Épi: A Persistent Agent for Long-Horizon Autonomous Research

Introduction

Large language model agents can often produce a plausible line of reasoning in one session, but they tend to lose failed experiments, useful tool strategies, and unresolved questions when that session ends. Épi, introduced by BakeLab, addresses this continuity problem by treating autonomous research as a long-running process. Its contribution is less about presenting a new foundation model than about building an operating loop for research: preserve history, run experiments, branch into alternatives, and submit results to an external verifier.

Key points

• Persistent research history: Épi carries prior investigations into later sessions, allowing new work to build on earlier attempts instead of restarting from the original prompt.
• Hierarchical memory: Different layers of memory retain background context, local experiments, and accumulated methodological experience.
• Branching and review: The agent can explore multiple hypotheses, return to abandoned directions, and use independent review to challenge proposed results.
• Mathematical output: The paper reports 120 record-setting results: 20 numerical-bound records across 12 families and 100 construction records across eight families. Examples include an approximately 84-fold narrower interval for the n-queens constant and a rigorous interval for the quartic Feigenbaum attractor dimension that is more than 3.6 million times narrower.
• Beyond numerical search: Épi produced a computer-assisted proof for a stated class of split graphs related to Tuza’s conjecture. It also generated exact dual certificates showing that the Cohn–Elkies linear-programming bound cannot be sharp in dimensions 10 and 11.
• Algorithmic coding: Épi-turbo applies the same research style to executable solvers for vehicle routing, scheduling, and production planning. On 30 FrontierOR tasks, it achieved a 5.4% higher total benchmark score than Codex while using 64% of its model cost.

Why it matters

Épi points to an alternative path for improving agents: not only increasing model size or context length, but giving the system a durable workspace and a structured way to learn from outcomes. In mathematics and scientific computing, progress frequently depends on many discarded approaches. A persistent research tree can preserve those lessons, reduce duplicated effort, and make it easier to resume promising directions.

The claims should still be read with care. “Record-setting” performance depends on the precise problem definitions, baselines, and verification procedures, while computer-assisted proofs are bounded by their formal assumptions and checking pipeline. The system is therefore better understood as a research collaborator than as an autonomous replacement for researchers. It can search broadly, run computations, and propose constructions; people still need to inspect assumptions, interpret the results, and decide what constitutes meaningful knowledge. If the approach generalizes, long-horizon task management and reliable verification may become as important to research agents as single-turn reasoning quality.

Source: Hugging Face Daily Papers