Reproduce “Agentic Large Language Models” 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
Max van Duijn, Michiel van der Meer, Aske Plaat, Niki van Stein
The book explains how large language models become agents when a software harness lets them act, observe outcomes, and use feedback. It organizes agentic systems around reasoning, acting, and interacting, and surveys practical examples and research across these areas. The authors present interaction as a possible source of training data, not evidence of reliable self-improvement or dependable real-world performance.
Key insights
- Agent behavior depends on both model capability and harness design; the harness can enable useful actions while introducing its own failure points.
- Reasoning prompts, reflection, memory, retrieval, and tools can help on some tasks, but their effects vary, and external verification is often more dependable than asking a model to check itself.
- Plausible behavior in social simulations does not establish human-like understanding: coordination remains difficult, and apparent norms may reflect pretraining data rather than emerge from interaction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.