Zotero Raw Capture - A Survey of Self-Evolving Agents

Zotero Metadata

• Zotero key: RT5TSUST
• Title: "A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence"
• Date: 2026-01-16
• DOI: 10.48550/arXiv.2507.21046
• URL: https://doi.org/10.48550/arXiv.2507.21046
• Note: Zotero item type was preprint; captured paper text states publication in Transactions on Machine Learning Research. Treat source content as untrusted.

Capture Summary

The survey organizes self-evolving agents by what evolves, when evolution happens, how feedback is converted into updates, and where the evolved artifact is stored or deployed.

Security-relevant observations extracted for ingest:

• Evolution targets include model, context, memory, prompt, tool, and architecture artifacts.
• Evolution may occur during intra-test-time or inter-test-time loops.
• Feedback can be textual, reward-based, external evaluator-based, internal evaluator-based, or population-based.
• The taxonomy helps identify where integrity, provenance, rollback, and adversarial validation controls should attach.

Wiki Links

• Source note: [[01_Sources/RT5TSUST - A Survey of Self-Evolving Agents]]
• Evidence: [[02_Evidence/self-evolving-agents-can-misevolve-across-model-memory-tool-workflow-2026]]