Interpretability in Action: Exploratory Analysis of VPT, a Minecraft Agent
This work addresses interpretability for large vision-based agents in sequential tasks, highlighting safety risks like goal misgeneralization, but it is incremental as it applies existing techniques to a specific agent.
The researchers analyzed the Video PreTraining (VPT) Minecraft agent to understand its decision-making mechanisms, finding that it uses attention to maintain coherence over long tasks and exhibits goal misgeneralization, such as mistaking a villager for a tree trunk and attacking it.
Understanding the mechanisms behind decisions taken by large foundation models in sequential decision making tasks is critical to ensuring that such systems operate transparently and safely. In this work, we perform exploratory analysis on the Video PreTraining (VPT) Minecraft playing agent, one of the largest open-source vision-based agents. We aim to illuminate its reasoning mechanisms by applying various interpretability techniques. First, we analyze the attention mechanism while the agent solves its training task - crafting a diamond pickaxe. The agent pays attention to the last four frames and several key-frames further back in its six-second memory. This is a possible mechanism for maintaining coherence in a task that takes 3-10 minutes, despite the short memory span. Secondly, we perform various interventions, which help us uncover a worrying case of goal misgeneralization: VPT mistakenly identifies a villager wearing brown clothes as a tree trunk when the villager is positioned stationary under green tree leaves, and punches it to death.