MOAE: Multi-Objective Agent Evolution with Pareto-Preserving Search
Hengle Jiang, Qijun Cai, Ziying Luo, Ke Tang
cs.AI · cs.LG
Quick verdict
MOAE applies Pareto-preserving evolutionary search to per-query LLM-agent optimization; gains look real but are modest and partly within CI overlap.
Summary
The paper introduces MOAE, a training-free framework that treats per-query improvement of a frozen LLM agent as a small multi-objective evolutionary search over compact execution policies. Given a budget of seven complete rollouts per query, MOAE maintains an archive of empirically non-dominated candidates, selects objective-specialized parents under a retention guard protecting the primary task objective, mutates policies using evaluator diagnostics, and applies a constraint-aware, lexicographic deployment rule only after search. Objectives are task performance (weighted constraint satisfaction), trajectory quality (execution validity, coverage, evidence support, discipline), and a probe-based safety risk score, evaluated on TravelPlanner (180 validation queries; Gemma-4-31B-it, Qwen3-30B-A3B, DeepSeek-V4-Pro) and AgentDojo. Full MOAE improves over the base agent on all three objectives across models (e.g., Qwen3 task 0.493→0.585, risk 0.533→0.400), beats Weighted-Sum on task and trajectory for all models, reduces attack success to zero where baselines were nonzero, and shows better search behavior (normalized hypervolume gain 0.52 vs 0.43). Ablations attribute the largest effect to directed mutation. Limitations include main results on the validation split, overlapping confidence intervals for several mechanism comparisons, evaluator-coupling and metric-overfitting risk, baselines available only as matched-budget online adaptations, and roughly sevenfold rollout cost per query.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under a matched seven-rollout budget, MOAE jointly improves task performance, trajectory quality, and safety over the unoptimized base agent across three LLMs on TravelPlanner (Tables 1 and 8).
- Pareto-preserving search expands the attainable objective region (normalized hypervolume gain 0.52 vs 0.43 for the strongest alternative) and raises joint-improvement query rate from 9.4% to 13.9% (Figure 3).
- A single MOAE search retains objective-specialized operating points (MOAE-Task/-Trajectory/-Safety) so deployment preferences can be applied afterward without re-running optimization (Table 1).
- The same procedure transfers to AgentDojo, improving utility and trajectory quality and reducing observed attack success to zero where baselines were nonzero (Table 2).
- Ablations attribute the largest effect to directed mutation, with the archive mattering mainly when objectives compete (Table 3).
Strengths
Clean separation of search-time preservation from deployment-time preference
The archive (Eq. 5, 19) retains incomparable candidates while the constraint-aware rule (Eq. 20–21) filters candidates no worse than the base rollout in task and safety, then applies a lexicographic preference — with c0 as guaranteed fallback, actually invoked on 79/180 TravelPlanner queries, an honest disclosure of how often search wins.
Matched-budget comparison with internally fair baselines
Every search method gets seven complete rollouts per query on the same agent scaffold, evaluator, and deployment rule; Weighted-Sum, Best-of-7, and online adaptations of GEPA, MOCHA, and EvoAgent are re-implemented in the identical per-query policy space (Appendix 'Implementation of compared methods').
Above-average statistical and robustness protocol
Unusually candid limitations and failure-mode analysis
Mechanism-level evidence, not just end metrics
Weaknesses
Several headline margins sit within or near CI overlap
Main results on the validation split with calibration possibly sharing that split
Preference commitment is postponed, not removed
Adapted baselines limit external validity
Substantial cost asymmetry surfaced only in the appendix
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MOAE (arXiv:2609.05992, SUSTech) tackles per-query optimization of frozen LLM agents under a limited rollout budget, arguing that fixed scalarization during search can discard useful deployment trade-offs. Central claims: (1) framing per-query refinement as Pareto-preserving evolutionary search over compact execution policies jointly improves task performance, trajectory quality, and safety across three models on TravelPlanner; (2) the procedure transfers to AgentDojo; (3) Pareto preservation measurably changes search behavior; (4) mutation, diagnostics, and the archive each contribute (Table 3). Claims are mostly carefully scoped; the abstract's "consistently improves" is the strongest wording and is supported at the mean level but not always by shown significance.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 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
What is MOAE in one sentence?
MOAE is a training-free method that improves a frozen LLM agent per query by running a small evolutionary search (seven rollouts) that keeps non-dominated candidates across task performance, trajectory quality, and safety, and applies deployment preferences only after the search ends.
Does MOAE require fine-tuning the LLM?
No. Both the agent model and the mutation model stay frozen; only a compact textual execution policy evolves. The method needs no parameter updates but costs about seven agent rollouts per query instead of one.
What benchmarks and models does the paper use?
TravelPlanner (all 180 validation queries) with Gemma-4-31B-it, Qwen3-30B-A3B, and DeepSeek-V4-Pro, plus AgentDojo v1.2.2 in Tool Knowledge, direct-injection, and official-attack settings with Gemma and Qwen3.
Does zero attack success rate mean the agent is safe?
No. The authors state that ASR=0 only means no successful attack was observed in the 32–64 evaluated cases under the tested attack families; adaptive or unseen injection strategies, compromised tools, and multi-turn attacks are not covered by the result.
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.
- Confidence intervals for several comparisons (ablations, some baseline cells) are stated to be in supplementary material that is not part of the extracted text, so per-comparison significance could not be verified.
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.