Zhe Zhao and colleagues at City University announced ResearchEVO, a new end-to-end framework that mimics the iterative process of scientific discovery, starting with experimentation and followed by theoretical explanation. The system uniquely combines performance-driven algorithmic advances and automated research paper generation to ensure factual accuracy and avoid fabricated citations. ResearchEVO examined the problem of quantum error correction using real hardware data and physics-based neural networks, discovered previously unproposed algorithmic mechanisms, and succeeded in autonomously producing a complete publishable manuscript. This is an important step towards fully automated scientific research and documentation.
Two-dimensional coevolution and search augmentation generation for automated scientific discovery
ResearchEVO, a new end-to-end framework, bridges the gap between algorithmic discovery and scientific explanation. A research question is specified as a triple consisting of a reference code, a seed bibliography, and a domain dataset. Then, the evolution phase explores the space of algorithm implementations through two-dimensional co-evolution, optimizing both the internal logic of the algorithm modules and the overall algorithm architecture at the same time. This two-dimensional search, combined with reflective feedback and domain-adaptive sandbox evaluation that provides structured error diagnostics beyond scalar suitability scores, enables the system to discover new algorithms without being restricted to predefined templates.
The writing phase employs the best-performing algorithms to autonomously build a complete publication-ready research paper through three stages: literature crawling and vector indexing, automated experiment design and execution, and sentence-level RAG-enhanced section writing with explicit anti-hallucination validation. Because the paper’s claims are fixed in the code that is actually deployed and evaluated, fabricating results is structurally more difficult than in systems that generate papers independently of algorithm development. This framework validates its capabilities for two interdisciplinary scientific problems of real-world importance.
For quantum error correction (QEC), the evolution phase discovered topology-aware edge weights for surface-coded MWPM decoders that were validated on real Google quantum hardware data. In Physics-Informed Neural Networks (PINNs), the evolution phase evolved a trust-region loss adapter with residual connectivity that consistently reduces the approximation error. Early research on automated machine learning established the concept of algorithm evolution by demonstrating that entire learning algorithms could evolve from mathematical primitives.
We found that neural architecture searches can discover network topologies without human intervention. Large-scale language models break new ground because they can directly generate, modify, and recombine executable code, effectively acting as intelligent evolutionary operators across the algorithmic space. These LLM-guided evolutionary systems have achieved impressive results, from discovering new mathematical solutions to improving the classic matrix multiplication algorithm for the first time in more than 50 years.
Most methods limit the search to a single dimension, evolve fixed-function logic within a predefined template, and leave the overall algorithm architecture untouched. Existing evolutionary systems end up with code artifacts, lacking scientific narratives, connections to existing theories, and explanations of why the solutions discovered work. Parallel research has tackled the problem of automated scientific writing, developing end-to-end research automation systems capable of generating ideas, performing experiments, and producing complete research papers.
Multi-agent frameworks for hypothesis generation employ discussion, knowledge graph inference, and iterative peer review to generate research ideas. The long-form description system uses multi-perspective search and RAG pipelines to generate structured documents. However, these systems automate the explanation stage without actually performing the discovery stage, since their “discovery” originates from the parametric memory of the LLM rather than a principled search across the algorithmic space.
Systems often struggle to identify truly novel solutions beyond their training data, and many focus on machine learning benchmarks rather than problems of actual scientific relevance. ResearchEVO provides a framework for combining discovery and explanation through a two-step process of experimentation and retrospective analysis. Optimize both algorithm logic and architecture using LLM-guided coevolution to search code purely based on performance.
Then, during the writing phase, you create a complete research paper, test your claims, and autonomously design your experiments. Verification of quantum error correction and physics-based neural networks revealed previously unknown algorithmic mechanisms, which were documented in an editable manuscript with verified citations. There are important patterns in scientific progress. That is, an unexpected discovery is made early in the experiment, which then explains why it works.
ResearchEVO instantiates this discover-then-explain paradigm through LLM-guided two-dimensional coevolution, while optimizing algorithm logic and architecture by fitness alone. The system then generates research papers through search-enhanced generation with verification to discover previously unproposed mechanisms in quantum error correction and physics-based neural networks. ResearchEVO employs LLM-guided two-dimensional co-evolution to simultaneously optimize the algorithm logic and the entire architecture, finding code implementations based purely on performance.
