Can AI Dream Up Truly New Science?

Author: Denis Avetisyan

New research reveals that complex AI workflows are capable of generating research plans assessed by experts as both novel and realistically achievable.

Multi-step large language model pipelines, employing decomposition and cross-domain ideation, show promise in addressing concerns around plagiarism and augmenting scientific discovery.

Despite growing integration of artificial intelligence into scientific research, concerns remain regarding the originality of AI-generated ideas and the potential for subtle plagiarism. This study, ‘Evaluating Novelty in AI-Generated Research Plans Using Multi-Workflow LLM Pipelines’, investigates whether advanced, multi-step AI workflows-employing techniques like iterative refinement, evolutionary search, and recursive decomposition-can overcome these limitations and produce genuinely novel research proposals. Our benchmarking of five distinct reasoning architectures reveals that decomposition-based and long-context workflows consistently generate ideas rated as highly novel and feasible by human experts, exceeding the performance of reflection-based approaches. These findings suggest that carefully designed agentic workflows hold promise for enhancing AI’s contribution to scientific discovery-but can these methods consistently yield impactful research across diverse domains?

The Limits of Intuition: Navigating the Complexity of Modern Science

Scientific advancement has long been fueled by the insights of experts and fortunate accidents, yet this reliance on intuition and serendipity presents inherent limitations. While experienced researchers possess valuable domain knowledge, this expertise can also introduce bias, subtly steering investigations toward familiar avenues and potentially overlooking genuinely novel approaches. Furthermore, these processes are often remarkably slow; breakthroughs aren’t consistently predictable, and the pace of discovery can’t keep up with the ever-increasing volume of scientific data. The subjective nature of interpreting evidence, coupled with the inherent unpredictability of chance encounters, suggests that traditional methods, though historically successful, are becoming increasingly insufficient for navigating the complexities of modern scientific challenges and capitalizing on the potential for rapid, data-driven innovation.

The sheer volume of published scientific research is now expanding at a rate that far outpaces any individual’s capacity for comprehensive review. Each year, millions of new papers flood databases, creating a knowledge landscape too vast for even the most dedicated expert to fully navigate. This isn’t simply a matter of time constraints; the combinatorial explosion of information means that novel connections and potentially groundbreaking syntheses are increasingly likely to be overlooked. While human researchers excel at focused inquiry, the ability to identify relevant insights scattered across this immense corpus is becoming a critical bottleneck, hindering discovery and demanding innovative approaches to knowledge aggregation and analysis.

Scientific advancement increasingly demands the integration of knowledge from diverse and often unconnected disciplines, yet the sheer volume of published research presents a formidable barrier to this synthesis. Breakthroughs in fields like systems biology, climate science, and materials discovery rarely emerge from single areas of expertise; instead, they require researchers to identify and combine insights scattered across numerous publications, databases, and experimental results. However, the exponential growth of this information overwhelms human capacity for comprehensive review, leading to overlooked connections and potentially stifling innovation. This limitation is not merely a matter of time or effort; it represents a fundamental bottleneck in the scientific process, hindering the translation of accumulated knowledge into genuinely novel discoveries and solutions.

Agentic Workflows: A New Paradigm for Idea Generation

Agentic workflows represent a paradigm shift in idea generation through the iterative application of multi-step reasoning, internal debate, and iterative mutation of concepts. This process moves beyond single-prompt LLM responses by enabling a sequence of operations where an LLM first decomposes a research question, then generates hypotheses, critically evaluates those hypotheses through self-debate – identifying weaknesses and potential counterarguments – and finally, mutates promising concepts by exploring variations and edge cases. This cyclical process of reasoning, evaluation, and modification allows the system to navigate the research landscape more thoroughly and generate novel insights that would be difficult to achieve with a static, single-pass approach. The workflow is designed to simulate aspects of human scientific inquiry, facilitating a more robust and comprehensive exploration of complex topics.

The agentic workflow relies on Large Language Models (LLMs) as its core processing unit, leveraging their capacity to analyze and integrate information from diverse sources. LLMs facilitate the handling of complex data by converting natural language inputs into structured representations, enabling the identification of relevant patterns and relationships. This capability extends to synthesizing information – combining data points to formulate novel insights and hypotheses. The models’ parameter scale and training datasets allow for generalization across a broad range of topics, supporting exploration of the research landscape without requiring pre-defined knowledge boundaries. Furthermore, LLMs provide a probabilistic framework for evaluating the validity and relevance of generated content, crucial for iterative refinement within the workflow.

Decomposition of complex problems into discrete subquestions is a core component of this workflow, enabling focused investigation and the generation of testable hypotheses. This process involves breaking down an overarching research question into a series of smaller, independent queries that can be addressed individually by the Large Language Model (LLM). By systematically addressing these subquestions, the LLM can build a more comprehensive understanding of the problem space and identify potential avenues for exploration that might be obscured when considering the problem as a whole. The resulting subquestion responses then serve as building blocks for formulating and evaluating hypotheses, increasing the efficiency and depth of the research process.

