IDEAgent: A Multi-Agent Framework for Scientific Idea Generation via Quality-Diversity Search
Existing large language models in automated scientific discovery often fall into the trap of optimizing either idea quality or diversity in isolation, producing concepts that are either too similar or too trivial. To address this, we introduce IDEAgent, a multi-agent framework that casts scientific idea generation as a Quality-Diversity (QD) search problem. IDEAgent manages the evolutionary trajectory of ideas through a genealogy-based structure, employs multi-objective feedback-driven repair and refinement to enhance quality, and leverages lightweight sequence memory with explicit comparison mechanisms to ensure diversity. The research team developed a joint metric called Yield for systematic evaluation. In large-scale experiments spanning 8 domains and 32 topics in computer science, IDEAgent surpasses the best baseline by 3.89× on the Yield metric and achieves non-zero Yield across 8× more topics. Ablation studies confirm that the repair and refinement mechanisms are critical for building logical rigor. The project is open-sourced.
Background and Context
Large language models have significantly accelerated the generation of scientific ideas in automated discovery, yet they suffer from a fundamental architectural limitation. Current systems typically optimize for idea quality and diversity as independent, siloed objectives. This decoupled approach creates a critical trade-off: models either produce concepts that are highly similar to existing literature, lacking novelty, or generate a high volume of trivial, scientifically unsound hypotheses that lack rigorous justification. The research community has identified this inability to balance innovation with validity as a major bottleneck in AI-assisted science.
To address this gap, the research team introduces IDEAgent, a multi-agent framework that redefines scientific idea generation as a Quality-Diversity (QD) search problem. Unlike traditional generative models that treat each idea as an isolated output, IDEAgent manages the evolutionary trajectory of ideas through a genealogy-based structure. This lineage approach allows the system to track the development of concepts over time, ensuring that the search process simultaneously maximizes the scientific value of individual ideas while maintaining a broad coverage of the conceptual space. The framework is designed to solve the long-standing challenge of achieving both high rigor and high diversity in automated research ideation.
Deep Analysis
The technical architecture of IDEAgent relies on a specialized multi-agent collaboration mechanism that separates the concerns of quality enhancement and diversity maintenance. For quality optimization, the framework employs a multi-objective feedback loop dedicated to repairing and refining generated ideas. This process goes beyond simple linguistic polishing; it involves rigorous logical validation and scientific plausibility checks. By actively identifying and correcting flawed assumptions, the system ensures that every final output possesses a high degree of academic integrity and logical coherence, effectively filtering out noise before it enters the final repository.
To ensure diversity, IDEAgent integrates a lightweight sequence memory module coupled with an explicit comparison strategy. The system continuously compares newly generated ideas against historical records, including completed ideas, ancestral nodes in the genealogy tree, and previously rejected proposals. This explicit comparison mechanism prevents the agents from converging on local optima or repeating previous outputs. By forcing the agents to explore under-developed regions of the concept space, the framework guarantees that the generated ideas remain distinct and non-obvious, thereby preserving the novelty essential for breakthrough research.
The evaluation of this joint search strategy is facilitated by a novel metric called Yield. Yield is defined as the maximum size of a set of distinct ideas that all meet a predefined quality threshold. This metric elegantly balances the dual goals of quantity and quality, avoiding the biases inherent in single-metric evaluations. The experimental setup was comprehensive, spanning eight distinct sub-domains within computer science and covering 32 specific research topics. This broad scope ensures that the results are representative of the diverse challenges faced in automated scientific discovery.
Industry Impact
The performance of IDEAgent in large-scale experiments demonstrates a substantial leap forward in automated ideation capabilities. The framework surpassed the best existing baseline by a factor of 3.89 on the Yield metric. More significantly, IDEAgent achieved non-zero Yield values across eight times as many topics as the baseline models. This result highlights the framework's superior generalization ability and its effectiveness across a wide variety of research contexts. The ablation studies further confirmed that the repair and refinement mechanisms are critical for building logical rigor, proving that quality enhancements do not come at the cost of non-obviousness or novelty.
For the open-source community and industrial researchers, IDEAgent offers a new methodological paradigm for AI-assisted science. By breaking the traditional dichotomy between quality and diversity, the framework provides a robust tool for rapidly screening and generating high-quality, diverse research hypotheses. This capability is particularly valuable in emerging interdisciplinary fields that require rapid iteration and broad exploration. The open-sourcing of the IDEAgent framework encourages further innovation in quality-diversity search techniques, potentially accelerating the pace of scientific discovery across multiple domains.
Outlook
The introduction of IDEAgent marks a significant milestone in the evolution of AI for Science. By systematically handling the multiple constraints of idea generation, the framework serves as a crucial infrastructure for future research tools. The genealogy-based management and multi-agent collaboration mechanisms offer valuable insights for other complex decision-making problems that require balancing exploration and exploitation. As large language models become increasingly integrated into the scientific workflow, frameworks like IDEAgent will play a pivotal role in helping human scientists navigate the vast ocean of knowledge.
Future developments in this area are likely to focus on expanding the scope of applicable domains and refining the efficiency of the multi-agent interactions. The success of IDEAgent suggests that treating idea generation as a structured search problem, rather than a purely generative task, is a promising direction. Continued research into quality-diversity search algorithms could lead to even more sophisticated tools that not only generate ideas but also assist in experimental design and hypothesis validation. This shift towards more rigorous and diverse automated discovery processes will ultimately enhance the productivity and creativity of the global scientific community.
The open-source nature of IDEAgent invites collaboration and extension by researchers worldwide. As the framework matures, it is expected to become a standard component in the toolkit of computational scientists. The ability to systematically explore the space of scientific possibilities with both rigor and creativity will likely unlock new avenues for discovery that were previously inaccessible to purely human or purely algorithmic approaches. This represents a transformative step towards a more efficient and innovative scientific enterprise.