🛰️ Daily AI Frontier
‹ back to 2026-08-23

Loreley: Repository-Scale Program Evolution with Quality-Diversity Search

Research LLM Agents

Ranking

Overall 82
Content 100
Popularity 41

Observed public metrics from 1 member.

Merged summary

TL;DR - Loreley applies quality-diversity search to repository-scale software evolution by retaining and reusing complete Git states rather than following one edit sequence or repeatedly restarting. It successfully preserved stepping-stone branches, but did not demonstrate better final performance within a 48-job budget.

  • Candidates are Git commits created in isolated worktrees and scored by a project-defined evaluator.
  • A controlled Zstandard experiment evaluated 1,008 candidates across Loreley, sequential champion editing, and independent root proposals.
  • At 48 jobs, Loreley scored 0.135% below sequential editing and 0.320% above independent proposals; neither comparison established a QD advantage.
  • Non-incumbent states contributed to four of seven final Loreley winners by primary-parent ancestry, showing that archive reuse occurred even without a proven endpoint benefit.

Sources (1)

Loreley: Repository-Scale Program Evolution with Quality-Diversity Search

arXiv cs.SE Mohan Chen 2026-08-20 arXiv:2608.19703
Public signals Semantic Scholar citations 0 · Semantic Scholar influential citations 0
Providers: Hugging Face · N/A OpenAlex · N/A Publisher · N/A Semantic Scholar · Citations 0 · Influential citations 0 X · N/A Fetched 2026-09-10 14:24:24.630473 UTC

TL;DR - Loreley applies quality-diversity search to repository-scale software evolution by retaining and reusing complete Git states rather than following one edit sequence or repeatedly restarting. It successfully preserved stepping-stone branches, but did not demonstrate better final performance within a 48-job budget.

  • Candidates are Git commits created in isolated worktrees and scored by a project-defined evaluator.
  • A controlled Zstandard experiment evaluated 1,008 candidates across Loreley, sequential champion editing, and independent root proposals.
  • At 48 jobs, Loreley scored 0.135% below sequential editing and 0.320% above independent proposals; neither comparison established a QD advantage.
  • Non-incumbent states contributed to four of seven final Loreley winners by primary-parent ancestry, showing that archive reuse occurred even without a proven endpoint benefit.
item →