Parallel Domain-Scoped Exploration Improves LLM Agents' Ability to Locate Files for Software Changes
Researchers compared linear versus non-linear exploration strategies for LLM-based agents tasked with localizing files that need to be changed to resolve software issues. The study found that domain-scoped parallel agent spawning significantly outperforms sequential exploration, particularly for changes spanning multiple subsystems. This matters because more efficient file localization could improve AI-assisted software engineering tools' ability to handle complex, multi-system code changes.
A new arXiv paper evaluates how LLM agents explore software repositories to identify which files need modification for bug fixes and feature requests. The authors argue that most current AI agents use inefficient linear exploration—visiting one directory or file at a time—which is poorly suited for changes affecting multiple subsystems. They tested their non-linear, domain-scoped parallel agent approach against several baselines using the SWE-Bench Pro benchmark and additional GitHub issues from 2025-2026. Their domain-agent system achieved the highest micro F1 score among smaller Haiku-class models by a significant margin and ranked second overall, behind only the much larger Codex 5.5 High model. The research also identified three key limitations: documentation evolution creates unresolved dependencies, naive file system access can introduce test-file prediction errors, and multi-agent consultation increases token costs without measurable benefits.
What's missing
The study's own limitations include: the approach was tested primarily on the Ansible project as an exemplar rather than across diverse codebases; the comparison between Haiku and Sonnet models shows trade-offs between precision and recall that may not generalize; and the paper does not discuss computational costs or latency implications of parallel agent spawning versus sequential approaches.
What different sources said
- arXiv cs.AICenter
Exploration Structure in LLM Agents for Multi-File Change Localization
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.