New AI Framework Enables Automated Design of Scientific Objectives

Global: New AI Framework Enables Automated Design of Scientific Objectives

In December 2025, a multidisciplinary team of researchers released a preprint describing a novel framework aimed at automating the formulation of objective functions for scientific discovery agents. The work, titled “Scientific Autonomous Goal‑evolving Agent (SAGA),” addresses the longstanding challenge that many scientific optimization problems rely on imperfect proxy metrics.

Bi‑Level Architecture

SAGA is built on a two‑tier design. An outer loop composed of large‑language‑model (LLM) agents evaluates the outcomes of optimization runs, generates revised scientific goals, and translates them into computable scoring functions. Meanwhile, an inner loop carries out the actual solution search using the current set of objectives.

Outer Loop: Objective Evolution

The outer loop continuously monitors performance metrics, identifies shortcomings in the existing objectives, and proposes alternative formulations. By converting these proposals into executable scoring functions, the system can iteratively refine what it seeks to optimize without human intervention.

Inner Loop: Solution Optimization

Within the inner loop, conventional optimization algorithms operate under the objectives supplied by the outer loop. This separation allows the inner loop to focus on finding high‑performing candidates while the outer loop steers the direction of the search.

Demonstrated Applications

The authors illustrate SAGA’s versatility across four domains: antibiotic molecule design, inorganic material synthesis, functional DNA sequence engineering, and chemical process optimization. In each case, automated objective redesign led to measurable improvements over baseline approaches that used static objectives.

Potential Impact

By treating objective design as a learnable component, SAGA opens a pathway for more adaptive and efficient scientific discovery agents. Critics note that the approach may increase computational demand, but proponents argue that the gains in discovery speed could outweigh the costs.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.