Algorithms Now Discover and Write Scientific Papers Autonomously

Zhe Zhao and colleagues at City University present ResearchEVO, a new end-to-end framework that mimics the iterative process of scientific discovery, beginning with experimentation and followed by theoretical explanation. The system uniquely combines algorithm evolution, driven by performance, with automated research paper generation, ensuring factual accuracy and avoiding fabricated citations. Validated on problems in quantum error correction using real hardware data and physics-informed neural networks, ResearchEVO successfully discovered previously unproposed algorithmic mechanisms and autonomously produced complete, publication-ready manuscripts. It represents a key step towards fully automated scientific research and documentation.

Bi-dimensional co-evolution and retrieval-augmented generation for automated scientific discovery

A novel end-to-end framework, ResearchEVO, bridges the gap between algorithm discovery and scientific explanation. A research problem is specified as a triple comprising reference code, seed bibliography, and domain dataset. The Evolution Phase then searches the space of algorithmic implementations through bi-dimensional co-evolution, simultaneously optimising both the internal logic of algorithmic modules and the overall algorithm architecture. This two-dimensional search, combined with reflective feedback and domain-adaptive sandbox evaluation providing structured error diagnostics beyond scalar fitness scores, enables the system to discover novel algorithms without being confined to predefined templates.

The Writing Phase takes the best-performing algorithm and autonomously constructs 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 verification. Because the paper’s claims are anchored in actually evolved and evaluated code, fabrication of results is structurally harder than in systems that generate papers independently of algorithm development. The framework validates its capabilities on two cross-disciplinary scientific problems with real-world significance.

In Quantum Error Correction (QEC), the Evolution Phase discovered topologically-aware edge weights for surface-code MWPM decoders, validated on real Google quantum hardware data. In Physics-Informed Neural Networks (PINN), the Evolution Phase evolved a trust-region loss adaptor with residual connections that consistently reduces approximation error. Early work on automated machine learning demonstrated that entire learning algorithms could be evolved from mathematical primitives, establishing the concept of algorithm evolution.

Neural architecture search showed that network topologies could be discovered without human intervention. Large language models have opened a new frontier, as they can directly generate, mutate, and recombine executable code, effectively serving as intelligent evolutionary operators over the space of algorithms. These LLM-guided evolution systems have achieved remarkable results, from discovering novel mathematical solutions to producing the first improvement over a classical matrix multiplication algorithm in over fifty years.

Most methods confine their search to a single dimension, evolving the logic of a fixed function within a predefined template, leaving the overall algorithm architecture untouched. Existing evolution systems terminate at a code artifact, lacking a scientific narrative, connection to existing theory, and explanation of why the discovered solution works. A parallel line of research has tackled the problem of automated scientific writing, with systems for end-to-end research automation capable of generating ideas, executing experiments, and producing complete research papers.

Multi-agent frameworks for hypothesis generation employ debate, knowledge-graph reasoning, and iterative peer review to produce research ideas. Long-form writing systems use multi-perspective retrieval and RAG pipelines for structured document generation. However, these systems automate the explanation stage without genuinely performing the discovery stage, as their “discoveries” originate from the LLM’s parametric memory rather than from principled search over an algorithm space.

Systems often struggle to identify genuinely novel solutions beyond their training data, and many focus on machine learning benchmarks instead of problems with real scientific relevance. ResearchEVO presents a framework that combines discovery and explanation through a two-stage process of experimentation and retrospective analysis. It uses LLM-guided co-evolution to optimise both algorithmic logic and architecture, searching for code based purely on performance.

The Writing Phase then generates a complete research paper, verifying claims and designing experiments autonomously. Validation on Quantum Error Correction and Physics-Informed Neural Networks revealed previously unknown algorithmic mechanisms, documented in compilable manuscripts with verified citations. An important pattern appears in scientific breakthroughs: an initial phase of experimentation yielding an unexpected finding, followed by explaining why it works.

ResearchEVO instantiates this discover-then-explain paradigm through LLM-guided bi-dimensional co-evolution, simultaneously optimising algorithmic logic and architecture by fitness alone. The system then generates a research paper via retrieval-augmented generation with verification, discovering mechanisms not previously proposed in quantum error correction and physics-informed neural networks. ResearchEVO employs LLM-guided bi-dimensional co-evolution, simultaneously optimising algorithmic logic and overall architecture, to search for code implementations based purely on performance.

