CuraLoc: Repository-Level Fault Localization via Planned Code Inspection

Sangeeta Warrier; Anush Sankaran; Ryen W. White

cs.LGOct 22, 2026

CuraLoc: Repository-Level Fault Localization via Planned Code Inspection

Authors: Sangeeta Warrier, Anush Sankaran, Ryen W. White

Organizations: Microsoft

Abstract

A bug report describes what went wrong, but rarely where the code should change. Existing repository-level localization methods often let language models iteratively decide what code to inspect next, requiring repeated model calls and substantial token usage. We investigate whether these inspection decisions can instead be planned before source retrieval. In this research, we introduce CuraLoc, which first selects candidate files and represents their symbols using compact behavioral and structural descriptions. In a single planning decision, the model jointly selects starting symbols, chooses code relationships to inspect from each, and forms hypotheses about likely edits. Deterministic retrieval collects the requested source for a final ranking informed by those hypotheses, concentrating model reasoning on choosing and comparing code rather than directing each retrieval step. We demonstrate the efficacy of our proposed approach on the SWE-Bench Lite and Verified datasets. On SWE Bench Verified with GPT-5.2, CuraLoc raises Func@1 from LocAgent's 53.8% to 59.3% while using 90.3% fewer recorded tokens, and also outperforms RepoSearcher (56.2%) and Agentless (39.6%) on Func@1. On SWE-bench Lite, CuraLoc achieves a high 83.8% Func@5 accuracy improving on Agentless by more than 15% in relative accuracy. Additionally, CuraLoc supports different model backbones, with strong localization accuracy demonstrated using both GPT-5.2 and GPT-6 Astra. These results provide evidence that repository-level code inspection can be planned effectively without iterative model-controlled navigation.

Figures & tables

Figure 1: CuraLoc uses three primary-model decisions. Symbol selection specifies both starting locations and retrieval actions; deterministic retrieval supplies source code for final ranking.

Method	Func@1	Func@5	Func@10	MRR	File@1	File@10	Tokens (K)
CuraLoc	$\mathbf{59.3}_{\pm 0.9}$	$\mathbf{77.2}_{\pm 0.6}$	$\mathbf{79.6}_{\pm 1.1}$	$\mathbf{0.840}_{\pm 0.006}$	$\mathbf{75.8}_{\pm 0.3}$	$88.3_{\pm 0.6}$	$125.7_{\pm 0.8}$
LocAgent	$53.8_{\pm 0.5}$	$74.8_{\pm 0.3}$	$79.1_{\pm 0.3}$	$0.791_{\pm 0.003}$	$73.7_{\pm 0.6}$	$\mathbf{94.4}_{\pm 0.7}$	$1292.7_{\pm 4.4}$
RepoSearcher	$56.2_{\pm 0.4}$	$74.4_{\pm 1.0}$	$74.4_{\pm 1.0}$	$0.796_{\pm 0.004}$	$\mathbf{75.8}_{\pm 0.8}$	$88.6_{\pm 0.8}$	$\mathbf{72.1}_{\pm 2.9}$
Agentless	$39.6_{\pm 1.0}$	$61.6_{\pm 1.9}$	$63.1_{\pm 1.6}$	$0.613_{\pm 0.022}$	$70.0_{\pm 1.2}$	$81.9_{\pm 1.1}$	–

Table 1: GPT-5.2 localization on SWE Bench Verified (

n=454

). Entries are three-run means with sample standard deviations. Accuracy is in percent; tokens are in thousands per task. Bold marks the best reported mean. Agentless token totals are unavailable. RepoSearcher returns at most five symbols, so Func@10 equals Func@5.

