Large Language Models for Scientific Idea Generation: A Creativity-Centered Survey

Abstract: Scientific idea generation is central to discovery, requiring the joint satisfaction of novelty and scientific soundness. Unlike standard reasoning or general creative generation, scientific ideation is inherently open-ended and multi-objective, making its automation particularly challenging. Recent advances in LLMs have enabled the generation of coherent and plausible scientific ideas, yet the nature and limits of their creative capabilities remain poorly understood. This survey provides a structured synthesis of methods for LLM-driven scientific ideation, focusing on how different approaches trade off novelty and scientific validity. We organize existing methods into five complementary families: External knowledge augmentation, Prompt-based distributional steering, Inference-time scaling, Multi-agent collaboration, and Parameter-level adaptation. To interpret their contributions, we adopt two complementary creativity frameworks: Boden taxonomy to characterize the expected level of creative novelty, and Rhodes 4Ps framework to analyze the aspects or sources of creativity emphasized by each method. By aligning methodological developments with cognitive creativity frameworks, this survey clarifies the evaluation landscape and identifies key challenges and directions for reliable and systematic LLM-based scientific discovery.

• The paper introduces five distinct methods for leveraging LLMs in scientific idea generation, aligning them with cognitive creativity frameworks.
• It demonstrates how knowledge augmentation and prompt-driven techniques improve idea novelty while mitigating hallucinations.
• It emphasizes multi-agent collaboration and inference-time scaling as innovative strategies to enhance interdisciplinary ideation and overcome computational challenges.

"LLMs for Scientific Idea Generation: A Creativity-Centered Survey"

Introduction

The paper "LLMs for Scientific Idea Generation: A Creativity-Centered Survey" (2511.07448) explores the intersection of LLMs with scientific ideation, focusing on strategies that balance novelty and scientific validity. Scientific ideation is characterized by its open-ended and multi-objective nature, demanding novel yet plausible hypotheses. The authors categorize current methodologies into five families, aligning them with cognitive creativity frameworks to clarify the evaluation landscape and identify key challenges for LLM-driven scientific discovery.

Knowledge Augmentation

The paper emphasizes the role of Knowledge Augmentation to ground LLMs in external data, reducing hallucinations and enhancing combinatorial creativity by recombining known elements. This is achieved through semantic retrieval systems and knowledge graphs (KGs), as depicted in the pipeline overview (Figure 1a).

Figure 1: Overview of the training-free methods and requirements for scientific idea generation using LLMs.

Semantic retrieval incorporates relevant literature into model prompts, while relational retrieval uses KGs to explore cross-domain links, fostering interdisciplinarity. However, the risk of conservatism arises if models over-rely on retrieved content, necessitating hybrid strategies that incorporate exploration.

Prompt-Driven Creativity

Prompt-driven techniques are explored (Figure 1b), where different styles of prompts steer LLMs towards originality. These include structured prompts, role priming, and adversarial queries. Persona priming aligns model outputs with pre-defined roles, and constraint-based prompts challenge existing assumptions, leading to innovative solutions. Despite their potential, these methods face limitations, such as alignment systems restricting output diversity to maintain safety and helpfulness.

Inference-Time Scaling

Inference-time scaling, illustrated in Figure 1c, allows models to expand their hypothesis space dynamically. Techniques like sequential, parallel, and tree-based searches facilitate exploration beyond single-decoding paths, engaging methods like Self-Refine and beam search. Although test-time scaling significantly enhances creativity, it poses computational challenges and often relies on sound evaluation signals to ensure idea plausibility.

Figure 2: Comparison of sequential, parallel, and tree-based search mechanisms for the ideation task of reducing LLM hallucinations in medical question answering.

Multi-Agent Collaboration

The survey highlights Multi-Agent Collaboration (Figure 1f) as a means to simulate collaborative dynamics akin to human scientific teams. These systems enable debate, critique, and synthesis in a shared context, allowing LLMs to generate more creative and scientifically valid ideas through emergent interactions. By doing so, they advance beyond single-agent capabilities to address the social aspect of creativity.

Parameter-Level Adaptation

Finally, the role of Parameter-Level Adaptation in enhancing LLM capabilities is discussed. Techniques such as supervised fine-tuning and reinforcement learning internalize search strategies and domain knowledge, fostering ideation closer to human-level creativity. This approach shifts the emphasis from external processes to intrinsic model qualities, striving for transformative creativity.

Figure 3: Conceptual mapping of LLM-driven scientific ideation methods to their primary source of creativity, inspired by Rhodes’ 4Ps framework.

Conclusion

The paper provides a comprehensive overview of approaches for leveraging LLMs in the creative scientific process. While significant progress is noted in balancing novelty with value, challenges remain in achieving transformational creativity. Future developments may focus on enhancing model architectures and creating robust, standardized evaluation frameworks to better measure and validate scientific ideas. With continual refinements, LLMs hold promise as invaluable tools in scientific discovery, potentially collaborating on innovative breakthroughs.