The framework extends this process to multi-objective settings via management of Pareto fronts. It discovers algorithms without requiring templates, optimising both functional logic and structural architecture. An important pattern in scientific breakthroughs involves an initial phase of experimentation yielding unexpected findings, followed by explanation and theoretical integration. ResearchEVO is a framework computationally instantiating this discover-then-explain paradigm.

Its Evolution Phase employs LLM-guided co-evolution, optimising algorithmic logic and architecture by fitness alone. Subsequently, the Writing Phase autonomously generates a research paper through retrieval-augmented generation with verification and experiment design. To date, no system jointly performs algorithm evolution and literature-grounded documentation. Validation across Quantum Error Correction and Physics-Informed Neural Networks revealed novel algorithmic mechanisms, documented in compilable manuscripts with zero fabricated citations.

Automated algorithm evolution delivers novel solutions and complete scientific manuscripts

The framework has achieved a 30000-fold improvement in algorithmic discovery, exceeding the scale of experimentation undertaken by Gregor Mendel with his pea plants. This breakthrough surpasses previous automated systems limited to incremental refinements of pre-defined templates, unlocking genuinely novel algorithmic mechanisms. ResearchEVO uniquely combines algorithm evolution with automated scientific documentation, autonomously generating complete, publication-ready manuscripts in LaTeX format with zero fabricated citations.

Validated on Quantum Error Correction and Physics-Informed Neural Networks, ResearchEVO instantiates a discover-then-explain paradigm by computationally mirroring a two-stage process of experimentation followed by theoretical explanation. The Evolution Phase employs LLM-guided bi-dimensional co-evolution, simultaneously optimising algorithmic logic and architecture, to search for code implementations based purely on performance. The Writing Phase then generates a complete research paper through sentence-level retrieval-augmented generation, incorporating explicit verification and automated experiment design.

Validated using data from Google quantum hardware, the framework discovered human-interpretable algorithmic mechanisms not previously proposed in existing literature for both Quantum Error Correction and Physics-Informed Neural Networks. In Physics-Informed Neural Networks, the system devised a trust-region loss adaptor, reducing approximation error. The generated manuscripts, compiled in LaTeX format, included detailed experiment designs autonomously executed to verify the functionality of the newly discovered algorithms. Scaling this approach to tackle genuinely open-ended scientific challenges and broader, less-structured datasets remains a hurdle.

Evolving algorithms and explaining their reasoning through automated scientific publication

Systems capable of automated discovery are being built, yet a critical gap remains between finding a working solution and understanding why it works. ResearchEVO addresses this by not only evolving algorithms but also generating a complete scientific paper explaining the process, a feat unmatched by prior work. However, a trade-off exists with AlphaEvolve, a system focused on sheer scale and practical impact within Google’s infrastructure, demonstrably delivering practical gains in areas like matrix multiplication and data centre efficiency, benefits ResearchEVO’s smaller experiments cannot yet rival.

Acknowledging concerns about comparing academic research with Google’s industrial-scale achievements is important. This system prioritises understanding alongside performance; it does not simply find a solution, but articulates why it works via a complete, verifiable research paper. Researchers have created a system capable of not only discovering algorithms but also writing a scientific paper explaining how they work, delivering both performance and traceability from code to theory, unlike systems such as AlphaEvolve which prioritises performance gains over understandable solutions.

Establishing a complete cycle of automated discovery and explanation fundamentally alters the field of scientific research. ResearchEVO uniquely integrates algorithm evolution, a process of refining code based on performance, with the automated generation of fully-formed scientific papers, bridging a critical gap in existing artificial intelligence systems. This framework not only uncovers novel algorithmic mechanisms in fields like quantum error correction and physics-informed neural networks, but also articulates their function and connects them to established scientific knowledge via literature review.

The research successfully demonstrated an end-to-end framework, ResearchEVO, capable of both evolving algorithms and autonomously writing a complete scientific paper detailing their function. This matters because it addresses the crucial need to understand why an algorithm works, not just that it produces a result, a step missing in many automated discovery systems. In two areas, quantum error correction and physics-informed neural networks, the system discovered previously unproposed algorithmic mechanisms and then explained them with reference to existing literature. The authors suggest this approach establishes a complete cycle of automated discovery and explanation, integrating code refinement with scientific documentation.