CoChat

When Does Diversity Rescue Evolutionary Search of LLM Agent Strategies? A Controlled Landscape Study of Monoculture Collapse and the Novelty Escape

Open access
ALOBED, MOHAMMAD · 2026 · SSRN Electronic Journal

The DOI matches its Crossref record; 19 of the 19 references it cites were found in a registry or an index.

Abstract

Population-based and evolutionary search is increasingly used to design LLM-agent strategies, motivated by the promise of robustness through behavioural diversity — yet such populations tend to lose diversity first. We ask when maintaining behavioural diversity rescues evolutionary search of natural-language agent strategies, and when it does not. To isolate search dynamics from the cost and noise of LLM rollouts, we use a fast deterministic surrogate: deceptive two-basin and tunable NK fitness landscapes that share the combinatorial structure of a modular agent-strategy genome. We run paired, equal-budget comparisons of eight selection strategies across up to 28 seeds, reporting Wilson intervals and paired-bootstrap tests. Objective-only selection escapes the deceptive trap in 11% of runs and collapses behavioural diversity from 0.75 to 0.17, whereas adaptive novelty-blending and novelty with local competition escape 46% and 43% of the time (p=0.0015, p=0.0007) while retaining 0.65-0.67 diversity. Four further results delimit the mechanism: a learned behaviour space (AURORA) backfires when the descriptor is already task-aligned; the advantage appears only as epistasis rises (K=6); it vanishes when a landscape offers multiple routes upward; and a real-LLM probe shows the effect requires both a fitness gradient and behavioural variance, neither of which two natural tasks supplied. Diversity maintenance is therefore a targeted remedy, not a universal one. We map where it helps and where it cannot apply, and pre-register an engineered-deception real-LLM experiment as the decisive transfer test. All data, generators, and protocols are released; no agent-benchmark results are claimed.

Overview

Integrity signals

Registry

Crossref · 10.2139/ssrn.7291548

Cites

19 of 19 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Verification

Claims the paper makes

We couldn’t retrieve this paper’s full text, so its claims and self-check weren’t run. The registry and citation checks still apply.

Citations

Cites

Show all 19 references

Research works printed but not linked: 14

  1. Maxence Bahlous-Boldi, 2025, Dominated novelty search: Parameter-free quality-diversity
  2. Hao-An, 2025, A survey of self-evolving agents
  3. Qingyan Guo, Connecting large language models with evolutionary algorithms yields powerful prompt optimizers, International Conference on Learning Representations (ICLR)
  4. John H Holland, 1975, Adaptation in Natural and Artificial Systems
  5. Shengran Hu, Automated design of agentic systems, International Conference on Learning Representations (ICLR)
  6. A Stuart, 1993, The Origins of Order: Self-Organization and Selection in Evolution
  7. Omar Khattab, Dspy: Compiling declarative language model calls into self-improving pipelines, International Conference on Learning Representations (ICLR)
  8. Sung Joon, Generative agents: Interactive simulacra of human behavior, Proceedings of the ACM Symposium on User Interface Software and Technology
  9. Timo Schick, Toolformer: Language models can teach themselves to use tools, Advances in Neural Information Processing Systems (NeurIPS)
  10. Guanzhi Wang, Voyager: An open-ended embodied agent with large language models, Transactions on Machine Learning Research
  11. Xuezhi Wang, Self-consistency improves chain of thought reasoning in language models, International Conference on Learning Representations (ICLR)
  12. Jason Wei, Chain-of-thought prompting elicits reasoning in large language models, Advances in Neural Information Processing Systems (NeurIPS)
  13. Shunyu Yao, React: Synergizing reasoning and acting in language models, International Conference on Learning Representations (ICLR)
  14. Jiayi Zhang, Aflow: Automating agentic workflow generation, International Conference on Learning Representations (ICLR)

Metadata

Authors
ALOBED, M.
Affiliations
Idlib University (SY)
Corresponding author
ALOBED, MOHAMMAD
Venue
SSRN Electronic Journal
Publisher
Elsevier BV
Published
2026
ISSN
1556-5068 (ISSN-L)
OpenAlex
W7203539670
Cited by
0
References
6
Type
posted content
Language
en
Open access
Locations
SSRN Electronic Journal (acceptedVersion)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Evolutionary Algorithms and Applications
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Evolution and Genetic Dynamics
  • Physical Sciences › Computer Science › Artificial Intelligence › Metaheuristic Optimization Algorithms Research
Keywords
behavioral diversity, evolutionary search, NK fitness landscapes, selection strategies, epistasis, diversity maintenance

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.