Source-linked AI summary
Scientific Algorithm Discovery by Augmenting AlphaEvolve with Deep Research
Gang Liu, Yihan Zhu, Jie Chen, Meng Jiang
TL;DR
LLM-driven algorithm discovery faces a gap between generating scientific hypotheses and implementing them as working code; pure evolution and untested research each have limitations. DeepEvolve combines research, implementation, and evaluation in iterative algorithm evolution. Across nine benchmarks, it consistently improves baseline algorithms and produces executable programs with higher performance and efficiency.
Problem
LLM agents face challenges both in advancing algorithms through internal knowledge alone and in implementing generated scientific hypotheses as working code.
Source: §1 · p. 1Method
Results
Takeaways & Limitations
Ablations and case studies indicate that research contributes domain-specific design insights and debugging improves robustness in complex implementations.
Source: §6 · p. 9On Open Vaccine, training often exceeded the 30-minute runtime budget, constraining DeepEvolve's search space.
Source: §4.2 · p. 7
Abstract
from arXiv · showhide
Large language models 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.
1 INTRODUCTION
LLM agents still struggle to turn scientific hypotheses into working code, while pure algorithm evolution can plateau in broader domains. DeepEvolve combines research with implementation and evaluation, and improves algorithms across nine scientific problems.
1 INTRODUCTION
- LLM agents can generate scientific hypotheses, but implementing them as working code remains challenging.Scientific problems that are difficult to solve but easy to evaluate motivate algorithm discovery through LLM reasoning, planning, and execution.Source: §1 · p. 1
- Pure AlphaEvolve-style evolution improved molecular property prediction by only 0.6% after 100 iterations, with its best candidate appearing in the first generation.The reproduced algorithm's score rose from 0.791 to 0.797; deeper refinements generally yielded only marginal gains.Source: §1 · p. 1
- Figure 1 contrasts pure evolution's marginal later gains with DeepEvolve's sustained progress and clear jumps at key iterations.DeepEvolve aims to avoid both shallow evolution and overly deep but unproductive refinements.Source: Figure 1 · p. 2, §1 · p. 2
- DeepEvolve combines research proposals with cross-file code editing, debugging, and evaluation feedback to refine subsequent proposals.Research is grounded in existing algorithms; early proposals favor feasible ideas, while later ones target higher impact.Source: §1 · p. 2
- Across nine problems in chemistry, mathematics, biology, materials, and patents, DeepEvolve consistently improves existing algorithms and produces original, promising methods with high performance scores.The benchmarks cover diverse data modalities, including molecules, geometries, partial differential equations, and images.Source: §1 · p. 3
2 PROBLEM DEFINITION FOR ALGORITHM DISCOVERY
The paper formalizes algorithm discovery as optimizing an algorithm's evaluation score on a scientific problem. A problem comprises evaluation data and a function that compares predicted with ground-truth answers.
2 PROBLEM DEFINITION FOR ALGORITHM DISCOVERY
- A scientific problem is represented as P = (D, g), where D contains question-answer pairs and g compares ground-truth answers with predictions.For circle packing, the objective is to maximize the sum of radii for n circles inside a unit square; molecular property prediction is another example.Source: §2 · p. 3
- An algorithm f maps each question to a predicted answer, and its score s is computed by applying g to predictions and ground-truth answers.The goal is to optimize f for a higher score, with computation and evaluation completed within bounded time.Source: §2 · p. 3
3 DEEPEVOLVE FOR ALGORITHM DISCOVERY
DeepEvolve repeatedly updates an initial algorithm through six modules, selecting the highest-scoring candidate across iterations. Its inputs specify the problem, starting algorithm, and user requirements; its framework couples research, coding, evaluation, and evolutionary selection.