This framework extends this process to multi-objective settings through Pareto front management. Discover algorithms and optimize both functional logic and structural architecture without the need for templates. An important pattern in scientific progress involves an initial stage of experimentation that leads to an unexpected discovery, followed by explanation and theoretical synthesis. ResearchEVO is a framework that computationally instantiates this discover-then-explain paradigm.
Its evolution phase employs LLM-guided co-evolution to optimize the algorithm logic and architecture by fitness alone. Then, in the writing phase, research papers are generated autonomously through search expansion generation with verification and experimental design. To date, there is no system that jointly performs algorithm evolution and literature-based documentation. Validation across quantum error correction and physically informed neural networks reveals a novel algorithmic mechanism, documented in an editable manuscript free of fabricated citations.
Advances in automated algorithms provide new solutions and complete scientific papers
The framework achieves a 30,000-fold improvement in algorithmic discovery, surpassing the scale of Gregor Mendel’s experiments with peas. This breakthrough surpasses previous automation systems, which were limited to incremental refinement of predefined templates, and unlocks truly novel algorithmic mechanisms. ResearchEVO uniquely combines algorithmic advances and automated scientific documentation to autonomously generate complete, publication-ready manuscripts in LaTeX format, free of fabricated citations.
Validated with quantum error correction and physics-based neural networks, ResearchEVO instantiates a discover-then-explain paradigm by computationally reflecting a two-step process: experiment followed by theoretical explanation. The evolution phase employs LLM-guided two-dimensional co-evolution to simultaneously optimize algorithm logic and architecture and search for code implementations purely based on performance. The writing phase then generates a complete research paper through sentence-level search-enhanced generation, incorporating explicit validation and automated experimental design.
The framework, validated using data from Google quantum hardware, discovered human-interpretable algorithmic mechanisms for both quantum error correction and physically-informed neural networks that had not been previously proposed in the existing literature. In Physics-Informed Neural Networks, the system devised a trust region loss adapter to reduce the approximation error. The resulting manuscript, edited in LaTeX format, included a detailed experimental design that was run autonomously to verify the functionality of the newly discovered algorithm. Hurdles remain to scale this approach to tackle truly open-ended scientific questions and broader, less structured datasets.
Evolving algorithms and explaining their reasoning through automated scientific publishing
Although systems are being built to enable automatic detection, there remains a significant gap between finding an effective solution and understanding why it works. ResearchEVO addresses this problem by not only evolving the algorithm, but also by creating a complete scientific paper explaining the process. This is a feat unmatched in previous studies. However, there are trade-offs with AlphaEvolve. AlphaEvolve is a system focused on sheer scale and practical impact within Google’s infrastructure, delivering clear practical benefits in areas such as matrix multiplication and data center efficiency, benefits that ResearchEVO’s small-scale experiments cannot yet match.
It is important to acknowledge the concerns about comparing academic research with Google’s industrial-scale results. This system prioritizes understanding along with performance. We don’t just find a solution, we clearly demonstrate why it works through a complete and verifiable research paper. Researchers have created a system that can not only discover algorithms but also write scientific papers explaining how they work, delivering both performance and traceability from code to theory, unlike systems like AlphaEvolve that prioritize performance improvements over easy-to-understand solutions.
Establishing a complete cycle of automated discovery and explanation will fundamentally change the field of scientific research. ResearchEVO uniquely integrates algorithmic evolution, the process of improving code based on performance, with the automatic generation of fully formed scientific papers, filling a critical gap in existing artificial intelligence systems. This framework not only reveals new algorithmic mechanisms in areas such as quantum error correction and physics-based neural networks, but also clarifies their functionality and connects them to established scientific knowledge through a literature review.
This work successfully demonstrated ResearchEVO, an end-to-end framework capable of both evolving algorithms and autonomously producing full scientific papers detailing their functionality. This is important because it addresses an important need to understand. why Algorithms work by doing more than just producing results, a step that many automated detection systems lack. In two areas, quantum error correction and physics-based neural networks, the system discovered previously unproposed algorithmic mechanisms and referenced existing literature to explain them. The authors propose that this approach establishes a complete cycle of automated discovery and explanation, integrating code refinement and scientific documentation.
