AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery
Junhao Qiu, Qinglong Hu, Xialiang Tong, Mingxuan Yuan, Liyong Lin, Qingfu Zhang
cs.AI
Quick verdict
A well-motivated agentic framework for LLM-driven algorithm design whose sample-efficiency story is real but measured in evaluations only, while its winning margins on two of six tasks sit inside seed noise.
Summary
AlgoEvo reframes automated algorithm discovery as an agent-driven search loop instead of a fixed evolutionary pipeline. An autonomous agent interleaves reading, editing, diagnosing and evaluating code, guided by two mechanisms. First, a design skill hub externalises paradigm-specific knowledge: each skill declares the roles, code interfaces, modification principles and evaluation conventions for a paradigm, so the same engine handles single-heuristic, multi-objective and multi-component design by swapping the activated skill. Second, a hierarchical experience bank distils each evaluation into a structured experience card, organises cards from one task into an MCTS-style tree selected by situational UCB (Eq. 5), and promotes design patterns observed in at least two tasks back into the skill hub. Four search situations (stagnation, sparse front, bottleneck, coupling) with fixed thresholds bias retrieval. Experiments cover six tasks (TSP, CVRP, Bi-TSP, Bi-FJSP, CVRP-DR, FJSP 4-Ops) against FunSearch, EoH, ReEvo, MCTS-AHD, MEoH, NSGA-II, MOEA/D, MOTIF and E2OC under a shared 500-evaluation allowance and three seeds. AlgoEvo reaches its best TSP training cost (5.986) in about 39 evaluations versus the baselines' 500, and leads CVRP-DR and FJSP 4-Ops held-out results. But token savings are far smaller than evaluation savings (2.9M vs 4.1M on TSP), the Bi-FJSP hypervolume margin (0.911 vs 0.903) sits inside a 0.083 standard deviation, multi-objective results are reported on training instances only, and no significance test appears anywhere.
Review complete
Assessment at a glance
7
risk flags identified
● high 2
● medium 3
● low 2
5 strengths · 6 weaknesses · 6 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Replacing predefined evolutionary pipelines with an autonomous agent that decides when to read, diagnose, edit and evaluate yields better search control than static prompt-driven sampling.
- A pluggable design skill hub lets a single discovery engine span single-objective, multi-objective and multi-component algorithm design without structural modification.
- A hierarchical experience bank (cards, task-level MCTS tree with situational UCB, cross-task skill promotion) produces intra-task accumulation and cross-task transfer.
- AlgoEvo matches or surpasses FunSearch, EoH, ReEvo, MCTS-AHD, MEoH, NSGA-II, MOEA/D, MOTIF and E2OC on six benchmark tasks with substantially fewer evaluations and reduced token consumption.
- Transferring executable code from a previous run (warm start) beats injecting abstract strategy descriptions, reducing TSP mean cost from 4.961 to 4.808.
Strengths
The paradigm-decoupling argument is concrete, not rhetorical: one engine, six tasks, three paradigms, changed only by swapping the activated design skill.
Section 3.2 defines a design skill as declaring (i) strategy roles, (ii) code interfaces, (iii) modification principles and (iv) evaluation conventions, and states the search mechanism is unchanged. Table 1 confirms the same engine run on TSP/CVRP (K=1), Bi-TSP/Bi-FJSP (multi-objective) and CVRP-DR/FJSP 4-Ops (K=3 and K=4), with the evolvable units and fixed surrounding solver enumerated per task in Appendix D (Table 8) and Algorithms 3-4.
The sample-efficiency result on the two constructive tasks is large and internally corroborated.
AlgoEvo reaches TSP train 5.986 +/- 0.11 / test 7.007 +/- 0.04 using 39 evaluations, and CVRP train 12.698 / test 24.150 using 35, against baselines that consume the full 500-evaluation allowance and finish behind (best baseline MCTS-AHD 6.331 train / 7.337 test on TSP). Figure 5 shows the best-so-far curves flatten within the first tens of evaluations on all four single-unit tasks, and Section 4.3 explicitly rules out early stopping as the explanation.
Diagnosis is operationalised with fixed, checkable thresholds rather than left to the LLM.
Multi-component co-design results are the strongest part of the table and are reported honestly.
Reproducibility provisions are unusually explicit for this literature.
Weaknesses
Evaluation count is not a cost measure, and the token savings are an order of magnitude smaller than the evaluation savings.
