Original Technical Article and Study: https://arxiv.org/abs/2512.15567
1.0 The State of Play: The Promise and Peril of LLMs in Science
Large Language Models (LLMs) are rapidly evolving from general-purpose tools into specialized partners in scientific inquiry. Across disciplines, they are beginning to accelerate core stages of discovery, assisting researchers with tasks ranging from literature triage and hypothesis generation to code synthesis for computational simulations. Illustrative successes, such as the ChemCrow agent that couples language reasoning with chemistry tools and the Virtual Lab for autonomous nanobody design, demonstrate the tangible potential for LLMs to become integral components of the modern research workflow.
However, this rapid integration into the scientific enterprise has outpaced the methods used to validate these powerful tools. A critical gap has emerged between the perceived capabilities of LLMs, often showcased on broad academic benchmarks, and their actual readiness for the complex, iterative, and often ambiguous reality of scientific discovery. The prevailing evaluation standards have not kept pace with real-world applications, creating a risk of misallocating resources, misplacing trust, and ultimately hindering the very progress we seek to accelerate.
This paper introduces the Scientific Discovery Evaluation (SDE) framework, a new methodology designed to bridge this evaluation gap. By grounding assessment in real-world research scenarios, SDE provides a more robust and realistic measure of an LLM's utility for science. The findings presented here, derived from applying SDE to state-of-the-art models, reveal a more nuanced and complex picture of LLM readiness, challenging current development paradigms and charting a clearer path toward creating AI that can genuinely advance scientific discovery.
--------------------------------------------------------------------------------
2.0 The Flaw in the Status Quo: Why Decontextualized Benchmarks Fall Short
Guiding the development of increasingly powerful AI systems requires a clear and accurate measurement of their capabilities. For science, this is a matter of strategic importance. The benchmarks we use to measure progress directly influence the research and engineering priorities of the entire AI community. Unfortunately, the current standard for science benchmarks—such as GPQA and MMMU—is predominantly based on a decontextualized, perception-heavy question-and-answering (Q&A) format. These benchmarks function like academic exams, testing an LLM's ability to retrieve static, isolated facts.
The core limitation of this approach is that it mistakes static knowledge recall for dynamic scientific aptitude. Real scientific discovery is not a static quiz; it is a dynamic and iterative process. It involves formulating novel questions, reasoning under conditions of imperfect evidence, and continuously refining hypotheses in response to new observations. As the source analysis aptly puts it, "just as earning straight A’s in coursework does not indicate a great researcher," perfect scores on these Q&A benchmarks do not guarantee a model's readiness to contribute to discovery.
Consequently, these decontextualized evaluations provide a flawed and incomplete measure of an LLM's suitability for integration into complex scientific workflows. They fail to capture the very competencies that define scientific reasoning, leading to an overestimation of current capabilities and obscuring the specific areas where models need to improve. To build AI that can truly partner with scientists, we must first adopt a new paradigm for evaluation.
--------------------------------------------------------------------------------
3.0 A New Paradigm: The Scientific Discovery Evaluation (SDE) Framework
The Scientific Discovery Evaluation (SDE) framework was developed as a direct response to the limitations of conventional benchmarks. Its foundational principle is to move evaluation beyond abstract quizzes and ground it in the context of real-world research scenarios. By doing so, SDE is designed to measure the discovery-relevant capabilities that are overlooked by static Q&A formats, such as iterative reasoning, hypothesis generation, and the interpretation of experimental results.
The framework employs a multi-level structure that assesses LLM performance at both granular and holistic scales, providing a comprehensive view of a model's scientific reasoning abilities.
1. Question-Level Assessment At its base, SDE anchors evaluation in concrete research projects defined by domain experts across biology, chemistry, materials science, and physics. Each of these projects, which reflect genuine scientific interest, is decomposed into modular "research scenarios"—scientifically grounded and reusable reasoning units. From these scenarios, a set of 1,125 expert-vetted questions is derived. This hierarchical design creates a "tight connection among questions, scenarios, and projects," ensuring that every question is relevant to a practical scientific task rather than an exercise in decontextualized trivia.
2. Project-Level Assessment The second layer of evaluation assesses an LLM's performance on open-ended discovery projects. In this setting, the model is placed into a closed loop that imitates an end-to-end scientific process. It must autonomously propose testable hypotheses, design and execute simulations, and interpret the resulting observations to refine its approach. By evaluating the discovery-oriented outcomes (e.g., the properties of a proposed chemical compound), this assessment measures how well a model can orchestrate a complete research cycle.
