Decision Transformer Pocket trains an offline return-conditioned policy on 4,000 mixed-quality corridor trajectories. A nearby claim pays 0.4; the larger reward requires first moving away from it, retrieving a key, crossing a door, and claiming treasure.

The behavior-cloning control sees identical states and actions but no requested return. Evaluation asks each policy to realize both low- and high-return targets from three starting positions.

The 18,371-parameter Decision Transformer selected the nearby reward in 100% of 300 target-0.4 episodes and the key-door treasure in 100% of 300 target-1.0 episodes. The 595-parameter behavior-cloning policy could not switch intention, reaching only 33.3% and 66.7% desired-terminal rates.