Foundation models and intelligent decision-making: Progress, challenges, and perspectives

Abstract

Intelligent decision-making (IDM) is a cornerstone of artificial intelligence (AI) designed to automate or augment decision processes. Modern IDM paradigms integrate advanced frameworks to enable intelligent agents to make effective and adaptive choices and decompose complex tasks into manageable steps, such as AI agents and high-level reinforcement learning. Recent advances in multimodal foundation-based approaches unify diverse input modalities-such as vision, language, and sensory data-into a cohesive decision-making process. Foundation models (FMs) have become pivotal in science and industry, transforming decision-making and research capabilities. Their large-scale, multimodal data-processing abilities foster adaptability and interdisciplinary breakthroughs across fields such as healthcare, life sciences, and education. This survey examines IDM's evolution, advanced paradigms with FMs and their transformative impact on decision-making across diverse scientific and industrial domains, highlighting the challenges and opportunities in building efficient, adaptive, and ethical decision systems.

Keywords: agent; artificial intelligence; foundation models; intelligent decision-making; large language model.

Graphical abstract

Figure 1

The development history of intelligent decision-making Rule-based decision support system achieves decision support through the rule base and the fact base and is suitable for scenarios driven by clear rules. Data-driven decision support system, which combines technologies such as neural networks and decision trees, is applied to projects like Deep Blue and AlphaGo and has surpassed humans in multiple fields. Decision-making with foundation models, an emerging technology in data-driven decision-making, achieves decision optimization by utilizing large models (such as the GPT series and LLaMA) through steps like demonstration data collection, data annotation, and reward model training, with application cases including OpenVLA, RoboGen, and others.

Figure 2

Overview and development of foundation models

Figure 3

The critical roles that FM can play for intelligent decision-making

Figure 4

Different paradigms of reinforcement learning (A) Offline RL is a method that learns optimal or near-optimal policies using only existing historical data with little or no online interaction. (B) Meta RL enables agents to have the ability of “learning to learn” and quickly adapt when facing new tasks. (C) Hierarchical RL simplifies the learning process by introducing a hierarchical structure to decompose complex tasks into high-level “meta-actions” or “sub-tasks” and low-level specific execution policies. (D) Multi-agent RL is a reinforcement learning paradigm that studies multiple agents learning optimal policies through collaboration or competition in an environment.

Figure 5

Advanced intelligent decision-making paradigms with foundation models (A) Vision-Language-Action (VLA) integrates the hierarchical reasoning of LLM and the perceptual capabilities of vision models to decompose high-level tasks into executable subtasks, addressing computational and data bottlenecks in complex decision-making tasks caused by scenario diversity and partial observability. (B) Learning from Videos (LfV) leverages large-scale, inexpensive online video datasets to transform raw video into structured transition trajectories for policy learning, enabling the training of generalist decision-making agents by exploiting diverse, real-world contextual information from noisy, uncurated videos. (C) Generative Simulation (GenSim) utilizes simulation environments alongside components such as prompt-guided task proposal modules and agent modules to create diverse decision-making scenarios and optimize adaptive policies, thus minimizing dependence on expensive real-world data for training agents in complex tasks like robotics and autonomous systems.

Figure 6

Examples of key technologies for intelligent decision-making The large-scale IDM system can be viewed as an “Agent,” whose core technical framework can be broadly divided into four modules: Memory, Planning, Tools, and Action. The Agent acquires information about the external environment through the Perception module and stores it in both short-term and long-term memory. The Planning module generates decision plans based on the current environmental state and historical information, while the Tools module provides external resources, such as computation and search functions, to assist in the decision-making process. Finally, the Agent executes the corresponding actions based on the planning results. This process affects the state of the environment, and through environmental feedback it further adjusts the Agent’s behavior strategy.

Figure 7

Foundation-model-driven intelligent decision-making in multidisciplinary sciences featuring the core roles of FM in intelligent decision-making, supported by diverse data types for training, and showcasing applications in key scientific fields such as information science, mathematical science, life science, healthcare, dentistry, urban science, agricultural science, economic science, and educational science

Figure 8

Risks and challenges in LLM agent The left of the three subfigures exhibits the algorithm-level attack and mitigation from perspectives of intrinsic vulnerabilities and interactive environment, e.g., jailbreak, backdoor, and model interrogation. The middle panel describes the application-level privacy and risk from the viewpoint of different disciplinary fields, e.g., membership inference attack and information cocoon. The right panel presents system-level LLM trustworthiness and robustness from the aspects of PoT and intelligent decision-making environment, e.g., content-conflicting hallucination and fact-conflicting hallucination.