This paper introduces ScientistTwo, a multi-agent framework intended to autonomously conduct scientific research from problem formulation through experimentation, ablation studies, and simulated peer review. It evaluates the system on tasks associated with ICLR, ICML, and NeurIPS papers, reporting automated generation of research papers and executable codebases.
AnyJev is an open-source GitHub project that uses a single LLM prefill to produce typed decisions and calibrated probabilities without generation, parsing, or fine-tuning. It includes methods for reducing answer-order sensitivity and improving calibration with a small number of labels.
An academic paper examining the emergence of large language model agents. The accompanying figure traces developments from early transformer architectures and language models to agent behaviors such as planning, negotiation, deception, and theory of mind.