Workflow Validation: Assessing Performance Across Diverse Approaches

The evaluation encompassed three agentic workflows – Gemini 3 Pro, Sakana AI v2, and Google Co-Scientist – each engineered with a unique approach to idea generation. Gemini 3 Pro was designed to enhance existing proposals through long-context analysis and identification of informational gaps. Sakana AI v2 focused on generating novel ideas via a mutation-based process guided by a custom fitness function. Google Co-Scientist utilized an adversarial vetting system, simulating a collaborative scientific review process with specialized agents to refine and rigorously evaluate proposed concepts. These workflows were then comparatively assessed based on metrics including novelty and feasibility, as determined by expert evaluations.

Gemini 3 Pro and Sakana AI v2 employ differing strategies for idea generation. Gemini 3 Pro utilizes long-context modeling, processing extensive input to identify deficiencies in existing proposals, and multimodal learning to integrate information from various sources for comprehensive refinement. Conversely, Sakana AI v2 focuses on generating novel concepts through a process of mutation, systematically altering initial ideas, and evaluating them using a specifically designed fitness function that prioritizes originality and relevance. This function guides the mutation process, favoring variations that exhibit desired characteristics and discarding those that do not, thereby driving the evolution of innovative proposals.

Google Co-Scientist utilizes an adversarial vetting process to enhance idea refinement and ensure methodological rigor. This workflow simulates a laboratory meeting environment populated by specialized agents, each assigned a specific role – such as experimental design, statistical analysis, or domain expertise. These agents actively challenge proposed ideas, identify potential flaws in reasoning or methodology, and request further justification or data. The iterative process of challenge and response aims to expose weaknesses and strengthen the overall validity of the generated concepts, mimicking the critical evaluation inherent in peer review and collaborative research.

Expert evaluations of generated ideas revealed a significant performance difference between workflow types. Decomposition-based and long-context workflows, exemplified by Gemini 3 Pro, achieved a mean novelty score of 4.17 on a 5-point scale. In contrast, reflection-based approaches yielded a substantially lower mean novelty score of 2.33/5. This indicates a clear advantage for workflows prioritizing the breakdown of complex problems and utilization of extended contextual information in generating novel ideas, as compared to those relying on iterative self-critique and refinement.

Expanding the Horizon: The Broad Impact of Automated Ideation

Recent investigations reveal that agentic workflows – systems designed to autonomously explore and develop ideas – have been successfully implemented across diverse scientific fields. Applications extend from the rapidly evolving landscape of AI and technology, where these workflows demonstrate high potential, to the intricacies of chemistry and biotechnology, as well as critical areas like climate and environmental research and the demands of modern industry and manufacturing. This broad applicability suggests a fundamental shift in how research can be approached, enabling automated exploration of complex problems and the generation of novel insights in traditionally disparate fields. The successful integration into such varied domains underscores the versatility of this technology and its capacity to become a powerful tool for scientific advancement across multiple disciplines.

Analysis across multiple scientific disciplines revealed significant variation in the novelty of ideas generated by the agentic workflows. The AI/Tech domain consistently produced the most novel concepts, achieving a mean novelty score of 4.00, suggesting a particularly fertile ground for automated idea generation within that field. Conversely, the Chemistry/Biotech domain exhibited the lowest mean novelty score of 3.20, potentially indicating that initial concepts within this discipline require more nuanced refinement or that the current workflows are better suited to the more abstract challenges present in AI/Tech. This difference underscores the importance of domain-specific optimization and adaptation of these automated workflows to maximize their impact on scientific innovation.

The successful deployment of these agentic workflows across diverse scientific fields – from artificial intelligence to biotechnology, and environmental science to manufacturing – demonstrates a remarkable capacity for generalization. This adaptability isn’t merely a coincidental outcome; it suggests the underlying principles guiding idea generation are fundamental and transcend specific disciplinary boundaries. The approach doesn’t require substantial modification when shifting between domains, indicating a robust framework capable of identifying and exploring novel concepts regardless of the subject matter. This versatility promises to unlock innovation in areas previously constrained by the limitations of traditional, domain-specific methodologies, and positions the workflows as a broadly applicable tool for accelerating scientific progress.

The automation of initial idea generation represents a paradigm shift in scientific workflows, promising to dramatically accelerate the pace of discovery and innovation. These agentic systems don’t replace researchers, but rather function as powerful collaborators, rapidly exploring vast solution spaces and surfacing novel concepts that might otherwise remain unexplored. By handling the often time-consuming preliminary stages – brainstorming, literature review, and hypothesis formulation – scientists are freed to focus on critical analysis, experimental design, and the refinement of promising leads. This increased efficiency isn’t limited to a single discipline; the demonstrated versatility across diverse fields, from artificial intelligence to biotechnology, suggests a broadly applicable tool capable of unlocking breakthroughs across the scientific landscape and fostering a more dynamic, iterative approach to research.

