Reproduce “Decision Titan: Test-Time Training for Long-Term Memory in Offline Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jude Waide, Robert Lieck
Decision Titan adds an episode-updated MLP memory to an offline reinforcement-learning policy, using test-time gradient updates to store key–value associations while self-attention handles recent context. In a controlled T-Maze-like task with a five-step context and six-step dependency, it reaches mean return 0, while Decision Transformer averages about −0.5. X-Maze recall depends on cues; results do not establish general RL performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.