Reproduce “AvE: Assistance via Empowerment” 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
Yuqing Du, Stas Tiomkin, Emre Kiciman, Daniel Polani, Pieter Abbeel, Anca Dragan
A new paradigm for human-agent collaboration, Assistance via Empowerment (AvE), enables AI agents to assist humans by increasing their environmental control rather than inferring specific goals. This approach led to a 33% average success rate for human-agent teams in a continuous control task, outperforming baseline methods by 16 percentage points, and demonstrated superior robustness in ambiguous goal scenarios.
Key insights
- Assistance based on increasing human empowerment offers a robust, goal-agnostic alternative to traditional goal inference, especially when human goals are uncertain or misspecified.
- Optimizing for human empowerment can lead to the emergence of naturally helpful assistive behaviors, such as system stabilization, without requiring explicit goal models or domain-specific heuristics.
- This paradigm promotes human autonomy by increasing the human's options and control over the environment, rather than attempting to predict or dictate their actions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.