pith. sign in

arxiv: 2605.16352 · v1 · pith:P472Q75Bnew · submitted 2026-05-08 · 💻 cs.IR · cs.AI· cs.LG

LARGER: Lexically Anchored Repository Graph Exploration and Retrieval

Yuntong Hu , Tongli Su , Liang Zhao , Bowen Zhu , Hasibul Haque This is my paper

Pith reviewed 2026-05-20 23:31 UTC · model grok-4.3

classification 💻 cs.IR cs.AIcs.LG
keywords repository localizationcoding agentsgraph retrievallexical anchoringstructural localizationcodebase navigationsoftware engineering agents
0
0 comments X

The pith

LARGER anchors lexical searches to repository graph structures to improve file localization accuracy for coding agents.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Repository-level coding agents must first locate the relevant files and symbols for a task, yet current methods rely mainly on lexical search and miss structural relations such as imports, call chains, and type hierarchies. The paper formalizes this challenge as Lexically Anchored Structural Localization, which requires converting initial lexical matches into high-precision graph entry points and then exposing useful local neighborhoods inside the agent's normal search loop. LARGER implements the idea through an active-set retrieval process that begins with lexical hits, aligns them to graph anchors, and applies confidence-filtered expansion without any external graph database or separate interface. When added to existing CLI coding agents, the method raises file-level Acc@5 on LocBench by 13.9 points with tuned settings and 11.8 points with fixed settings over the strongest baseline, while also improving results on test generation and codebase QA tasks.

Core claim

We formalize repository context localization as Lexically Anchored Structural Localization, where success depends on turning lexical matches into high-precision structural entry points and exposing the most useful confidence-filtered local neighborhoods within the agent's existing search loop. We introduce LARGER, a lexically anchored active-set retrieval framework that starts from lexical matches, aligns them to graph anchors, and performs confidence-filtered local expansion within the agent's existing search loop, integrating directly into existing CLI coding agents without requiring external graph databases or specialized graph interfaces.

What carries the argument

Lexically Anchored Structural Localization, the formalization that converts lexical matches into graph anchors for confidence-filtered local expansion inside the agent's standard search loop.

If this is right

  • File-level Acc@5 on LocBench rises by 13.9 points with tuned hyperparameters and 11.8 points with fixed hyperparameters over the strongest baseline.
  • Consistent gains appear on MuLocBench, SWE-Atlas Test Writing, and SWE-Atlas Codebase QA.
  • The framework integrates into existing CLI coding agents without external graph databases or specialized interfaces.
  • Better localization reduces cascading failures in downstream tasks such as patch generation and test writing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Precise neighborhood retrieval could reduce the context size needed by agents for the same task accuracy.
  • The active-set approach may scale to larger or frequently changing codebases if the confidence filter adapts to new edits.
  • Similar anchoring of lexical results to structural graphs could improve retrieval in other large-document domains beyond software repositories.

Load-bearing premise

Lexical matches must reliably serve as high-precision entry points that allow useful structural neighborhoods to be selected and expanded inside the agent's existing search loop.

What would settle it

Integrating LARGER into agents and measuring file-level Acc@5 on LocBench yields no gain or a loss relative to the strongest baseline.

Figures

Figures reproduced from arXiv: 2605.16352 by Bowen Zhu, Hasibul Haque, Liang Zhao, Tongli Su, Yuntong Hu.

Figure 1
Figure 1. Figure 1: Effect of enabling LARGER in a matched CLI agent. (a) File-level Recall@5 across [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Accuracy–efficiency frontier on MuLocBench and LocBench. Each point is one method, with file-level Acc@5 on the y-axis and median runtime per instance on the x-axis. The two LARGER operating points lie on the upper frontier in both benchmarks. B Full Experimental Results B.1 Efficiency Tables and Cost Decomposition This appendix collects the full per-method efficiency comparison summarized in Section 5.2 (… view at source ↗
Figure 3
Figure 3. Figure 3: Per-commit cost of full graph reconstruction vs. commit-aware alignment on MuLocBench [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Oracle hyperparameters versus repository size. Each point is one repository; marker size is proportional to its issue count. Solid lines are issue-weighted least-squares fits in log10(LOC). (a) On MuLocBench (multi-file edits), the oracle top-k ⋆ trends upward with repository size (Spearman ρ=+0.36): larger repos benefit from wider neighborhoods. (b) On LocBench, the oracle confidence threshold θ ⋆ trends … view at source ↗
read the original abstract

