Fang, Wenkai, Liu, Shunyu, Luo, Wei, Wu, Zihao, Zhou, Yang, Zheng, Tongya, Yin, Wotao, Shen, Jialie, Rutkowski, Leszek, Song, Mingli, Tao, Dacheng · Nov 5, 2026
Wenkai Fang and coauthors propose NINJA for issue localization
Markdown · JSON · Original source
Wenkai Fang and coauthors propose NINJA, a hierarchical Navigator-Inspector Joint Architecture for issue localization: finding where in a software repository an issue may be addressed. They argue that a single agent’s growing context can compress effective reasoning space, while flat exploration makes it hard to balance file-level breadth with function-level depth. NINJA assigns file-level search and global search state to a navigator, which sends selected entry files to inspectors. Each inspector explores functions around its assigned files in a separate context and returns suspicious locations; the navigator uses that feedback to continue or finalize localization. The authors say this division balances broad repository search with closer inspection, while preventing local exploration traces from accumulating in one growing context. They also introduce a two-stage strategy to strengthen coordination and exploration. The abstract reports that NINJA outperforms competitive baselines across multiple and backbones, and says fine-tuned Qwen3-Coder-30B-A3B-Instruct surpasses -Haiku-4.5. These are the authors’ reported experimental results; the abstract does not name the benchmarks or give result measures, so their scope and size cannot be assessed from this evidence alone.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T18:03:58.182Z. First observed: 2026-11-10T08:36:27.828Z. Retrieved: 2026-11-10T18:03:12.687Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The abstract does not name the , provide result measures, or specify the conditions for the reported model comparison; the results’ scope and size cannot be assessed from this evidence alone.
Open question
Which and result measures support the authors’ reported comparisons, including the comparison between Qwen3-Coder-30B-A3B-Instruct and -Haiku-4.5?
Source-stated
The authors argue that single-agent methods with growing context and flat exploration limit effective reasoning and the balance of file-level breadth and function-level depth.
long-context accumulation compresses the effective reasoning space. Meanwhile, the flat exploration structure hinders the balance between file-level breadth and function-level depth.
Source-stated
NINJA assigns global search state and file-level search to a navigator, and function-level exploration around selected files to independent inspectors.
the navigator maintains the global search state, performs file-level search, and dispatches selected entry files, while inspectors independently conduct function-level exploration
Source-stated
Inspectors return suspicious locations as feedback for the navigator to decide whether to continue exploring or finalize localization.
inspectors return suspicious locations as feedback, and the navigator updates the global state to decide whether to continue exploration or finalize localization.
Source-stated
The authors report that NINJA outperforms competitive baselines across multiple benchmarks and LLM backbones, and that fine-tuned Qwen3-Coder-30B-A3B-Instruct surpasses Claude-Haiku-4.5.
Extensive experiments across multiple benchmarks and LLM backbones show that NINJA consistently outperforms competitive baselines. Notably, after fine-tuning, Qwen3-Coder-30B-A3B-Instruct surpasses the strong closed-source Claude-Haiku-4.5 model.
Source and contributor context
Fang, Wenkai, Liu, Shunyu, Luo, Wei, Wu, Zihao, Zhou, Yang, Zheng, Tongya, Yin, Wotao, Shen, Jialie, Rutkowski, Leszek, Song, Mingli, Tao, Dacheng · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: ff803f39-1464-4835-85d3-e8e85c0cca38.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T18:03:58.182Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 3736f2d0944448c5876379a1bfd05808476998980bf017551e14240a3644eb6e. Source text hash: e838014e2681ade2b158a32bfce47be085fb2ecf4043633a8f8c57ad376e2294.
Revision and source-check history
Version 1 · 2026-11-10T18:03:58.182Z
Wenkai Fang and coauthors propose NINJA for issue localization
Wenkai Fang and coauthors propose NINJA, a hierarchical Navigator-Inspector Joint Architecture for issue localization: finding where in a software repository an issue may be addressed. They argue that a single agent’s growing context can compress effective reasoning space, while flat exploration makes it hard to balance file-level breadth with function-level depth. NINJA assigns file-level search and global search state to a navigator, which sends selected entry files to inspectors. Each inspector explores functions around its assigned files in a separate context and returns suspicious locations; the navigator uses that feedback to continue or finalize localization. The authors say this division balances broad repository search with closer inspection, while preventing local exploration traces from accumulating in one growing context. They also introduce a two-stage strategy to strengthen coordination and exploration. The abstract reports that NINJA outperforms competitive baselines across multiple and backbones, and says fine-tuned Qwen3-Coder-30B-A3B-Instruct surpasses -Haiku-4.5. These are the authors’ reported experimental results; the abstract does not name the benchmarks or give result measures, so their scope and size cannot be assessed from this evidence alone.
