Exploration Structure in LLM Agents for Multi-File Change Localization

2026-06-10 · Source: Artificial Intelligence · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

A new study investigates exploration structures for LLM-based agents tasked with multi-file change localization in software issues, addressing the inefficiency of linear repository exploration for changes spanning multiple subsystems. The research compares linear sequential exploration against a novel non-linear, domain-scoped parallel agentic approach. Using SWE Bench Pro and an expanded benchmark including recent PRs from 2025 and 2026, the study found that domain-scoped parallel agent spawning with a small Haiku-class model achieved the highest micro F1 among Haiku models. This approach was second only to the larger Codex 5.5 High on the expanded benchmark. On the original 2020 SWE-bench Pro, a Sonnet plain LLM baseline showed higher micro F1 by predicting fewer files, yielding high precision but low recall. Additional findings highlight documentation evolution as an unresolved latent dependency, the degradation of localization from naive file system access due to test-file overprediction, and the ineffectiveness and high token cost of forced multi-agent consultation.

Key takeaway

For AI Engineers developing LLM agents for multi-file code changes, prioritize implementing non-linear, domain-scoped parallel exploration over traditional linear methods. Your agents will achieve higher localization accuracy, as demonstrated by Haiku-class models. Additionally, avoid naive file system access that can lead to test-file overprediction, and be wary of forced multi-agent consultation, which significantly increases token costs without measurable benefit.

Key insights

Non-linear, domain-scoped parallel exploration significantly improves LLM agent performance for multi-file change localization.

Principles

• Linear exploration mismatches multi-subsystem changes.
• Naive file system access degrades localization.
• Forced multi-agent consultation is ineffective.

Method

Spawn parallel agents, each exploring a specific domain concurrently, to localize multi-file changes.

In practice

• Adopt non-linear exploration for LLM agents.
• Guard against test-file overprediction.
• Reconsider multi-agent consultation costs.

Topics

• LLM Agents
• Multi-File Change Localization
• Software Engineering
• Exploration Structures
• SWE Bench Pro
• Haiku Model

Best for: Research Scientist, AI Scientist, AI Engineer, Machine Learning Engineer

Related on AIssential

• Exploration Structure in LLM Agents for Multi-File Change Localization — Linear LLM agent exploration is structurally mismatched for multi-subsystem software changes, benefiting from domain-scoped parallel approaches.
• SWE-Explore: Benchmarking How Coding Agents Explore Repositories — SWE-Explore benchmarks coding agents' line-level repository exploration, isolating it from end-to-end repair outcomes.
• Cross-Lingual Token Arbitrage: Optimizing Code Agent Context Windows via Local LLM Preprocessing — Proactive, edge-side prompt rewriting with a local LLM significantly reduces token costs for coding agents.
• Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agents — HORMA uses a hierarchical, file-system-like memory to enable efficient, detailed context retrieval for LLM agents in long-horizon tasks.
• Context-as-a-Service: Surfacing Cross-File Dependency Chains for LLM-Generated Developer Documentation — CaaS enhances LLM documentation agents by surfacing non-obvious cross-file dependencies, improving accuracy and efficiency.
• The Infrastructure Behind Making Local LLM Agents Actually Useful — Local LLM agents require deliberate infrastructure for performance and long-session stability beyond basic model serving.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.