Method	Func@1	Func@5	Func@10	MRR	File@1	File@10	Tokens (K)
CuraLoc	$\mathbf{64.4}_{\pm 1.6}$	$\mathbf{83.8}_{\pm 1.6}$	$\mathbf{87.1}_{\pm 0.2}$	$\mathbf{0.798}_{\pm 0.009}$	$\mathbf{85.9}_{\pm 0.9}$	$93.6_{\pm 0.2}$	$124.4_{\pm 0.6}$
LocAgent	$58.5_{\pm 1.1}$	$82.2_{\pm 0.4}$	$85.5_{\pm 1.1}$	$0.755_{\pm 0.006}$	$82.0_{\pm 1.4}$	$\mathbf{97.0}_{\pm 0.2}$	$1457.4_{\pm 21.4}$
RepoSearcher	$61.2_{\pm 1.1}$	$78.3_{\pm 1.7}$	$78.3_{\pm 1.7}$	$0.759_{\pm 0.011}$	$84.1_{\pm 1.3}$	$93.1_{\pm 0.7}$	$80.3_{\pm 3.8}$
Agentless	$46.8_{\pm 1.5}$	$69.8_{\pm 1.2}$	$70.9_{\pm 0.8}$	$0.625_{\pm 0.008}$	$79.9_{\pm 0.6}$	$89.7_{\pm 1.1}$	$\mathbf{72.7}_{\pm 2.8}$

Table 2: GPT-5.2 localization on SWE Bench Lite (

n=274

), using the conventions of Table 1 . SWE Bench Lite was used for configuration selection. RepoSearcher returns at most five symbols, so its Func@10 equals Func@5.

Figure 2: Func@1 versus recorded tokens on (a) SWE Bench Lite and (b) SWE Bench Verified. Points are three-run means; error bars show sample standard deviations. The token axis is logarithmic. Agentless is omitted from panel (b) because its exact token total is unavailable.

Figure 3: CuraLoc-minus-baseline Func@1 differences with pointwise 95% intervals from paired task and whole-repository resampling. Positive values favor CuraLoc. These intervals are unadjusted; SWE Bench Verified outside SWE Bench Lite is a subset of SWE Bench Verified.

Configuration	Func@1	Func@10	MRR
Full CuraLoc (reference)	67.2	87.6	0.823
Selection-stage readout	55.8	83.2	0.739
Source-order selection	42.3	55.1	0.518
Without summaries	64.6	86.1	0.807
Starting symbols only	67.2	83.9	0.822
Relevance-guided formulation	63.1	82.8	0.775

Table 3: Single-run SWE Bench Lite comparisons (

n=274

) against one fixed reference. Interventions are separate; the selection-stage ranking is an offline readout.

Figure 4: Evidence-budget sensitivity on SWE Bench Lite (

n=274

). (a) Func@1 and mean S2+S3 tokens across file settings; the outlined setting averages three development runs, while alternatives use one run each. (b) Source depth with S2 fixed to one reference run; labels show Func@1 and S3 tokens. Both panels exclude S1 and summary-generation tokens.

Method	Func@1	Func@3	Func@5	Func@10	MRR	File@1	File@10	Tokens (K)
CuraLoc	67.0	84.4	87.4	89.2	0.935	84.1	94.5	138.9
LocAgent	63.2	80.0	82.2	85.9	0.896	81.1	96.9	95.3
RepoSearcher	66.7	82.8	87.4	87.4	0.926	83.9	94.1	105.0
Agentless	59.7	76.2	78.4	79.7	0.846	82.2	90.3	28.2

Table 4: Single-run GPT-6 Astra localization on SWE Bench Verified (

n=454

). Tokens are in thousands per task and include all contributing model calls, including CuraLoc’s GPT-5.4-mini summaries. Bold marks the best score or lowest token total. RepoSearcher’s Func@10 equals Func@5; finalized empty rankings count as misses.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: CuraLoc-minus-baseline differences in (a) Func@1 and (b) Func@10 across all nine comparisons. Circles and squares show pointwise 95% task- and repository-bootstrap intervals, respectively. Positive differences favor CuraLoc. Intervals are unadjusted for multiple comparisons. RepoSearcher returns at most five symbols, so its Func@10 equals Func@5. The non-overlapping cohort is a subset of SWE Bench Verified.

