Exploration Structure in LLM Agents for Multi-File Change Localization

2026-06-10 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A study on LLM agents for multi-file change localization reveals that traditional linear repository exploration is inefficient for changes spanning multiple subsystems. Researchers compared this against a non-linear, domain-scoped parallel agentic exploration approach using SWE Bench Pro (ansible) and an expanded benchmark including 2025/2026 PRs. The domain-scoped parallel agent spawning, utilizing a small Haiku-class model, achieved the highest micro F1 among its class, and was second only to the larger Codex 5.5 High on the expanded benchmark. Additional findings indicate documentation evolution as an unresolved latent dependency, naive file system access degrading localization due to test-file overprediction, and forced multi-agent consultation increasing token cost without measurable benefit.

Key takeaway

For AI Engineers designing LLM agents for software issue resolution, prioritize implementing non-linear, domain-scoped parallel exploration. This approach significantly improves multi-file change localization, outperforming linear methods. You should also avoid naive file system access, which can degrade localization, and recognize that forcing multi-agent consultation increases token costs without measurable performance gains.

Key insights

Linear LLM agent exploration is structurally mismatched for multi-subsystem software changes, benefiting from domain-scoped parallel approaches.

Principles

• Linear exploration mismatches multi-subsystem changes.
• Domain-scoped parallel agents improve localization.
• Forced multi-agent consultation raises cost without benefit.

Method

Construct persistent-session evaluation of GitHub issues anchored at a single base commit, comparing linear versus non-linear domain-scoped parallel agentic exploration.

In practice

• Implement domain-scoped parallel agents for complex changes.
• Avoid naive file system access to prevent overprediction.
• Do not force multi-agent consultation for efficiency.

Topics

• LLM Agents
• Software Engineering
• Change Localization
• SWE Bench Pro
• Domain-scoped Agents
• Haiku-class Models

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

Related on AIssential

• LLM Consortium for Software Design Refinement: A Controlled Experiment on Multi-Agent Collaboration Topologies — Structural adversarial and cross-model review topologies significantly enhance multi-agent LLM software design.
• 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.
• AgentEscapeBench: Evaluating Out-of-Domain Tool-Grounded Reasoning in LLM Agents — LLM agents struggle with long-range tool-grounded reasoning and deep contextual dependencies, unlike local tool use.
• How an Unexpected Reddit Spike Forced Me to Learn Prompt Caching the Hard Way — Multi-agent LLM pipelines offer significant prompt caching benefits, especially when context is passed sequentially between agents.
• TREX: Automating LLM Fine-tuning via Agent-Driven Tree-based Exploration — TREX automates LLM fine-tuning through a multi-agent, tree-based exploration system, outperforming human experts on specific tasks.
• The Amazing Agent Race: Strong Tool Users, Weak Navigators — LLM agents struggle with navigation in complex, non-linear tasks more than with tool execution.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.