3 DEEPEVOLVE FOR ALGORITHM DISCOVERY
- DeepEvolve repeatedly applies a six-module update operator and returns the highest-scoring algorithm among the initial and intermediate versions.The operator takes a problem, initial algorithm, and user instructions as inputs.Source: §3 · p. 3
3.1 INPUT OF PROBLEM, ALGORITHMS, AND INSTRUCTIONS
- The input context includes the evaluation function, data, and textual problem description; user instructions can add preferences or constraints.The starting algorithm includes multi-file code and a description of its motivation, pseudo-code, performance, and qualitative assessments; Table 1 summarizes benchmark tasks, data types, domains, and metrics.Source: §3.1 · p. 3, Table 1 · p. 5
3.2 FRAMEWORK DESIGNS
- Deep research plans questions, searches online sources, and proposes an algorithm with pseudo-code, favoring feasibility early and higher-impact ideas later.Proposals integrate retrieved evidence with the problem, current algorithm, and inspiring algorithms.Source: §3.2 · p. 4
- A coding agent makes targeted edits across files, while a debugging agent uses execution feedback to address implementation failures.If execution remains unsuccessful after a limited number of attempts, the algorithm receives a score of zero.Source: §3.2 · p. 4
- Evaluated algorithms are stored for future candidate sampling and inspiration, while research, implementation, and evaluation are coupled across iterations.The evolutionary database uses populations and selects candidates while retaining exploration; reflection checkpoints help check research progress and code alignment.Source: §3.2 · pp. 4–5
4 EXPERIMENTS
Experiments evaluate DeepEvolve across nine scientific problems under a 30-minute development budget, examining performance, agent interactions, and design ablations. DeepEvolve improves both effectiveness and efficiency on six tasks, and improves the primary objective within the runtime limit on the other three.
4.1 SET-UPS
- Experiments cover nine problems with diverse data modalities, using initial algorithms as baselines and a 30-minute development budget on one GPU.The study standardizes problem-specific metrics into higher-is-better scores and treats efficiency as a secondary objective.Source: §4.1 · pp. 5–6
4.2 RQ1: EFFECTIVENESS AND EFFICIENCY FOR THE NEWLY DISCOVERED ALGORITHMS
- DeepEvolve improves effectiveness and efficiency on six of nine tasks; on the other three, it improves the primary objective within the 30-minute runtime constraint.Table 2 compares new algorithms with their initial versions by new score and runtime; efficiency is a secondary objective.Source: Table 2 · p. 6, §4.2 · p. 6
- Across tasks, performance gains range from 0.39% to 666.02% and vary with task type and initial-algorithm maturity.In Circle Packing, the fixed-26-circle baseline often produced invalid solutions, while DeepEvolve found valid packings across a broader range of circle counts.Source: §4.2 · p. 6
- DeepEvolve proposes ideas judged for originality, future potential, and implementation difficulty; automatic debugging supports complex implementations.For Molecular Prediction, the new model adds contrastive learning and motif-aware masking; on Open Vaccine, execution success rises from 0.13 to 0.99.Source: §4.2 · p. 7, Figure 4 · p. 7, Table 3 · p. 8
4.3 RQ2: ITERATIVE SYNERGY BETWEEN DEEP RESEARCH AND CODING AGENTS
- Deep research uses domain-specific priors, including molecular motifs, polymer periodicity, chemical grammars, CDEs, LoRA, CPC embeddings, and physiological waveforms.These priors inform task-specific algorithm choices across molecular, polymer, Parkinson’s Disease, and USP P2P problems.Source: §4.3 · p. 8
- Performance feedback guides subsequent research from heuristic tuning toward methods with theoretical or physical guarantees, including certified packing optimization and Krylov solvers.The paper also describes physics-informed regularization for disease dynamics as an example of this progression.Source: §4.3 · p. 8
- Across tasks, DeepEvolve finds reusable patterns such as uncertainty estimation, dynamic loss reweighting, and self-supervised representation learning in task-specific modules.Examples include uncertainty-guided refinement for molecular motifs and nuclei-image boundaries, and adaptive loss weighting in Open Vaccine, Parkinson’s Disease, and USP P2P.Source: §4.3 · p. 8
4.4 RQ3: ABLATION AND CASE STUDIES FOR ALGORITHM IMPROVEMENT
- Deep-research evolution achieves stronger improvements within about five generations than LLM-internal-knowledge-only evolution in Molecular Prediction and Circle Packing.All evolved candidates in both cases outperform the initial algorithms; score improvements often occur as sudden jumps, and debugging raises execution success.Source: Table 4 · p. 8, Figure 5 · p. 8, §4.4 · p. 9
5 RELATED WORK
Prior work explores algorithm discovery through evolutionary coding systems, while scientific agents and deep-research systems address other stages of discovery; implementing ideas as executable code remains a challenge.
5.1 AUTOMATED ALGORITHM DISCOVERY
- Algorithm discovery combines evaluation feedback with evolutionary optimization, but AlphaEvolve's broader-domain generalization and external grounding remain uncertain [Novikov et al., 2025].Related work also notes challenges in editing and understanding code across files.Source: §5.1 · p. 9
5.2 AGENT FOR SCIENTIFIC DISCOVERY
- Scientific agents span idea generation through code execution, yet EXP-Bench reports below 1% full-pipeline success [Kon et al., 2025].LLM-generated ideas were found more novel than experts' ideas but less feasible (Si et al., 2024).Source: §5.2 · p. 9
6 CONCLUSION
DeepEvolve combines deep research with algorithm evolution, cross-file editing, and iterative debugging. Across nine diverse scientific benchmarks, it improves baseline algorithms, with ablations and case studies highlighting the roles of research insights and debugging.