SWE Bench Verified outside SWE Bench Lite ( $n=370$ )
Method	Func@1	Func@3	Func@5	Func@10	MRR	File@1	File@10
CuraLoc	$55.5_{\pm 0.7}$	$69.4_{\pm 0.3}$	$73.9_{\pm 0.6}$	$76.8_{\pm 1.1}$	$0.837_{\pm 0.006}$	$72.9_{\pm 0.6}$	$86.8_{\pm 0.7}$
LocAgent	$51.0_{\pm 0.4}$	$67.7_{\pm 0.7}$	$73.1_{\pm 0.4}$	$77.1_{\pm 0.3}$	$0.797_{\pm 0.003}$	$72.3_{\pm 0.8}$	$94.1_{\pm 0.8}$
RepoSearcher	$52.5_{\pm 0.4}$	$66.7_{\pm 0.6}$	$71.5_{\pm 1.0}$	$71.5_{\pm 1.0}$	$0.793_{\pm 0.001}$	$73.8_{\pm 0.7}$	$87.5_{\pm 0.9}$
Agentless	$37.0_{\pm 1.0}$	$53.8_{\pm 1.4}$	$58.4_{\pm 2.0}$	$60.0_{\pm 1.6}$	$0.605_{\pm 0.023}$	$67.7_{\pm 1.2}$	$79.9_{\pm 0.8}$
SWE Bench Verified overlapping SWE Bench Lite ( $n=84$ )

Appendix

Table 5: GPT-5.2 localization on fixed SWE Bench Verified subsets. Entries are means across three runs, with

\pm

sample standard deviations in subscripts. Accuracy values are percentages; MRR is first-hit reciprocal rank on a 0–1 scale. These subsets reuse the full SWE Bench Verified executions.

Method	Run	A	B	C	Total failures
CuraLoc	1	94	4	0	98
	2	79	9	0	88
	3	88	4	0	92
LocAgent	1	23	63	10	96
	2	23	59	13	95
	3	20	60	13	93

Appendix

Table 6: Func@10 failures on SWE Bench Verified. A: incomplete coverage under the method-specific definition; B: complete coverage but incomplete final recovery; C: unknown coverage. All 454 tasks remain in each run’s denominator.

Run	Tasks	CuraLoc	LocAgent	Difference
1	342	338	316	+22
2	359	351	327	+24
3	345	341	319	+22

Appendix

Table 7: Final recovery on tasks with complete coverage under both methods’ respective definitions. Counts are Func@10 successes.

Configuration	Func@1	Func@3	Func@5	Func@10	MRR	File@1
Full CuraLoc (reference)	67.2	81.8	85.0	87.6	0.823	88.0
File scoping only	–	–	–	–	–	74.1
Selection-stage readout	55.8	76.3	79.6	83.2	0.739	80.3
Source-order selection	42.3	51.1	53.3	55.1	0.518	75.9
Without summaries	64.6	81.0	84.3	86.1	0.807	86.1
Starting symbols only	67.2	81.0	83.2	83.9	0.822	88.3

Appendix

Table 8: Stage diagnostics, component interventions, and source-depth comparisons against the fixed SWE Bench Lite reference. Accuracy values are percentages; MRR uses the first correct symbol. File scoping reports native file accuracy, whereas other File@1 values are derived from symbol predictions. Source-depth settings retain the reference selection-stage outputs.

Stage A: file policy; tokens cover S2+S3
$F$	$T$	$L$	Runs	Func@1	Func@3	Func@5	Func@10	MRR	Tokens/task
5	1	60	1	59.1	73.4	75.9	77.4	0.730	29,636 †
5	3	60	1	64.2	79.9	83.2	85.4	0.797	41,461
5	5	60	1	62.8	78.5	81.8	85.8	0.784	46,635
10	1	60	1	58.8	73.7	77.0	78.5	0.731	38,131
10	3	60	3	$\mathbf{65.5}_{\pm 1.8}$	$\mathbf{80.4}_{\pm 1.3}$	$\mathbf{84.2}_{\pm 0.8}$	$\mathbf{86.7}_{\pm 0.8}$	$\mathbf{0.807}_{\pm 0.014}$	51,352

Appendix

Table 9: Hyperparameter sensitivity on SWE Bench Lite (

n=274

). Bold identifies the selected configuration, not the largest value in each column. Subscripts show sample standard deviations for the replicated setting. Token counts are means over tasks and cover different stages in the two panels; both exclude S1 and summary generation. † The

F=5,T=1

setting retains one deterministic empty-pool miss and its 1,375 recorded tokens without stage attribution, contributing approximately five tokens per task.