This dual-layered approach provides a far richer and more realistic picture of an LLM's capabilities. It not only measures what a model knows but also how it reasons within the dynamic context of a research pipeline, revealing critical capability gaps and failure modes that single-level Q&A benchmarks cannot capture.
--------------------------------------------------------------------------------
4.0 Key Findings from Question-Level Evaluation: A Reality Check for Frontier Models
Applying the SDE framework to state-of-the-art LLMs yields a fine-grained analysis of their true strengths and weaknesses in scientific tasks. The results from the question-level evaluation provide a crucial reality check, challenging some of the prevailing assumptions in the AI development community. Below are four critical findings that emerge from this analysis.
4.1 The Discovery Deficit: A Consistent Gap Between Quizzes and Research Tasks
When comparing LLM scores on SDE to their scores on general-science benchmarks, a consistent performance gap emerges. For example, while a frontier model¹ achieves high scores of 0.86 on GPQA-Diamond and 0.84 on MMMU, its performance on the scenario-grounded SDE benchmark is notably lower, scoring 0.75 in materials and 0.60 in physics. This "discovery deficit" holds true across top-tier models and demonstrates that proficiency on decontextualized Q&A tests does not automatically translate to mastery of the nuanced, context-dependent reasoning required for practical research tasks.
4.2 The Granularity Imperative: Unmasking Scenario-Specific Strengths and Weaknesses
The SDE framework reveals that a model's performance can vary drastically across different research scenarios, even within the same scientific domain. For instance, the same frontier model (gpt-5) achieves an impressive score of 0.85 in the chemistry scenario of retrosynthesis planning but struggles significantly with NMR structure elucidation, where it scores only 0.23. This finding proves that conventional, domain-level scores are insufficient, as they mask critical weaknesses. This level of variance suggests that deploying a single, general-purpose LLM for a multi-stage discovery workflow is a high-risk strategy; a deficit in any one constituent scenario can introduce compounding errors that derail an entire project. A fine-grained, scenario-level assessment is essential for understanding where LLMs can be reliably applied.
4.3 The Plateau Effect: Diminishing Returns from Scaling and Reasoning
Two of the primary strategies for improving LLM performance—scaling up model size and increasing test-time reasoning—appear to be yielding diminishing returns for scientific discovery tasks. The analysis shows that accuracy gains for gpt-5 between medium and high reasoning efforts were often statistically negligible (e.g., 0.74 vs. 0.75 in materials). Similarly, when the factor of reasoning is isolated, the performance improvement from gpt-4o to gpt-5 is also negligible, indicating a potential convergence for pretrained foundation models. Further evidence of this plateau is the marginal performance gain of gpt-5 over a predecessor, o3, with gpt-5 showing significantly worse performance in eight scenarios. This suggests that further progress in scientific AI may require a strategic shift away from simply making models bigger or having them "think" longer.
4.4 A Consensus of Failure: Shared Weaknesses Among Top Models
A striking observation is that top-performing models from different providers—including gpt-5, grok-4, deepseek-R1, and claude-sonnet-4.5—exhibit highly correlated accuracy profiles. They tend to succeed and fail on the same sets of difficult questions. This correlation is particularly strong in chemistry and physics, with Spearman's rank correlation coefficients greater than 0.8. This consensus of failure indicates shared, systemic weaknesses, likely inherited from similar pre-training data and objectives. This finding has significant practical implications, as it suggests that naive ensemble strategies, such as majority voting across different models, will offer limited benefit on the hardest scientific problems.
These question-level insights provide a detailed map of current model capabilities, but to understand how these skills translate into practice, we must turn to the more dynamic evaluation at the project level.
--------------------------------------------------------------------------------
5.0 From Theory to Practice: Insights from End-to-End Discovery Projects
Scientific discovery is an iterative process, not a single-turn interaction. To truly assess an LLM's readiness for science, it is essential to move beyond static Q&A and evaluate its performance within the closed loop of hypothesis, experiment, and observation. The project-level evaluation within the SDE framework does just that, revealing how question-level competencies translate—or fail to translate—to practical, end-to-end research workflows. The results paint a complex picture where high scores do not always predict success, and surprising capabilities can emerge.