6 CONCLUSION
- Across nine benchmarks, DeepEvolve consistently improved baseline algorithms, producing executable programs with higher performance and efficiency.Ablations and case studies found that deep research guided design with domain-specific insights, while debugging improved robustness in complex implementations.Source: §6 · p. 9
- DeepEvolve integrates deep research, cross-file code editing, and iterative debugging to combine idea generation with reliable execution.The conclusion describes the approach as having potential for future AI-driven scientific discovery.Source: §6 · p. 9
A DETAILS ON DEEPEVOLVE METHOD
The appendix assigns distinct LLM roles to DeepEvolve's deep-research and coding agents.
A.1 LLMS
- Deep research uses o4-mini for planning and reflection, gpt-4o for searching, and o3-mini for proposal writing.Source: §A.1 · p. 12
- The coding agent uses o3-mini for code development and o4-mini for code debugging.Source: §A.1 · p. 12
- o4-mini serves in both agents: as the deep-research planner and reflection agent, and as the coding debugger.Source: §A.1 · p. 12
A.2 EVOLUTIONARY DATABASE
The evolutionary database retains past discoveries for inspiration and candidate selection, while the appendix also describes research-agent and coding-agent templates.
A.2 EVOLUTIONARY DATABASE
- Past discoveries serve as inspirations and next candidates, using MAP-elites for inspiration sampling and the island algorithm for candidate sampling.Source: §A.2 · p. 12
- MAP-elites archives the best algorithms, scores them by performance, diversity, and complexity, and samples five inspirations per iteration, including the current best.The archive defaults to size 10; each dimension is normalized and divided into 10 bins.Source: §A.2 · p. 12
- The island algorithm maintains up to 25 algorithms across five islands, balancing exploitation and exploration at probabilities 0.7 and 0.3.Islands can migrate their best program to neighboring islands every 25 iterations by default.Source: §A.2 · p. 12
A.3 TEMPLATES FOR DEEP RESEARCH AGENTS
- Deep-research prompts for planning, searching, reflection, and report writing are shown in Figures 6–9; the user-input template appears in Figures 10–11.Source: §A.3 · p. 12
A.4 TEMPLATES FOR THE CODING AGENT
- The coding-agent appendix provides coding and debugging prompts and an input template, then describes code reflection before evaluation and error-guided debugging.Reflection uses the coding agent's LLM with a different prompt; execution errors are captured for debugging.Source: §A.4 · p. 12
B DETAILS ON THE BENCHMARKING PROBLEMS
The benchmark suite covers nine research problems across multiple scientific domains, data modalities, and evaluation metrics standardized to higher-is-better scores.
B DETAILS ON THE BENCHMARKING PROBLEMS
- Nine problems span chemistry, mathematics, biology, and materials, with diverse inputs including molecules, images, mRNA, text, time series, geometric structures, and multimodal data.Their metrics, including AUC-ROC, RMSE, precision, and Pearson correlation, are standardized so higher scores indicate better performance.Source: B DETAILS ON THE BENCHMARKING PROBLEMS · p. 12
- Evaluation metrics include AUC-ROC, RMSE, precision, and Pearson correlation, converted to a common higher-is-better score.Source: B DETAILS ON THE BENCHMARKING PROBLEMS · p. 12
- The benchmark inputs include molecules, images, mRNA, text, time series, geometric structures, and multimodal data.Source: B DETAILS ON THE BENCHMARKING PROBLEMS · p. 12
B.1 MOLECULAR PREDICTION
This section describes nine benchmark tasks and their initial algorithms, spanning molecular prediction and translation, geometry, PDE solving, clinical forecasting, microscopy, RNA degradation, polymers, and patent semantics.