Fault Localization

Fault localization identifies the code, component, event, or location responsible for a failure so it can be diagnosed and corrected.

See all 42 papers

Jun 23, 2026

SHERLOC: Structured Diagnostic Localization for Code Repair Agents

LLM agents solve repository-level coding tasks through multi-turn tool use, but utilize half their budget on locating faults before editing. Dedicated localization frameworks have emerged, yet are still evaluated as file retrieval rather than actionable diagnosis, producing locations without the diagnostic context a repair agent needs. We introduce SHERLOC (Structured Hypothesis-driven Exploration and Reasoning for Localization), a training-free framework pairing a reasoning LLM with compact repository tools and self-recovery, without fine-tuning or multi-agent orchestration. SHERLOC reaches state-of-the-art localization across model scales: 84.33% accuracy@1 on SWE-Bench Lite and 81.27% recall@1 on SWE-Bench Verified; at ~30B parameters, it matches or outperforms other agentic methods. Injecting our locations and diagnostic findings into repair agents yields an average +5.95 pp resolve-rate gain from the best SHERLOC result per setting on SWE-Bench Verified. SHERLOC cuts localization and total tokens by 36.7% and 23.1% on average.

Hovhannes Tamoyan, Sean Narenthiran, Erik Arakelyan +2

NVIDIA, Santa Clara, CA 95051, USA · 2TU Darmstadt, Darmstadt, Germany · 3hessian.AI & National Research Center for Applied Cybersecurity ATHENE

Software Engineering Agents AI Coding Agents Fault Localization Repository-Level Program Repair

Ask PDF

Sep 25, 2026

Semantic Navigation for Issue Localization in Code Repository

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%.

Yunxiang Wei, Zhenyu Lei, Jundong Li

Independent Researcher · University of Virginia

Software Engineering Agents LLM Grounding

Ask PDF

May 18, 2026

BLAgent: Agentic RAG for File-Level Bug Localization

Bug localization remains a key bottleneck for large language model (LLM)-based software maintenance, where accurately identifying faulty code is essential for debugging, root cause analysis, triage, and automated program repair (APR). File-level bug localization is especially critical in hierarchical localization and repair pipelines, where incorrect file selection can propagate to downstream stages such as function-level localization and patch generation. While Retrieval-Augmented Generation (RAG) offers a promising way to ground LLMs in repository context, existing RAG pipelines often rely on static retrieval and lack the reasoning needed to accurately identify faulty code. In this work, we present BLAgent, a novel agentic RAG framework for file-level bug localization that integrates three key ideas: (i) code structure-aware repository encoding with path-augmented AST-based chunking, (ii) dual-perspective query transformation that captures both structural and behavioral signals from bug reports, and (iii) two-phase agentic reranking that combines symbolic inspection with evidence-grounded reasoning. Unlike prior graph-based or multi-hop agentic approaches, BLAgent adopts a bounded reasoning strategy that limits LLM-based inspection and reranking to a compact, retrieval-filtered set of candidate files, avoiding open-ended repository traversal. This design balances localization accuracy with computational cost. On SWE-bench-Lite, BLAgent attains over 78% Top-1 accuracy with open-source models and over 86% with a closed-source model, while being over 18x cheaper than the strongest baseline using the same model. When integrated into an APR framework, BLAgent improves end-to-end repair success by up to 25%.

Md Afif Al Mamun, Gias Uddin

University of Calgary, Canada · York University, Canada

Software Engineering Agents LLM Reranking Agentic RAG

Ask PDF

Abstract

Fault Localization

Explore similar work

SHERLOC: Structured Diagnostic Localization for Code Repair Agents

Semantic Navigation for Issue Localization in Code Repository

BLAgent: Agentic RAG for File-Level Bug Localization