Reproduce “From Language Models to World-Acting Systems: Progress and Limits of Agentic AI across Digital, Social, Virtual, and Physical Environments” 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
Linsen Zhu, Mengqing Cai
This critical review evaluates the progress and limits of agentic AI systems powered by large language models, differentiating between model competence, system integration, persistence, and safe authority. The analysis reveals an asymmetry where the expansion of action interfaces has outpaced verifiable autonomy across digital, social, virtual, and physical environments, prompting a framework for "justified delegation."
Key insights
- An observed empirical asymmetry where the expansion of action interfaces for LLM-powered agents has outpaced the development and evidence for verifiable autonomy and reliable task completion.
- The concept of "agency" in AI is best understood as a property of a configured system (model, harness, environment, delegator) rather than an intrinsic capability of the language model itself.
- The paper advocates for a research agenda focused on "justified delegation," meaning the expansion of an agent's action scope should be evidenced by provenance, bounded authority, failure detection, safe recovery, and calibrated human control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.