Future Directions: Charting a Course for Robust and Autonomous Discovery

Maintaining the integrity of scientific output requires careful attention to originality as AI increasingly contributes to content creation. The potential for unintentional plagiarism arises from the models’ training on vast datasets of existing literature, necessitating robust mechanisms to verify the novelty of generated text and ideas. Current strategies focus on comparing outputs against established databases, but more sophisticated techniques are needed to detect subtle forms of paraphrasing or recombination of existing concepts. Addressing this challenge isn’t merely about avoiding academic misconduct; it’s fundamental to ensuring that AI serves as a tool for genuine discovery, building upon-rather than replicating-prior knowledge and fostering innovation within the scientific community.

Accurately gauging the feasibility of scientifically novel ideas remains a significant challenge, demanding further investigation into robust assessment methodologies. Current approaches often struggle to differentiate between genuinely promising, yet complex, concepts and those that are impractical given existing resources or technological limitations. Consequently, research efforts should prioritize developing more nuanced metrics that evaluate not just technical hurdles, but also factors such as resource availability, potential for iterative development, and alignment with established scientific principles. Refining this balance between ambitious innovation and pragmatic execution is crucial; a system that consistently favors only incremental advancements risks stifling disruptive breakthroughs, while one that overlooks practical considerations may lead to a proliferation of unachievable concepts. Ultimately, improved feasibility assessment will enable a more efficient allocation of research resources and accelerate the translation of theoretical ideas into tangible scientific progress.

Analysis revealed a surprisingly weak correlation of 0.23 between the novelty of a proposed scientific idea and its practical feasibility. This finding challenges the conventional assumption that truly innovative concepts are inherently more difficult to implement. The data suggests that while groundbreaking ideas aren’t necessarily easy to realize, their novelty doesn’t automatically preclude them from being feasible. Consequently, researchers and innovation systems shouldn’t prematurely dismiss potentially transformative concepts simply because they deviate significantly from existing knowledge; a rigorous assessment of practicality independent of novelty is crucial for fostering genuine scientific advancement. This decoupling of innovation and implementation difficulty opens avenues for exploring a broader range of ideas and potentially accelerating the pace of discovery.

The progression towards fully autonomous scientific discovery hinges on equipping agentic workflows with the capacity for comprehensive experimental design and robust data analysis. Currently, these systems excel at generating hypotheses, but translating those ideas into actionable experiments – defining controls, selecting appropriate methodologies, and interpreting resulting data – remains a significant hurdle. Integrating these capabilities would allow an agent to not only propose novel research directions, but also independently validate or refute them through iterative experimentation and analysis. This self-sufficient cycle, driven by algorithms capable of both creativity and critical evaluation, promises to accelerate the pace of scientific advancement by minimizing human intervention and maximizing the efficiency of the discovery process. Such systems envision a future where agents can independently formulate research questions, design and execute experiments, and disseminate findings, ultimately reshaping the landscape of scientific inquiry.

The pursuit of novelty, as demonstrated by this work on agentic workflows, echoes a sentiment articulated by Alan Turing: “Sometimes people who are unaware of their own potential are driven by external forces.” The research highlights how decomposed, multi-step Large Language Model pipelines, capable of cross-domain ideation, move beyond simple imitation. This isn’t merely about generating text; it’s about constructing novel research plans judged feasible by human experts. The system’s ability to navigate complex problem spaces, breaking down tasks into manageable components, suggests an overcoming of limitations – a realization of potential previously unseen. This directly addresses the challenge of plagiarism, as genuine novelty emerges from the process itself, not from replication.

Where Do We Go From Here?

The demonstrated capacity of these agentic workflows to generate ostensibly novel research plans does not, of course, resolve the underlying problem. A system that requires novelty assessment has already failed to truly innovate. The metrics employed-human ratings of feasibility and originality-remain inherently subjective, and thus, susceptible to the very biases the pursuit of automated ideation seeks to circumvent. The question is not whether an AI can mimic discovery, but whether it can operate beyond the constraints of pre-existing knowledge-a feat no assessment can accurately gauge.

Future work must therefore shift from evaluating output to understanding process. Deeper analysis of the decomposition strategies employed by these large language models-the ways in which problems are broken down and recombined-may reveal not genuine creativity, but simply a more efficient traversal of existing conceptual space. The focus should be on minimizing the reliance on initial prompts, and maximizing the system’s capacity for self-directed exploration-even if that exploration yields only further questions.

Ultimately, the true measure of success will not be the quantity of ‘novel’ ideas generated, but the degree to which these systems can expose the limitations of existing frameworks-revealing not what can be known, but what remains fundamentally unknowable. Clarity, it seems, is still the most courteous goal.

Original article: https://arxiv.org/pdf/2601.09714.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/