SweRank+: Multilingual, Multi-Turn Code Ranking for Software Issue Localization

TL;DR

SweRank+ addresses multilingual software issue localization by introducing SweRankMulti, a multilingual retrieve-and-rerank system trained on SweLocMulti, and SweRankAgent, an iterative, multi-turn agent that refines localization through successive searches and evidence aggregation. SweRankMulti uses a bi-encoder retriever (SweRankEmbedMulti) and a listwise LLM reranker (SweRankLLMMulti) trained with a weakly supervised objective on a large, cross-language dataset, achieving state-of-the-art multilingual performance while preserving Python task strength. SweRankAgent extends this with a ReAct-style loop (Search, Reasoning, Reformulation, Aggregate) and a memory of candidates, yielding additional gains over single-pass approaches, especially on hard/localization tasks with low lexical/semantic overlap or multiple target functions. The framework demonstrates that a lightweight, iterative reasoning approach can significantly improve localization accuracy across languages and complexity levels, suggesting practical impact for multilingual code maintenance and automated repair pipelines.

Abstract

Maintaining large-scale, multilingual codebases hinges on accurately localizing issues, which requires mapping natural-language error descriptions to the relevant functions that need to be modified. However, existing ranking approaches are often Python-centric and perform a single-pass search over the codebase. This work introduces SweRank+, a framework that couples SweRankMulti, a cross-lingual code ranking tool, with SweRankAgent, an agentic search setup, for iterative, multi-turn reasoning over the code repository. SweRankMulti comprises a code embedding retriever and a listwise LLM reranker, and is trained using a carefully curated large-scale issue localization dataset spanning multiple popular programming languages. SweRankAgent adopts an agentic search loop that moves beyond single-shot localization with a memory buffer to reason and accumulate relevant localization candidates over multiple turns. Our experiments on issue localization benchmarks spanning various languages demonstrate new state-of-the-art performance with SweRankMulti, while SweRankAgent further improves localization over single-pass ranking.

Figures (5)

• Figure 1: Comparison of function localization accuracy@10 against the SweRank baseline. SweRankMulti shows significant improvement on multilingual benchmarks (in red) while maintaining strong performance on Python-specific evaluations (in violet). SweRankAgent further improves over single-pass ranking approaches across the board.
• Figure 2: SweRankAgent interleaves Search actions, which retrieve candidate functions, and Aggregate actions, which selectively adds the most relevant functions to a persistent memory. The diagram illustrates a qualitative example of the agent solving a complex issue. The visualization details the agent's multi-turn trajectory (turns 1-3), showing how it iterates through thinking and tool calls to SweRankMulti to progressively narrow down the root cause from a broad search to the target function.
• Figure 3: Language-wise function localization accuracy for SWE-PolyBench and SWE-Bench-Multilingual.
• Figure 4: Comparison of SweRank, SweRankMulti, and SweRankAgent performance across varying levels of difficulty. The plots show Function Accuracy@10 separately for Python and other languages, broken down by (a) lexical overlap using ROUGE-1 scores and (b) semantic overlap using cosine similarity.
• Figure 5: Function Accuracy@10 breakdown by number of target localization functions.