Original source
cs.AI · Article
Source details
Original source date: Nov 5, 2026 (2026-11-05).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T18:03:58.182Z. First observed: 2026-11-10T08:36:27.828Z. Retrieved: 2026-11-10T18:03:12.687Z.
Limits
The abstract does not name the , provide result measures, or specify the conditions for the reported model comparison; the results’ scope and size cannot be assessed from this evidence alone.
Open question
Which and result measures support the authors’ reported comparisons, including the comparison between Qwen3-Coder-30B-A3B-Instruct and -Haiku-4.5?
Source-stated
The authors argue that single-agent methods with growing context and flat exploration limit effective reasoning and the balance of file-level breadth and function-level depth.
long-context accumulation compresses the effective reasoning space. Meanwhile, the flat exploration structure hinders the balance between file-level breadth and function-level depth.
Source-stated
NINJA assigns global search state and file-level search to a navigator, and function-level exploration around selected files to independent inspectors.
the navigator maintains the global search state, performs file-level search, and dispatches selected entry files, while inspectors independently conduct function-level exploration
Source-stated
Inspectors return suspicious locations as feedback for the navigator to decide whether to continue exploring or finalize localization.
inspectors return suspicious locations as feedback, and the navigator updates the global state to decide whether to continue exploration or finalize localization.
Source-stated
The authors report that NINJA outperforms competitive baselines across multiple benchmarks and LLM backbones, and that fine-tuned Qwen3-Coder-30B-A3B-Instruct surpasses Claude-Haiku-4.5.
Extensive experiments across multiple benchmarks and LLM backbones show that NINJA consistently outperforms competitive baselines. Notably, after fine-tuning, Qwen3-Coder-30B-A3B-Instruct surpasses the strong closed-source Claude-Haiku-4.5 model.
Source and contributor context
Fang, Wenkai, Liu, Shunyu, Luo, Wei, Wu, Zihao, Zhou, Yang, Zheng, Tongya, Yin, Wotao, Shen, Jialie, Rutkowski, Leszek, Song, Mingli, Tao, Dacheng · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Editorial provenance
Version 1. Publication ID: ff803f39-1464-4835-85d3-e8e85c0cca38.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-10T18:03:58.182Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 3736f2d0944448c5876379a1bfd05808476998980bf017551e14240a3644eb6e. Source text hash: e838014e2681ade2b158a32bfce47be085fb2ecf4043633a8f8c57ad376e2294.
passed · 2026-11-10T18:03:58.182Z · anthropic/claude-opus-5.5
Publication: ff803f39-1464-4835-85d3-e8e85c0cca38
attribution — pass: The draft correctly attributes NINJA to Wenkai Fang and coauthors, matching the arXiv author list. It frames the arguments and results as the authors' claims rather than established fact.
Authors:Wenkai Fang, Shunyu Liu, Wei Luo, Zihao Wu, Yang Zhou, Tongya Zheng, Wotao Yin, Jialie Shen, Leszek Rutkowski, Mingli Song, Dacheng Tao
summary — pass: The summary follows the abstract's problem, navigator/inspector design, feedback loop, fine-tuning and reported results. It glosses issue localization in plain language as finding where in a repository an issue may be addressed. It drops the word 'consistently' from the results claim, which is a minor softening and not a distortion. Terms such as context, LLM backbones and fine-tuning are kept exact and suit term-library tooltips. The title and question are neutral, and the question targets a real gap in the evidence.
issue localization, a critical prerequisite for software issue resolution
date — pass: The displayed original publication date of 2026-11-05 matches the v1 submission date. The later feed date of November 10 is correctly not used.
[Submitted on 5 Nov 2026]
support — pass: Every claim's excerpt is an exact quote from the abstract and supports its claim. The benchmark results are labeled as the authors' reported results. The draft adds no unsupported benefits, costs or real-world reliability claims. The limitation that the abstract names no benchmarks or result measures is accurate.
Notably, after fine-tuning, Qwen3-Coder-30B-A3B-Instruct surpasses the strong closed-source Claude-Haiku-4.5 model.
media — pass: The selected media is the verified arXiv abstract page, which is the direct original source. No better source media was found.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeAI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referenceClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore Claude