Reported metrics are not comparable across tables: three different hypervolume scales appear for nominally multi-objective TSP/FJSP tasks.
Two of six headline wins are inside seed noise, and no significance test is reported anywhere.
Half the benchmark has no held-out evaluation, contradicting the paper's own train/test framing.
The ablation does not establish that either knowledge mechanism matters.
The multi-component baselines are partly self-implemented and the classical-solver baseline is absent.
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
AlgoEvo (arXiv 2609.15820, City University of Hong Kong, Huawei Noah's Ark Lab, A*STAR) targets automated algorithm discovery, the problem of synthesising executable heuristics for optimisation problems without manual algorithm engineering. Its diagnosis of the field is clear and largely fair: existing LLM-AHD systems (EoH, FunSearch, ReEvo, MCTS-AHD, MEoH, MOTIF, E2OC) embed the model inside a rigid, expert-predefined control flow, couple the framework to one paradigm, and feed back only a scalar score. AlgoEvo proposes three changes: an autonomous agent that decides when to read, diagnose, edit or evaluate; a design skill hub that externalises paradigm-specific knowledge so one engine spans single-heuristic, multi-objective and multi-component design; and a hierarchical experience bank that distils evaluations into cards, organises them into a task-level MCTS tree and promotes cross-task patterns back into the skill hub.
The claimed outcome is that AlgoEvo 'matches or surpasses specialized methods with substantially fewer evaluations and reduced token consumption' across six benchmark tasks. This review finds the first half of that claim largely supported for four tasks, and the second half considerably weaker than stated.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 6 weaknesses, analyzed in detail
- 7 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 6 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Is AlgoEvo really more efficient, or does it just do fewer evaluations?
Both, but the two effects differ by an order of magnitude. On TSP, AlgoEvo uses 39 evaluations where EoH, ReEvo, MCTS-AHD and FunSearch each use 500, and it still wins on quality (train 5.986 versus 6.309-6.465; test 7.007 versus 7.337-7.699). But token consumption only falls from 4.1M (EoH) or 3.2M (FunSearch) to 2.9M, because the agent spends most of its steps on non-evaluation operations. Table 12 shows one CVRP-DR run performing 163 diagnose, 122 edit and 14 read operations alongside 81 evaluations. Evaluation count is therefore not a cost measure.
Does AlgoEvo beat every baseline on all six tasks?
No. It leads on TSP and CVRP (single heuristic), on both splits of CVRP-DR, and on the held-out split of FJSP 4-Ops. On Bi-FJSP it has the highest hypervolume (0.911 versus MEoH 0.903) but a worse IGD (0.162 versus 0.143), and the HV margin is smaller than the reported standard deviation of 0.083. On FJSP 4-Ops training it is third (646.4) behind Rotating-EoH (645.10) and Synergy (646.10). All comparisons use three random seeds with no significance testing.
What makes AlgoEvo different from FunSearch or EoH?
FunSearch and EoH embed the LLM in a fixed loop (island-model sampling, generate-then-evaluate cycles) with scalar fitness feedback. AlgoEvo lets an autonomous agent choose its next action among read, edit, diagnose, analyse-interaction and evaluate, adds a design skill hub that carries paradigm-specific interfaces and principles so one engine serves single-objective, multi-objective and multi-component tasks, and keeps a hierarchical experience bank that turns each evaluation into a card in an MCTS-style tree selected by situational UCB.
Are the reported results reproducible?
Substantially but not fully. Appendix C publishes instance distributions (Table 7), the 500-evaluation budget, three seeds and the HV/IGD normalisation formulas, and Appendix H states complete source code is provided, with instances, seeds and evaluation scripts publicly released. Two obstacles remain: the agent loop requires paid access to gpt-4o-mini (TSP, CVRP) and DeepSeek-V4-Flash (other tasks), and two of the four multi-component baselines (Synergy, Rotating-EoH) are the authors' own implementations rather than published code.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- I read the complete text (all 87,199 characters, including Appendices A-H), so no substantive content was missed.
- Figures 3-9 are described in extracted text but could not be visually inspected; claims about convergence curves and Pareto-archive coverage rest on the paper's own captions and numbers.
- Several cited works are 2026 preprints or vendor items that cannot be verified offline; their descriptions are taken at face value from the paper.
- The token column in Table 1 uses a superscript 'e' for values inferred from baseline runs; the extracted text did not preserve which specific cells carry that marker, so I could not distinguish measured from inferred token figures.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.