We apply mechinterp techniques on VPT, OpenAI's Minecraft agent. We also find a new case of goal misgeneralization - VPT kills a villager when we force one to stand under some tree leaves.
Abstract
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.
2min teaser video
Media
Feel free to use this GIF in your presentations on the importance of AI safety :) More formats and speeds in the website below.
TL;DR
We apply mechinterp techniques on VPT, OpenAI's Minecraft agent. We also find a new case of goal misgeneralization - VPT kills a villager when we force one to stand under some tree leaves.
Abstract
2min teaser video
Media
Feel free to use this GIF in your presentations on the importance of AI safety :) More formats and speeds in the website below.
Website
https://sites.google.com/view/vpt-mi