Resolving Java Code Repository Issues with iSWE Agent

Jatin Ganhotra; Sami Serhan; Antonio Abu Nassar; Avraham Shinnar; Ziv Nevo; Martin Hirzel

cs.SEMar 11, 2026

Resolving Java Code Repository Issues with iSWE Agent

Authors: Jatin Ganhotra, Sami Serhan, Antonio Abu Nassar, Avraham Shinnar, Ziv Nevo, Martin Hirzel

Organizations: IBM

Abstract

Resolving issues on code repositories is an important part of software engineering. Various recent systems automatically resolve issues using large language models and agents, often with impressive performance. Unfortunately, most of these models and agents focus primarily on Python, and their performance on other programming languages is lower. In particular, a lot of enterprise software is written in Java, yet automated issue resolution for Java is under-explored. This paper introduces iSWE Agent, an automated issue resolver with an emphasis on Java. It consists of two sub-agents, one for localization and the other for editing. Both have access to novel tools based on rule-based Java static analysis and transformation. Using this approach, iSWE achieves state-of-the-art issue resolution rates across the Java splits of both Multi-SWE-bench and SWE-PolyBench. More generally, we hope that by combining the best of rule-based and model-based techniques, this paper contributes towards improving enterprise software development.

Figures & tables

Figure 1: Overview of iSWE Agent

Bench mark	Base LLM	% Resolved	$ Cost Total (Avg.)	Avg. # Tokens	Edit Correct Location
					File		Node
					Recall	Precision	Recall	Precision
iSWE-Agent (Frontier Models)
Multi-SWE-Bench (Java)	Claude-4.5-Sonnet	43 (33.6%)	237.89 (1.86)	598k	62%	82%	44%	68%
	Claude-3.7-Sonnet	32 (25.0%)	154.51 (1.21)	386k	50%	72%	32%	63%
Leaderboard submissions (top other agents for comparison)

Table 1: Main results on Multi-SWE-Bench (Java) and SWE-PolyBench (Java). 1 1 1 Correct Location metrics are unavailable for a subset of Multi-SWE-Bench (Java) leaderboard submissions due to difficulties encountered when applying the publicly available output patches. 2 2 2 Cost and average token metrics are unavailable for SWE-PolyBench (Java) leaderboard submissions due to the lack of publicly available trajectories.

Panel of Figure 2 — Figure 2: Average cost vs. %resolved across models. Most datapoints are from iSWE with different models; two datapoints for Multi-SWE-bench (Java) are from MopenHands and MSWE-agent.

Panel of Figure 3 — Figure 3: Average cost vs. resolved across individual instances in the benchmark (cumulative).

Benchmark	Base LLM	Avg. # Turns			Avg. Cost
Benchmark	Base LLM	Loc.	Editing	Total	Loc.	Editing	Total
Multi-SWE- Bench (Java)	Claude-4.5-Sonnet (Loc & Edit)	31.5	1.7	33.3	$1.77	$0.09	$1.86
Multi-SWE- Bench (Java)	DeepSeek-V3-2 (Loc & Edit)	41.6	2.5	44.2	$0.24	$0.01	$0.25
SWE-PolyBench (Java)	Claude-4.5-Opus (Loc & Edit)	34.1	2.2	36.3	$2.41	$0.14	$2.55
SWE-PolyBench (Java)	DeepSeek-V3-2 (Loc & Edit)	54.8	3.3	58.0	$0.27	$0.01	$0.28

Table 2: Cost per sub-agent.

Benchmark	Base LLM	Localization Correct Location				Edit Correct Location
Benchmark	Base LLM	File		Node		File		Node
		Recall	Precision	Recall	Precision	Recall	Precision	Recall	Precision
iSWE-Agent (Frontier Models)
Multi-SWE-Bench (Java)	Claude-4.5-Sonnet	63%	83%	36%	59%	62%	82%	44%	68%
Multi-SWE-Bench (Java)	Claude-3.7-Sonnet	56%	77%	28%	53%	50%	72%	32%	63%
iSWE-Agent (Open Source Models)

Table 3: Localization metrics per sub-agent.

Bench mark	Base LLM	Tool Usage (Number of Calls)							Total
Bench mark	Base LLM	get file info	get class info	get method info	get symbol info	get inheritance hierarchy	get function callers	get call chain	Total
Frontier Models
Multi-SWE-Bench (Java)	Claude-4.5-Sonnet	87	124	403	181	2	19	33	849
Multi-SWE-Bench (Java)	Claude-3.7-Sonnet	163	136	265	75	1	7	5	652
Open Source Models
	DeepSeek-V3-2	497	549	1803	1113	22	93	49	4126

Table 4: Tool usage distribution across the entire benchmark (not per-instance).

Tool Set	$ Cost	Avg.	Turns	File Retrieval		Node Retrieval
Tool Set	Total (Avg.)	#Tokens	Turns	Recall	Precision	Recall	Precision
iSWE-tools	226 (1.77)	575k	4037	63%	83%	36%	59%
iSWE-tools & view-file	229 (1.79)	584k	3777	63%	83%	37%	57%
Bash	263 (2.06)	667k	5815	62%	82%	37%	57%
All-Tools (iSWE-tools & Bash & swea)	336 (2.63)	855k	5820	62%	81%	37%	58%

Table 5: Localization performance for different tool sets, using Claude-4.5-Sonnet on Multi-SWE-bench (Java).

