PaperlayerView exact revision on arXiv

Source-linked AI summary

Semantic Navigation for Issue Localization in Code Repository

Yunxiang Wei, Zhenyu Lei, Jundong Li

arXiv:2609.31176v1cs.AI

TL;DR

Repository issue localization must rank relevant files and functions, yet existing environments make agents search for unresolved relations, reconstruct entity meaning, and revise candidates without recorded evidence. SemNav seeds broad candidates with deterministic retrieval and supports iterative refinement through a Semantic Navigation Graph, Semantic Cards, and a Candidate Workspace. It outperforms existing methods across SWE-bench Lite and PLocBench, with Gemma4 4B File Hit@10 rising from 68.33% to 82.67% on SWE-bench Lite.

Abstract

from arXiv · show

Repository-level issue localization aims to identify and rank the files and functions relevant to resolving a reported issue. LLM agents approach this task iteratively: they identify a set of potentially relevant locations, inspect the corresponding code, and revise their judgments about these candidates as new evidence is acquired. Existing environments, however, provide limited support for this loop: agents must search for unresolved relation targets, reconstruct entity semantics from raw source code, and revise candidates without evidential basis. To address these limitations, we present SemNav, a framework that leverages deterministic retrieval to seed a broad candidate set and an LLM agent to continually refine that set, thereby combining initial coverage with evidence-guided revision. SemNav supports this process through three key components. A Semantic Navigation Graph resolves program relations on demand through a language server, enabling direct navigation to related entities across files. Issue-conditioned Semantic Cards provide compact, source-grounded interpretations of each entity's role and relevance to the issue. A persistent Candidate Workspace records each candidate together with its evidential basis, enabling grounded verification, revision, and ranking. Across SWE-bench Lite and PLocBench, SemNav outperforms existing baselines, improving File Hit@10 from 68.33\% to 82.67\% with Gemma 4B. Component ablations and trajectory analysis support the complementary roles of all three components, while Semantic Cards reduce working-context load by 48.2\% relative to full-source reading. SemNav further ranks first on all seven evidence-quality metrics on SWE-Explore and improves downstream issue resolution from 44.00\% to 52.33\%.

Introduction

Issue localization must narrow an oversized repository to relevant files and functions while preserving evidence that guides iterative candidate revision. SemNav supports this process with semantic navigation, compact entity interpretations, and an explicit evidential basis for candidate judgments.

Introduction

  • Repositories often exceed what an agent can inspect, so issue localization must rank potentially relevant files and functions that shape subsequent reasoning and editing.Source: Introduction · p. 1
  • Workflow-driven methods offer broad coverage but prune locations irreversibly, while agent-driven methods adapt search yet tie candidate rationales to a path-dependent trajectory.Source: Introduction · p. 1
  • SemNav combines broad deterministic retrieval with an LLM agent that revises candidates as repository evidence accumulates, making localization an evidence-guided loop.Source: Introduction · p. 2
  • Its Semantic Navigation Graph resolves relation targets online, Semantic Cards compactly interpret entities, and the Candidate Workspace preserves each candidate’s evidential basis for revision.Source: Introduction · p. 2
  • Across SWE-bench Lite and PLocBench, SemNav outperforms existing methods; with Gemma4 4B, SWE-bench Lite File Hit@10 rises from 68.33% to 82.67%.Source: Introduction · p. 2
  • 79.99% line-level precision (up from 53.35%) accompanies SemNav’s first-place rank on all seven SWE-Explore evidence-quality metrics; Qwen3.5 35B issue resolution rises from 44.00% to 52.33%.Source: Introduction · p. 2

Methodology

SemNav seeds a broad candidate set through deterministic retrieval, then lets a Localization Agent navigate repository evidence and revise its ranking in a persistent Workspace.

Background

  • ASTs identify repository entities and syntactic occurrences, while language servers can resolve symbols project-wide when static project information permits.Language-server results may remain unresolved for dynamic behavior or insufficiently constrained symbols.Source: Background · p. 2

Overview

  • The pipeline initializes an SNG and retrieval index offline, then seeds a Candidate Workspace and lets an agent investigate and revise candidates online.Language-server queries resolve requested relation targets and update the graph during investigation.Source: Overview · p. 3

Semantic Navigation Graph

  • The SNG represents repository entities and typed relations, retaining unresolved relation sites for on-demand language-server resolution during navigation.AST analysis initializes the graph; Access uses go-to-definition and find-references to resolve forward and reverse relations.Source: Semantic Navigation Graph · pp. 3–4

