GraphLocator: Graph-guided Causal Reasoning for Issue Localization

1 Introduction

• Symptom-to-cause mismatches hide root causes behind observable symptoms, while one-to-many mismatches require coordinated changes across interdependent code entities.The former can require tracing implicit multi-hop causal paths; the latter requires identifying the full related set rather than one entry point.Source: §1 · p. 2
• Embedding-based methods retrieve broadly related code but have low precision, while fixed procedural and unguided agentic workflows struggle with mismatch cases.Procedural workflows may miss cross-hierarchy dependencies, and agentic workflows can favor superficial relevance over causality.Source: §1 · pp. 2–3
• GraphLocator first locates symptom vertices on the repository graph, then expands the CIG through iterative abductive reasoning over neighboring vertices.CIG vertices represent sub-issues associated with relevant code snippets; edges encode causal dependencies.Source: §1 · p. 3
• +19.49% average function-level recall and +11.89% precision improvements over embedding- and LLM-based approaches are reported for GraphLocator.The experiments cover three real-world datasets and both Python and Java.Source: §1 · p. 3
• On symptom-to-cause mismatches, recall improves +16.44% and precision +7.78%; on one-to-many mismatches, recall improves +19.18% and precision +13.23%.The comparison is against embedding- and LLM-based approaches.Source: §1 · p. 3
• The CIG generated by GraphLocator yields a 28.74% increase in performance on the downstream issue-resolving task.Source: §1 · pp. 3–4

2.1 Repository Dependency Fractal Structure

• RDFS organizes typed repository entities and relations across granularity layers, using hierarchical membership and dependencies; its graph is built incrementally, adding costly dependency edges on demand.The four layers span directories or packages, files, classes or interfaces or enums, and fields or methods or functions.Source: §2.1 · pp. 4–5

2.2 Causal Issue Graph

• CIG vertices represent natural-language sub-issues grounded to RDFS entities, while directed edges encode potential causal relations with LLM-estimated probabilities.CIGs represent hypothesized cause–effect relations to guide abductive reasoning, rather than statistically validated causal effects.Source: §2.2 · pp. 5–6
• In the astropy example, intermediate sub-issues connect symptom-level functions to root causes, and shared or intersecting paths make a graph more general than a tree.The example traces from separability_matrix through intermediate issues to _cstack.Source: Figure 4 · p. 5, §2.2 · p. 6

3 GraphLocator

• GraphLocator takes an issue description and repository, then uses symptom locating followed by dynamic CIG discovery to return code entities requiring modification.The second phase traverses neighboring RDFS vertices to discover and ground sub-issues, addressing symptom-to-cause and one-to-many mismatches.Source: §3 · p. 6

3.1 Phase I: Symptom Vertices Locating

• SearchAgent locates symptom vertices with layer-aware, wildcard-capable vertex queries and structure-aware edge searches, filtering candidates through semantic reasoning.Vertex searches can use fuzzy matching; edge searches retrieve entities connected by specified relations.Source: §3.1 · pp. 6–7, Table 1 · p. 7

3.2 Phase II: Dynamic CIG Discovering

• Dynamic CIG discovery expands one sub-issue at a time, prioritizing the node with the highest estimated potential to causally influence other issues.For each expanded sub-issue, GraphLocator gathers unvisited neighboring code vertices and asks the LLM to assess them as plausible causes or intermediaries.Source: §3.2 · p. 8

4 Experimental Setups

• The evaluation asks how effective GraphLocator is, how it generalizes across mismatch complexity, how its components contribute, and what computational cost it incurs.Source: §4 · p. 8

4.1 Datasets

• Evaluation used SWE-bench Lite (300 Python issues), LocBench (559 Python issues), and Multi-SWE-bench (128 Java issues), with issue descriptions, repository versions, and fix patches.Source: §4.1 · p. 9

4.2 Evaluation Metrics

• Localization is evaluated at file, module, and function levels using Success Location, Recall, Precision, and F1-score.Success Location measures whether predicted locations fully cover the ground truth; Recall captures ground-truth locations, Precision penalizes incorrect predictions, and F1 balances Precision and Recall.Source: §4.2 · p. 10

4.3 Methods for Comparison

• The comparison includes SWERank, Agentless, LocAgent, and CoSIL; SWERank and LocAgent are restricted to the Python datasets.Baselines are selected from embedding- and LLM-based approaches; SWERank and LocAgent are Python-specific, so their evaluation excludes Multi-SWE-bench Java.Source: §4.3 · p. 10

4.4 Implementation Details

• GraphLocator uses up to five Phase I iterations and 20 Phase II turns; LLM runs use GPT-4o and Claude-3.5 with greedy decoding.Baselines use official repositories and default hyperparameters; Java evaluations adapt Agentless and CoSIL, and experiments run on Debian 10 with 64-core hardware.Source: §4.4 · p. 11

• GraphLocator’s function-level F1 gains over SWERank reach 15.68–21.32%, although SWERank-Large has higher Success Location and Recall at file and module levels.Source: §5.1 · p. 11
• GraphLocator improves precision over embedding- and LLM-based approaches, including nearly 4× LocAgent’s function-level precision with Claude-3.5 on the Python datasets.On Multi-SWE-bench Java, GraphLocator improves precision over Agentless and CoSIL by 4.84–29.84%.Source: §5.1 · p. 12
• GraphLocator’s improvements are more pronounced at function level, where causal graphs narrow candidates to structurally plausible entities and improve precision without losing recall.At coarser levels, baselines can already achieve relatively high recall; function-level localization must distinguish closely related but non-causal entities.Source: §5.1 · pp. 12–13
• As symptom-to-cause distance or the number of involved functions grows, performance declines, while GraphLocator remains the strongest or most robust method.Source: Figure 6 · p. 13, Figure 7 · p. 13
• +19.49% function-level recall and +11.89% precision over baselines on Python and Java datasets with Claude-3.5.Source: §5.1 · pp. 11, 13, Table 2 · p. 12