Bench mark	Complexity	# Instances	Base LLM	% Resolved	Avg. Cost	File Retrieval		Node Retrieval
Bench mark	Complexity	# Instances	Base LLM	% Resolved	Avg. Cost	Recall	Precision	Recall	Precision
Multi-SWE- Bench (Java)	Easy	27	Claude-4.5-Sonnet	16 (59.2%)	$1.65	65%	80%	42%	49%
	Easy	27	DeepSeek-V3-2	9 (33.3%)	$0.18	58%	74%	39%	47%
	Medium	65	Claude-4.5-Sonnet	26 (40.0%)	$1.71	66%	86%	39%	58%
	Medium	65	DeepSeek-V3-2	18 (27.6%)	$0.21	61%	70%	37%	47%
	Hard	36	Claude-4.5-Sonnet	1 (2.7%)	$2.26	55%	80%	22%	65%

Table 6: Performance by hand-annotated complexity level on Multi-SWE-Bench (Java) and by hand-annotated Verified or not label on SWE-PolyBench (Java).

Bench mark	Modification Type	# Instances	Base LLM	% Resolved	Avg. Cost	File Retrieval		Node Retrieval
Bench mark	Modification Type	# Instances	Base LLM	% Resolved	Avg. Cost	Recall	Precision	Recall	Precision
Multi-SWE-Bench (Java)	None	3	Claude-4.5-Sonnet	0 (0.0%)	$2.27	36%	100%	36%	100%
	None	3	DeepSeek-V3-2	0 (0.0%)	$0.14	25%	100%	25%	100%
	Function Only Single All	33	Claude-4.5-Sonnet	20 (60.6%)	$1.75	72%	87%	58%	70%
		33	DeepSeek-V3-2	10 (30.3%)	$0.23	56%	69%	50%	61%
		44	Claude-4.5-Sonnet	27 (61.3%)	$1.63	73%	90%	59%	74%

Table 7: Performance by localization-derived complexity attributes. These attributes are computed based on the gold patch and emulate the attributes defined by SWE-PolyBench.

# Hints Provided	# Instances	% Resolved	File Retrieval		Node Retrieval
# Hints Provided	# Instances	% Resolved	Recall	Precision	Recall	Precision
0	79	44%	68.33%	84.99%	52.03%	73.89%
1	23	30%	59.66%	86.96%	35.98%	59.50%
2	4	25%	50.00%	100.00%	43.15%	95.00%
3	2	0%	50.00%	62.50%	50.00%	66.67%
4	8	0%	62.35%	75.00%	18.31%	68.84%

Table 8: Performance by number of identifier hints about new entities provided in Multi-SWE-Bench (Java), using Claude-4.5-Sonnet.

Software Engineering Agents

Software Engineering Agents are AI systems that use code, tools, and feedback to perform software development tasks.

See all 344 papers

SWE-Adept: An LLM-Based Agentic Framework for Deep Codebase Analysis and Structured Issue Resolution
Mar 1, 2026Kang He, Kaushik RoyCode Localization Multi-Agent Coding
Electrical and Computer Engineering, Purdue University
BeyondSWE: Can Current Code Agent Survive Beyond Single-Repo Bug Fixing?
Mar 3, 2026Guoxin Chen, Fanzhe Meng, Jiale Zhao +12Retrieval-Augmented Code Generation Software Engineering Benchmarks
Kozuchi Agent: A Language-Agnostic Open-Weight Agent for Software Repair
Aug 16, 2026Mehdi Bahrami, Kosaku Kimura, Satoshi Munakata +24Software Engineering Agents Automated Program Repair
Fujitsu Research of America Santa Clara, CA, USA · Fujitsu Research Kawasaki, Japan · Fujitsu Research of Europe Slough, UK +1

Mar 1, 2026J/K move · Enter open · S save

SWE-Adept: An LLM-Based Agentic Framework for Deep Codebase Analysis and Structured Issue Resolution

Kang He, Kaushik Roy

Electrical and Computer Engineering, Purdue University

Code Localization Multi-Agent Coding Automated Program Repair Repository-Level Code Generation LLM Agent Planning Memory-Augmented Language Model Agents

Open paper AskPDF arXiv

Large language models (LLMs) exhibit strong performance on self-contained programming tasks. However, they still struggle with repository-level software engineering (SWE), which demands (1) deep codebase navigation with effective context management for accurate localization, and (2) systematic approaches for iterative, test-driven code modification to resolve issues. To address these challenges, we propose SWE-Adept, an LLM-based two-agent framework where a localization agent identifies issue-relevant code locations and a resolution agent implements the corresponding fixes. For issue localization, we introduce agent-directed depth-first search that selectively traverses code dependencies. This minimizes issue-irrelevant content in the agent's context window and improves localization accuracy. For issue resolution, we employ adaptive planning and structured problem solving. We equip the agent with specialized tools for progress tracking and Git-based version control. These tools interface with a shared working memory that stores code-state checkpoints indexed by execution steps, facilitating precise checkpoint retrieval. This design enables reliable agent-driven version-control operations for systematic issue resolution, including branching to explore alternative solutions and reverting failed edits. Experiments on SWE-Bench Lite and SWE-Bench Pro demonstrate that SWE-Adept consistently outperforms prior approaches in both issue localization and resolution, improving the end-to-end resolve rate by up to 4.3%.

Abstract

Software Engineering Agents

Explore similar work

SWE-Adept: An LLM-Based Agentic Framework for Deep Codebase Analysis and Structured Issue Resolution

BeyondSWE: Can Current Code Agent Survive Beyond Single-Repo Bug Fixing?

Kozuchi Agent: A Language-Agnostic Open-Weight Agent for Software Repair