B.1 MOLECULAR PREDICTION
- Molecular Prediction uses scaffold-split SIDER data to assess graph-rationale GNN generalization, with environment replacement augmenting latent-space examples.The task is evaluated by ROC AUC.Source: §B.1 · pp. 12–13
B.2 MOLECULAR TRANSLATION
- Molecular Translation maps molecular images to InChI strings using a ResNet-GRU sequence model, evaluated by mean Levenshtein distance.Source: §B.2 · p. 14
B.3 CIRCLE PACKING
- Circle Packing maximizes the radii sum for 26–32 circles in a unit square, using constrained SLSQP optimization that may return numerically invalid packings.Source: §B.3 · p. 14
B.4 BURGERS’ EQUATION
- Burgers’ Equation uses a one-dimensional viscous PDE solver on a periodic domain, with explicit Euler time stepping and finite-difference flux and diffusion terms.The time step is bounded by 0.2 ∆x^2/ν for stability.Source: §B.4 · pp. 14–16
B.5 PARKINSON’S DISEASE
- Parkinson’s Disease predicts current and 6-, 12-, and 24-month MDS-UPDRS scores from longitudinal CSF protein and peptide data using LightGBM and a neural network.The models’ outputs are averaged; performance is measured by SMAPE.Source: §B.5 · p. 16
B.6 NUCLEI IMAGE; B.7 OPEN VACCINE
- Nuclei Image segments microscopy masks with a U-Net, while Open Vaccine predicts position-specific RNA degradation rates using structure-aware graphs and a bidirectional GRU.Nuclei performance uses mean average precision across IoU thresholds; RNA predictions use MCRMSE.Source: §B.6 · pp. 16–17, §B.7 · pp. 18–19
B.8 POLYMER PREDICTION
- Polymer Prediction estimates five simulated polymer properties from SMILES, while its initial approach adapts graph rationalization and environment replacement.Source: §B.8 · pp. 19–20
B.9 USP P2P
- USP P2P predicts semantic similarity for patent phrase pairs by fine-tuning patent-pretrained BERT with a regression layer and technical context.Evaluation uses Pearson correlation.Source: §B.9 · pp. 20–21
C.1 SET-UPS
The setup specifies task-specific prompts and budgets, while the evolution history traces successive algorithmic changes across the benchmark tasks, including some ideas not implemented or used in the workflow.
C.1 SET-UPS
- Task prompts set iteration limits, compute or runtime budgets, evaluation constraints, and task-specific instructions across the benchmarks.Examples include a 30-minute budget for molecular translation and a 2080 Ti GPU limit for Burgers’ Equation.Source: §C.1 · pp. 21–23
C.2 SUMMARY OF ALGORITHMIC EVOLUTION HISTORY
- Molecular Prediction evolves from contrastive learning and adaptive loss weighting to motif-aware masking, reconstruction, and dual-phase adversarial training.Source: §C.2 · p. 23
- Molecular Translation progresses from a frozen ViT–GPT-2 model and molecule-aware tokenization to augmentation, grammar-constrained decoding, and a loss-weighting scheduler.The reported soft-edit-distance loss is a placeholder in the code.Source: §C.2 · p. 23
- Circle Packing develops from power-diagram placement and optimization toward multiple starts, controlled repairs, improved neighbor detection, and constraint certification.Source: §C.2 · p. 23
- Burgers’ Equation evolves from GPU-accelerated explicit finite differences to spectral methods and advanced ϕ-function evaluation, though some described components were not executed or were placeholders.Source: §C.2 · p. 23
- Parkinson’s Disease explores Neural CDEs, meta-learning, and PINN-inspired regularization; Nuclei Image adds PointRend and uncertainty-guided selective refinement.The wavelet preprocessing idea for Parkinson’s was not implemented; nuclei self-distillation was not used because no teacher model was available.Source: §C.2 · pp. 23–24
- Open Vaccine adds RNA-structure features, dynamic target-loss weighting, and self-supervised transformer embeddings, with the embeddings component described as a placeholder.Source: §C.2 · p. 24
- Polymer Prediction explores polymer-specific message passing and a degree-of-polymerization auxiliary loss, while USP P2P combines ordinal and contrastive objectives with CPC embeddings.The polymer data contain only one repeating unit; some proposed polymer parsing or pooling ideas were unsupported, unused, or unimplemented.Source: §C.2 · p. 24
C.3 DEEPEVOLVE PROPOSED ALGORITHM CODE FOR THE MOLECULAR PREDICTION TASK
The molecular-prediction code adds contrastive InfoNCE loss and motif masking to the model forward function, with adaptive motif selection informed by sampling uncertainty.
C.3 DEEPEVOLVE PROPOSED ALGORITHM CODE FOR THE MOLECULAR PREDICTION TASK
- The molecular-prediction forward function adds InfoNCE loss and motif masking as two components.Source: §C.3 · p. 24
- The InfoNCE implementation computes loss using batch negatives and can use supplied negatives for an alternate sampling path.Source: §C.3 · p. 24
- Motif masking samples motif probabilities with Gumbel-Softmax and uses their mean and variance to form uncertainty-aware masks.The code sets an adaptive uncertainty threshold of 0.05.Source: §C.3 · p. 25