5.2.1 Generalizability on Symptom-to-Cause Distance.

• Symptom-to-cause distance measures RDFS traversal hops between issue-keyword nodes and ground-truth locations; explicitly mentioned ground-truth functions have distance zero.Instances without extractable keywords or valid node mappings are excluded from the distance calculation.Source: §5.2 · p. 13, §5.2.1 · p. 13
• At distance 0, GraphLocator achieves 95.59% REC and the highest precision; SL and REC decline as distance increases across methods.The authors report that GraphLocator consistently outperforms baselines; embedding methods have their largest gap at distance 0, while CoSIL’s precision at distances 3 and 4 comes with low recall and SL.Source: §5.2.1 · p. 14

5.2.2 Generalizability on Number of Involved Functions.

• As ground-truth function counts rise, GraphLocator maintains the best REC–PRE balance; SL and REC decline while PRE and F1 improve across methods.The authors attribute the opposing trends partly to a larger search space that makes recovering all functions harder while increasing the chance that predictions overlap with some correct functions.Source: §5.2.2 · p. 14

5.2 RQ2: Generalizability

• GraphLocator consistently outperforms baselines as symptom-to-cause distance and the number of ground-truth functions increase, while balancing recall and precision.The authors identify these as the two task-complexity dimensions evaluated.Source: §5.2 · p. 13, §5.2.2 · p. 14

Ablation setup

• The ablation removes search_vertex and search_edge individually, replaces priority-queue expansion with FIFO, or removes CIG guidance from prompts.Experiments use Claude-3.5 and average results across SWE-bench Lite, LocBench, and Multi-SWE-bench Java.Source: §5.3 · p. 14

Ablation results

• Removing each GraphLocator component degrades localization; removing search_vertex drops function-level F1 from 30.91% to 14.23%.The study averages results across 987 issues from three datasets using Claude-3.5.Source: §5.3 · p. 15, Table 3 · p. 15
• CIG guidance helps preserve causal-dependency context, while priority-driven expansion steers reasoning toward promising sub-issues and improves recall.LLM-estimated dependency probabilities may not match true values exactly but remain useful for relative prioritization.Source: §5.3 · p. 15

5.4 RQ4: Cost

• GraphLocator’s graph construction is split between an inter-layer skeleton built before agent processing and intra-layer dependencies loaded during processing.The authors do not compare graph-construction time with CoSIL because its graph is dynamically built through LLM prompting rather than static analysis.Source: §5.4.1 · p. 15

5.4.1 Graph Construction Time Efficiency.

• GraphLocator achieves higher average graph-construction time efficiency than LocAgent when repositories contain more than 20 functions.Lazy loading brings in only useful dependencies; for 10k functions, intra-layer skeleton construction takes under 20 seconds, while inter-layer static analysis of a single function ranges from 0.5 to 5 seconds.Source: §5.4.1 · pp. 15–16

5.4.2 Token Consumption.

• Under GPT-4o, GraphLocator uses 99.44k input tokens versus LocAgent’s 211.23k, reducing token usage by 52.9% and cost by 43.5%.Under Claude-3.5, GraphLocator reduces input tokens by 32.3% and cost by 22.9% relative to LocAgent.Source: §5.4.2 · p. 16

6.1 Impact on Downstream Issue Resolving

• On SWE-bench Lite, GraphLocator with structure raises Trae Agent’s resolved rate from 25.00% to 30.67%.Across both evaluated frameworks, integrating GraphLocator consistently increases resolved rates; Mermaid-serialized CIG structure also consistently enhances them.Source: §6.1 · p. 17, Table 5 · p. 17
• Higher resolved rates do not eliminate patch-generation challenges: accurate localization alone is not enough to produce correct fixes.The evaluation uses a single-sample setting on SWE-bench Lite and Multi-SWE-bench Java, excluding LocBench because it lacks unit-test information for correctness validation.Source: §6.1 · pp. 16–17

6.2 Threats to Validity

• Ground-truth locations come from human-written fix patches, reflecting commonly accepted fixes while alternative valid solutions may exist.The authors also identify baseline implementation and potential LLM data leakage as internal-validity threats.Source: §6.2 · pp. 17–18
• Findings are limited to open-source repositories in Python and Java, and the evaluated repository selection is not exhaustive.The study includes 164 LocBench repositories, 12 SWE-bench Lite repositories, and 9 Multi-SWE-bench Java repositories.Source: §6.2 · p. 18

7 Related Work

• Embedding-based methods rank code entities by semantic similarity but overlook repository structure and dependencies, limiting their ability to trace root causes or interdependent changes.They also require indexing infrastructure with storage and maintenance overhead.Source: §7 · p. 18
• Graph-based agentic localization improves structural awareness, yet existing approaches still rely on superficial relevance and lack explicit causal modeling of interdependent code entities.Earlier workflows also include fixed procedural hierarchies and autonomous agentic exploration.Source: §7 · p. 18
• Traditional causal frameworks target tabular data, while LLM-based causal reasoning offers opportunities for text and code but remains limited in counterfactual reasoning.GraphLocator leverages LLM causal discovery to infer latent structures from code snippets and identify interdependent sub-issues.Source: §7 · pp. 18–19

8 Conclusion

• Across three real-world Python and Java datasets, GraphLocator consistently outperforms existing baselines in function-level recall and precision and in both mismatch scenarios.The authors describe the approach as a promising direction for more accurate and interpretable automated issue localization.Source: §8 · p. 19