Semantic Navigation and Semantic Cards

  • Search locates entities, Expand follows typed relations, and Read provides structural summaries, source code, or issue-conditioned Semantic Cards.Cards combine an entity’s definition, behavior, issue role, and supporting snippets; the agent can request source code when more implementation detail is needed.Source: Semantic Navigation and Semantic Cards · p. 4

Candidate Workspace

  • The persistent Workspace orders candidate records by priority and preserves each candidate’s status and evidence-backed rationale as the agent updates and compares them.Submission places confirmed plausible candidates first, fills remaining positions with the highest-ranked unverified candidates, and excludes discarded candidates.Source: Candidate Workspace · pp. 5–6

Experiments

SemNav improves ranked localization across SWE-bench Lite and PLocBench, while component analyses show complementary contributions from broad seeding, agent investigation, the Candidate Workspace, and Semantic Cards. It also ranks first on all seven SWE-Explore evidence metrics and improves downstream repair outcomes with two repair models.

Experimental Setup

  • The evaluation covers 300 SWE-bench Lite issues, 821 PLocBench issues, and 368 SWE-Explore instances, using Gemma 4B for cross-benchmark comparisons.SWE-bench Lite is also evaluated with Qwen3.5 9B and Qwen3.5 35B-A3B.Source: Experimental Setup · p. 6

Overall Localization Performance

  • SemNav ranks first in 24 of 27 SWE-bench Lite model–metric settings and on all five PLocBench metrics, raising Gemma 4B Function Recall@10 from 32.92 to 40.72 over the strongest baseline.On PLocBench, Acc@5 rises from 60.41 to 67.36 and Macro Recall@5 from 42.07 to 47.45.Source: Overall Localization Performance · p. 6

Component Contributions

  • Without Semantic Cards or online resolution, File Hit@10 falls by 6.00 or 7.00 points, respectively, while average output size remains unchanged.The Seed supports coverage while agent investigation improves ranking; removing the Seed lowers Function Recall@10 by 13.20 points, and removing the Workspace lowers it by 19.20 points.Source: Component Contributions · p. 6
  • Semantic Cards reduce working context by 48.2% versus full-source reading, while total cost remains 25.4% lower than full-source reading and 4.8% above source truncation.The context comparison replays identical SemNav trajectories and tool actions while varying only Read-tool observations.Source: Component Contributions · p. 6

Evidence Acquisition Quality

  • SemNav ranks first on all seven SWE-Explore metrics, with line-level precision rising from 53.35% to 79.99% while Recall@500 also improves.The reported pattern indicates higher precision without narrowing evidence coverage.Source: Evidence Acquisition Quality · p. 6, Figure 3 · p. 7

Downstream Repair Utility

  • On SWE-bench Lite, SemNav raises resolved rate over the strongest baseline by 4.33 points with Gemma 4B and 8.33 points with Qwen3.5 35B-A3B.It also achieves the highest patch application rate; results remain below oracle context with both repair models.Source: Downstream Repair Utility · p. 7, Table 3 · p. 7

Related Work

Prior localization methods organize repository exploration through predefined workflows or agent-directed search, while repository representations range from code fragments to entity graphs and semantic descriptions. SemNav combines broad deterministic initialization with evidence-backed revision and unifies resolved relations with issue-conditioned entity interpretations.

Repository-Level Issue Localization

  • Workflow-driven methods retrieve, filter, and rerank locations, while agent-driven methods acquire evidence through repository search and navigation; SemNav combines deterministic initialization with agent revision.The described prior approaches include hierarchical localization, lexical and learned retrieval, task-specific reranking, call-graph search, AST search, graph traversal, language-server navigation, and shell tools.Source: Repository-Level Issue Localization · p. 7

Repository Representation and Navigation

  • Repository representations include code fragments, entity graphs, syntax-derived relations, and natural-language entity descriptions, but relation targets may remain ambiguous and evidence types separate.SemNav unifies on-demand language-server-resolved relations and issue-conditioned Semantic Cards around shared graph entities, enabling reuse of both evidence types throughout localization.Source: Repository Representation and Navigation · pp. 7–8

Conclusion

SemNav combines broad deterministic retrieval with agent-driven investigation through a Semantic Navigation Graph, Semantic Cards, and a Candidate Workspace. Experiments show improved localization accuracy and evidence quality, reduced agent context requirements, and better downstream issue resolution.

Conclusion

  • SemNav makes repository relations, entity interpretations, and evidence-grounded candidate judgments explicit and revisable throughout localization.The framework combines broad deterministic retrieval with agent-driven investigation; experiments report improved localization accuracy and evidence quality, reduced context requirements, and better downstream issue resolution.Source: Conclusion · p. 8
Loading 2609.31176v1…