5.1 Beyond Rote Knowledge: The Critical Role of Serendipitous Exploration
A surprising disconnect often exists between a model's performance on granular questions and its success in an open-ended project. The optimization of Transition Metal Complexes (TMCs) provides a key example. Despite demonstrating low proficiency scores on question scenarios related to TMC properties, models like gpt-5 and deepseek-R1 showed excellent efficiency in the project, rapidly identifying optimal candidates. This presents a fascinating paradox: the models failed the "textbook" questions about TMC properties but excelled at the "research" task of finding novel TMCs. It suggests that for certain optimization problems, an LLM's ability to effectively navigate a vast hypothesis space is more valuable than its explicit domain knowledge.
5.2 The Fragility of Execution: When High Scores Don't Translate to Success
Conversely, high question-level scores do not guarantee project success, particularly in tasks requiring rigorous, long-horizon planning. In the retrosynthesis project, top models scored well on related questions but struggled to generate valid multi-step synthesis routes, frequently failing molecule or reaction validity checks. This fragility of execution prevented them from outperforming traditional models. Notably, gpt-4o, a relatively older model, outperformed its more advanced successors on this project. This counterintuitive result suggests that current reasoning enhancements, while beneficial for some tasks, may be misaligned with problems that demand strict adherence to physical and chemical constraints, potentially sacrificing execution robustness for reasoning flexibility.
5.3 No Single Winner: The Absence of a Scientific "Superintelligence"
Across the eight discovery projects evaluated, no single LLM emerged as a definitive winner. Leadership rotates depending on the task, with different models excelling at protein design, symbolic regression, or TMC optimization. This variability demonstrates that scientific discovery is a composite activity requiring a blend of different skills, and a deficit in any one constituent scenario can introduce compounding errors that derail a project. The absence of a single dominant model proves that all contemporary LLMs are distant from a true, general-purpose scientific "superintelligence." Instead, they are specialized tools with unique strengths and weaknesses, underscoring the need for balanced capabilities to effectively orchestrate the full discovery loop.
Ultimately, project-level evaluation reveals that an LLM's utility in science is more complex than a simple function of its Q&A accuracy. This understanding is crucial for shaping the next generation of science-ready AI.
--------------------------------------------------------------------------------
6.0 The Path Forward: A Blueprint for Developing Science-Ready LLMs
The findings from the Scientific Discovery Evaluation (SDE) framework, spanning both question-level and project-level assessments, provide a clear and urgent message: to build better AI for science, we must change how we measure it. The insights gathered offer a strategic roadmap for the AI development community, moving beyond the current paradigms of indiscriminate scaling and generic reasoning. The following recommendations represent a blueprint for creating LLMs that are genuinely equipped to accelerate scientific discovery.
1. Shift Focus from Indiscriminate Scaling to Targeted Training The diminishing returns observed from simply increasing model size and test-time compute suggest a need for more focused training. The SDE framework reveals specific gaps in scientific methodology, such as problem formulation and hypothesis refinement. Future development should prioritize targeted training on these core competencies to build models that can reason more like scientists, not just larger knowledge bases.
2. Diversify Pre-Training Data and Inductive Biases The high correlation of errors among top models points to a critical bottleneck: a lack of diversity in pre-training data. This homogeneity creates shared blind spots that limit progress on the most difficult scientific questions. There is an urgent need to move beyond current data sources, explore novel inductive biases, and curate domain-specific datasets that can mitigate these shared failure modes and foster more robust and original reasoning.
3. Integrate Robust Tool Use into Fine-Tuning Many research scenarios are intractable without coupling linguistic reasoning with computational tools like simulators and data analysis libraries. Therefore, training and evaluation must expand beyond textual accuracy to prioritize executable actions. Fine-tuning should focus on a model’s ability to reliably invoke external tools, debug execution failures, and iteratively refine its approach based on computational feedback.
4. Develop Scientific Reasoning-Specific Reinforcement Learning Reasoning enhancements optimized for mathematics and coding yielded negligible gains in many SDE projects, indicating that scientific reasoning is a distinct capability. This highlights a promising frontier for research: developing reinforcement learning strategies tailored specifically for the scientific discovery process. Such methods could train models to navigate complex hypothesis spaces, balance exploration and exploitation, and adapt to noisy or imperfect evidence.
To unlock the profound potential of artificial intelligence in science, we must first get the measurement right. The prevailing benchmarks, modeled on academic quizzes, have guided the field to a plateau. The SDE framework offers a more robust, multi-scale methodology that reflects the dynamic reality of research. By adopting such scenario-grounded evaluations, the AI community can gain the necessary navigational insights to steer development away from a plateau of quiz-solving and toward the true frontier of genuine scientific discovery.
#AI #AIScience #AIDiscovery #AIBenchmarks #AIScienceBenchmarks