajectories, demo clips, and a step-by-step guide with cost information on GitHub (hsrvc/adapt1-mario).
The developer shared on Reddit a validation scenario of Adapt-1 learning from scratch and completing Super Mario Bros. level 1-1.
The test was a proof of concept, not a complete benchmark of Adapt-1's general capabilities.
All Facts (5)
Adapt-1 is a learning system launched by Rei Labs that is neither LLM-based (non-LLM) nor uses PPO.
50%UnverifiedAdapt-1's reactive policy mode starts in an untrained state, accumulates experience through approximately 36 minutes of actual gameplay to form a policy, then disables learning and directly uses the acquired policy to complete the level.
50%UnverifiedAdapt-1's Machina sequence engine searches for a key-press sequence to reach the goal flagpole, finding a usable sequence after 403 attempts and approximately 11 minutes of real time.
50%UnverifiedThe developer publicly released the code, complete experimental data, run trajectories, demo clips, and a step-by-step guide with cost information on GitHub (hsrvc/adapt1-mario).
50%UnverifiedThe test was a proof of concept, not a complete benchmark of Adapt-1's general capabilities.
50%