Reproduce “Atropos: Improving Cost-Benefit Trade-off of LLM-based Agents under Self-Consistency with Early Termination and Model Hotswap” 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
Naryeong Kim, Shin Yoo
Atropos predicts whether an agent trajectory will succeed using a Semantic Flow Graph of tool-call steps and a graph convolutional network. It can stop predicted failures or replay context on a stronger model. On 428 AutoFL tasks, midpoint parallel hotswapping achieved 171 top-1 localizations for estimated $69.01, versus 230 for $288.80 with GPT-4o alone; this is one tested configuration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.