Scientific Algorithm Discovery by Augmenting AlphaEvolve with Deep Research (2510.06056v1)

Abstract: LLMs hold promise as scientific assistants, yet existing agents either rely solely on algorithm evolution or on deep research in isolation, both of which face critical limitations. Pure algorithm evolution, as in AlphaEvolve, depends only on the internal knowledge of LLMs and quickly plateaus in complex domains, while pure deep research proposes ideas without validation, resulting in unrealistic or unimplementable solutions. We present DeepEvolve, an agent that integrates deep research with algorithm evolution, uniting external knowledge retrieval, cross-file code editing, and systematic debugging under a feedback-driven iterative loop. Each iteration not only proposes new hypotheses but also refines, implements, and tests them, avoiding both shallow improvements and unproductive over-refinements. Across nine benchmarks in chemistry, mathematics, biology, materials, and patents, DeepEvolve consistently improves the initial algorithm, producing executable new algorithms with sustained gains. By bridging the gap between unguided evolution and research without grounding, DeepEvolve provides a reliable framework for advancing scientific algorithm discovery. Our code is available at https://github.com/liugangcode/deepevolve.

• The paper presents DeepEvolve, which augments traditional algorithm evolution using deep research to sustain progress in diverse scientific tasks.
• DeepEvolve leverages external knowledge retrieval and cross-file editing to incorporate domain-specific inductive biases into algorithm design.
• Experimental evaluations across nine domains demonstrate substantial performance improvements, validating the framework’s capacity to discover high-impact algorithms.

Scientific Algorithm Discovery by Augmenting AlphaEvolve with Deep Research

Motivation and Limitations of Pure Algorithm Evolution

The paper addresses the challenge of automating scientific algorithm discovery using LLM-based agents. Prior approaches such as AlphaEvolve rely on pure algorithm evolution, where LLMs generate and mutate code based solely on their internal knowledge. This strategy is effective for narrow domains (e.g., $4 \times 4$ matrix multiplication) but quickly plateaus in broader scientific fields due to limited idea generation and lack of external grounding. Empirical results show that AlphaEvolve-style evolution yields only marginal improvements after the first generation, with the best candidate often appearing early and subsequent refinements offering little gain.

Figure 1: Comparison of AlphaEvolve (top) and DeepEvolve (bottom) showing stagnation in pure evolution and sustained progress with deep research integration.

DeepEvolve: Integrating Deep Research with Algorithm Evolution

DeepEvolve is introduced as a modular agentic framework that alternates between deep research and algorithm evolution. The system is composed of six collaborative modules: planning, searching, proposal writing, coding, evaluation, and evolutionary selection. Deep research leverages external knowledge retrieval, synthesizes domain-specific hypotheses, and grounds proposals in prior successful algorithms. Algorithm evolution then implements, debugs, and evaluates these proposals, feeding performance feedback back into the research loop.

Figure 2: DeepEvolve architecture with six modules alternating between deep research and algorithm evolution.

Key innovations include:

• External Knowledge Retrieval: Deep research agents query scientific databases (e.g., PubMed, arXiv) to inform hypothesis generation.
• Cross-file Code Editing: Coding agents parse and modify multi-file codebases, enabling complex algorithmic changes beyond single-file edits.
• Automated Debugging: Debugging agents iteratively resolve execution errors, substantially increasing implementation success rates.
• Evolutionary Database: MAP-Elites and island-based population strategies maintain diverse candidate pools and sample inspirations for future generations.

Experimental Evaluation Across Nine Scientific Domains

DeepEvolve is benchmarked on nine tasks spanning chemistry, mathematics, biology, materials science, and patents. Each task is defined by a problem $P = (D, g)$, initial algorithm $f$, and user instructions $u$. The agent iteratively evolves $f$ to maximize a standardized performance score $s = g(\{a_i\}, \{\hat{a}_i\})$ under compute and runtime constraints.

Quantitative results demonstrate consistent improvements over initial algorithms, with performance gains ranging from 0.39% to 666.02% depending on domain and baseline maturity. For example, in the Circle Packing task, DeepEvolve discovers algorithms that generalize to variable-sized packings and maintain validity, yielding a 666% improvement over the baseline. In molecular property prediction, DeepEvolve introduces contrastive learning and motif-aware masking, improving both accuracy and interpretability.

Figure 3: DeepEvolve-generated model.forward() for molecular prediction, highlighting contrastive learning and motif-aware masking.

Figure 4: Score trajectories over iterations, showing non-monotonic but substantial jumps in performance.

Ablation studies confirm that deep research is critical for sustained progress. Pure algorithm evolution without external grounding fails to generate substantive improvements, while the addition of deep research yields stronger candidates within fewer generations.

Synergy Between Deep Research and Coding Agents

Analysis of evolutionary trajectories reveals that deep research and coding agents reinforce each other:

• Domain-specific Inductive Biases: Deep research extracts priors (e.g., molecular motifs, polymer periodicity) and guides coding agents to instantiate these as algorithmic modules.
• Feedback-driven Design Shifts: Performance feedback transitions development from heuristic tuning to principled, physically grounded methods (e.g., certified global optimization, physics-informed regularization).
• Reusable Methodological Patterns: DeepEvolve consistently discovers generalizable strategies such as uncertainty estimation, dynamic loss reweighting, and self-supervised learning, which are instantiated across tasks.

  Figure 5: LLM-as-a-judge evaluation of initial and new algorithmic ideas, showing improved originality and future potential.

Implementation Details and System Prompts

DeepEvolve orchestrates multiple LLMs for different agent roles (planning, searching, coding, debugging, reflection). System prompts are carefully engineered for each module to ensure focused, high-quality outputs. The evolutionary database uses MAP-Elites for inspiration sampling and island-based populations for candidate selection, balancing exploitation and exploration.

Figure 6: System prompts for planning in the deep research agent.

Figure 7: System prompts for coding in the coding agent (part 1 of 2).

Figure 8: System prompts for debugging in the coding agent.

Performance, Resource Requirements, and Scaling

DeepEvolve operates under strict runtime (30 minutes) and hardware (single GPU) constraints. Efficiency is not the primary optimization objective but can be specified in user instructions. The system demonstrates robust scaling across diverse modalities (molecules, images, text, time series, PDEs) and complex codebases. Automated debugging increases execution success rates from 0.13 to 0.99 in challenging tasks (e.g., Open Vaccine).

Practical and Theoretical Implications

Practically, DeepEvolve provides a reliable framework for advancing scientific algorithm discovery, automating the full pipeline from hypothesis generation to executable code. Theoretically, the integration of deep research with evolutionary feedback enables agents to escape local optima and discover novel, high-impact algorithms. The modular design supports extensibility to new domains and problem types.

Future Directions

Potential future developments include:

• Scaling to Larger Codebases: Extending cross-file editing and debugging to complex, multi-repository scientific software.
• Integration with Experimental Automation: Coupling algorithm discovery with lab automation for closed-loop scientific research.
• Meta-learning and Transfer: Leveraging evolutionary history and domain priors for rapid adaptation to new tasks.
• Benchmark Expansion: Systematic evaluation on open-ended scientific problems and real-world experimental datasets.

Conclusion

DeepEvolve demonstrates that augmenting algorithm evolution with deep research enables sustained, reliable progress in scientific algorithm discovery. By combining external knowledge retrieval, cross-file code editing, and iterative debugging, DeepEvolve consistently improves baseline algorithms across diverse domains. The framework advances both the practical automation of scientific research and the theoretical understanding of agentic algorithm discovery, with strong implications for future AI-driven science.