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

Epi. Towards a Persistent, Self-Improving Agent for Autonomous Research. By Yue Huang and eighteen others, from Bake AI, the University of Notre Dame, and eleven other institutions. Published on the eighteenth of October. Its code repository is public but slight: one source file.

Research rarely fits inside one sitting. A real investigation has to continue across many model sessions, picking up where it left off, and its methods ought to get better as experience accumulates rather than starting from nothing each time. It also has to be able to try one approach, set it aside, and return to it. The goal here is an autonomous research agent that holds on to all of that.

Epi is that agent. It sustains investigations across model sessions and evolves its methods through accumulated experience. Four ingredients carry the design: a persistent research history, hierarchical memory, branching, and independent review. Together they serve three purposes. They maintain research continuity. They let the agent learn from its experiments. And they let it explore alternative approaches, then revisit them later.

Experiments run across session boundaries, so a single line of work is not cut off when one session ends. A fixed verifier evaluates their outputs, which means that what counts as a result is settled by something outside the agent rather than by the agent's own judgment. The same research process is then carried over to algorithmic coding through Epi-turbo, a lightweight configuration that develops executable solvers for tasks including vehicle routing, scheduling, and production planning.

In mathematics, Epi produced one hundred and twenty record-setting results. Twenty of them are numerical bound records, spread across twelve families. The other one hundred are construction records, across eight families. Two examples stand out. One is an interval for the n-queens constant that is roughly eighty-four times narrower. The other is a rigorous interval for the quartic Feigenbaum attractor dimension that is more than three and a half million times narrower.

Exact dual certificates establish that the Cohn-Elkies linear-programming bound cannot be sharp in dimensions ten and eleven. And beyond numerical improvements, Epi gives a computer-assisted proof of Tuza's conjecture for a class of split graphs: those whose clique part has eight vertices, with at most three triangle-active neighborhood types, and with arbitrary multiplicities.

On the coding side, across thirty FrontierOR tasks, Epi-turbo achieves a total benchmark score five point four percent higher than Codex. It does so at sixty-four percent of Codex's model cost, using the same underlying model and the same reasoning effort, so the difference lies in the research process wrapped around the model.

Those records are not presented as answers delivered from a black box. Epi is designed to support human understanding, tying each computational result back to explicit methods and mathematical arguments that researchers can examine and then build on themselves. The machine extends the search, and the people reading it are meant to be able to follow every step.