AI coding agents
Kozuchi Takes the Same 27B Code-Repair Agent to Java, Solving 41 of 128 Multi-SWE-bench Tasks
Fujitsu ported its fine-tuning-free Qwen3.5-27B setup, eight-stage repair workflow, and cross-agent test selector to Java unchanged. It solved 374 of 500 tasks on the Python SWE-bench Verified benchmark, but achieved only a 32.03% success rate on Java, showing that agent-framework portability does not imply equal language proficiency.
Fujitsu’s research team has formally published its paper on Kozuchi Agent, extending its previously reported Python code-repair results to Java. The system is built around a locally hosted Qwen3.5-27B model with no additional fine-tuning. Rather than using a single ReAct loop, it surrounds the model with an eight-stage workflow, file-based persistent state, deterministic software-engineering tools, and a model-agnostic action interface. A CI pipeline handles inference startup, evaluation, and trajectory preservation.
Kozuchi runs each task eight times, with every run producing both a patch and tests, then cross-applying those tests to the other candidate patches. This selector neither reads hidden evaluation results nor requires a separately trained verifier. Across the 500 Python tasks in SWE-bench Verified, the eight runs achieved an average Pass@1 of 67.7%, while the theoretical upper bound in which at least one run succeeds was 81.6%. The patch ultimately submitted by the selector solved 374 tasks, or 74.8%—14 more than a fixed strategy that always selected the first candidate. Even so, it missed 34 cases for which a correct candidate already existed.
The cross-language results are even more noteworthy. The researchers retained the same model, stage graph, state format, and eight-candidate strategy, changing only the test environment to Maven/Gradle. On Multi-SWE-bench Java, the system solved 41 of 128 tasks, or 32.03%. The share of messages used by each stage differed by no more than five percentage points between the two languages, indicating that the workflow structure did transfer. However, the absolute success rate on Java was more than 42 percentage points lower, and all 38 elastic/logstash cases failed. The authors therefore claim only that the framework works across languages, not that the model has equal proficiency in both languages.
Engineering teams should note two points. First, candidate generation and the selector itself have become major performance levers, rather than performance depending solely on the foundation model. Second, the paper conducts matched ablations only on candidate count and selection signals; it does not isolate the causal contributions of staging, persistent state, or the tool suite. The results also come from public benchmarks, with no developer studies on private codebases or production telemetry yet available.