METTA
A meta-reinforcement learning framework that employs dual-timescale memory decomposition: causal neural memory handles within-episode adaptation, while compact episodic-level summaries support cross-episode long-range inference. It achieves long-term memory and test-time training capability while maintaining linear complexity.
Timeline (last 90 days)
METTA (Memory-Enhanced Test-Time Adaptation) is a meta-reinforcement learning framework that, by decomposing the time scales of fast adaptation, achieves long-term memory while maintaining linear complexity.
METTA decomposes the fast adaptation phase into two time scales: causal neural memory handles fine-grained within-episode adaptation and runs with linear sequence complexity, while compact episode-level summaries support cross-episode long-horizon task inference.
In the Meta-World ML10/45 test environments, METTA achieves over 50% success rate on unseen tasks while maintaining a low training budget and short runtime.
On the long-context, sparse-reward PointMaze task, METTA exhibits more structured exploration behavior than Transformer-based baselines.
METTA's current experiments focus on simulated benchmarks such as Meta-World and PointMaze, and whether the memory compression and summarization strategies remain effective in more complex, higher-dimensional real-world tasks still requires further validation.
All Facts (5)
METTA (Memory-Enhanced Test-Time Adaptation) is a meta-reinforcement learning framework that, by decomposing the time scales of fast adaptation, achieves long-term memory while maintaining linear complexity.
50%UnverifiedMETTA decomposes the fast adaptation phase into two time scales: causal neural memory handles fine-grained within-episode adaptation and runs with linear sequence complexity, while compact episode-level summaries support cross-episode long-horizon task inference.
50%UnverifiedIn the Meta-World ML10/45 test environments, METTA achieves over 50% success rate on unseen tasks while maintaining a low training budget and short runtime.
50%UnverifiedOn the long-context, sparse-reward PointMaze task, METTA exhibits more structured exploration behavior than Transformer-based baselines.
50%UnverifiedMETTA's current experiments focus on simulated benchmarks such as Meta-World and PointMaze, and whether the memory compression and summarization strategies remain effective in more complex, higher-dimensional real-world tasks still requires further validation.
50%