Epi: Towards a Persistent, Self-Improving Agent for Autonomous Research

[arXiv] Yue Huang, Zichen Chen and other authors have published a paper on arXiv introducing Épi, an autonomous research agent that can sustain research across model sessions and improve its methods by accumulating experience. According to the paper, Épi combines a persistent research record, hierarchical memory, branch exploration and independent review, with experimental results evaluated by a fixed verifier; it achieved a total of 120 record results, including 20 numerical bound records across 12 categories and 100 construction records across 8 categories, among which the n-queens constant interval was narrowed by about 84 times, and the rigorous interval for the dimension of the fourth Feigenbaum attractor was narrowed by more than 3.6 million times. The authors state that Épi also provided a computer-assisted proof of the Tuza conjecture on split graphs under specific conditions. Its lightweight configuration Épi-turbo, on 30 FrontierOR tasks, under the same underlying model and reasoning effort, achieved a benchmark total score 5.4% higher at 64% of Codex's model cost.