Repository-level coding agents must first localize the files and symbols relevant to a task; failures at this stage can cascade across downstream objectives ranging from patch generation to test writing and codebase question answering. Existing agents navigate repositories primarily through lexical search, often missing structural relations such as imports, call chains, type hierarchies, and code-test links. Graph-based retrieval can recover such dependencies, but existing approaches often require separate graph tools or traversal stages that fragment the agent's interaction loop. We formalize repository context localization as Lexically Anchored Structural Localization, where success depends on turning lexical matches into high-precision structural entry points and exposing the most useful confidence-filtered local neighborhoods within the agent's existing search loop. We introduce LARGER (Lexically Anchored Repository Graph Exploration and Retrieval), a lexically anchored active-set retrieval framework that starts from lexical matches, aligns them to graph anchors, and performs confidence-filtered local expansion within the agent's existing search loop. LARGER integrates directly into existing CLI coding agents without requiring external graph databases or specialized graph interfaces. Across four benchmarks spanning localization, test generation, and codebase understanding, LARGER improves file-level Acc@5 on LocBench by +13.9 points with tuned hyperparameters and still gains +11.8 points with fixed hyperparameters over the strongest baseline, while delivering consistent gains on MuLocBench, SWE-Atlas Test Writing, and SWE-Atlas Codebase QA.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces LARGER, a lexically anchored active-set retrieval framework for repository context localization in coding agents. It formalizes Lexically Anchored Structural Localization, where lexical matches serve as high-precision structural entry points followed by confidence-filtered local expansion inside the agent's existing search loop. The method integrates directly into CLI coding agents without external graph databases. Empirical results claim file-level Acc@5 gains of +13.9 points (tuned hyperparameters) and +11.8 points (fixed hyperparameters) on LocBench over the strongest baseline, plus consistent improvements on MuLocBench, SWE-Atlas Test Writing, and SWE-Atlas Codebase QA.

Significance. If the reported gains are attributable to the anchoring-plus-expansion logic rather than auxiliary structural information, the approach could meaningfully improve localization for repository-scale coding agents while preserving a single interaction loop. The explicit fixed-hyperparameter control is a methodological strength that helps isolate the contribution of the proposed technique.

major comments (1)
  1. [Formalization of Lexically Anchored Structural Localization and method integration description] The central claim that LARGER performs confidence-filtered local expansion strictly inside the agent's existing search loop using only lexical matches as entry points is load-bearing for attributing the +11.8 / +13.9 Acc@5 gains to the method itself. The manuscript does not explicitly state whether the underlying graph (imports, call chains, type hierarchies, code-test links) is constructed via a separate pre-loop parsing/indexing pass or built dynamically within the loop; if the former, the gains may derive from an implicit structural oracle unavailable to the baselines rather than from the lexically anchored expansion logic.
minor comments (2)
  1. [Experimental evaluation] Baseline details, exact statistical significance tests, and any post-hoc selection criteria for the reported numbers should be expanded in the experimental section to permit independent verification of the quantitative claims.
  2. [Abstract] The abstract's phrasing 'without requiring external graph databases or specialized graph interfaces' does not directly address the possibility of an in-memory pre-computed graph; a clarifying sentence would reduce ambiguity.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and valuable feedback on our manuscript. We address the major comment regarding the formalization and integration details below. We agree that explicit clarification is needed and will revise the paper accordingly.

read point-by-point responses

  1. Referee: The central claim that LARGER performs confidence-filtered local expansion strictly inside the agent's existing search loop using only lexical matches as entry points is load-bearing for attributing the +11.8 / +13.9 Acc@5 gains to the method itself. The manuscript does not explicitly state whether the underlying graph (imports, call chains, type hierarchies, code-test links) is constructed via a separate pre-loop parsing/indexing pass or built dynamically within the loop; if the former, the gains may derive from an implicit structural oracle unavailable to the baselines rather than from the lexically anchored expansion logic.

    Authors: We appreciate the referee's point on the importance of clearly distinguishing the graph construction from the runtime search loop to properly attribute the observed gains. Upon review, the manuscript indeed does not explicitly detail the timing of graph construction. In our implementation, the structural graph (capturing imports, call relations, type hierarchies, and code-test links) is built via a single preprocessing pass using language-specific parsers before the agent begins its task. This pre-built graph is then accessed during the agent's standard search loop for the anchoring and expansion steps. This preprocessing is lightweight, repository-specific, and does not involve external databases or additional agent interactions, aligning with our claim of direct integration into CLI coding agents. The baselines were implemented with access only to lexical search capabilities, without this structural information, to isolate the effect of our lexically anchored expansion. We believe the gains stem from the proposed anchoring and confidence-filtered expansion logic rather than mere access to the graph, as evidenced by the fixed-hyperparameter results. We will add a dedicated subsection in the Method section to describe the graph construction process and its separation from the search loop, ensuring the formalization of Lexically Anchored Structural Localization is unambiguous. revision: yes

Circularity Check

0 steps flagged

No derivation chain present; empirical claims are self-contained

full rationale

The paper introduces the LARGER framework for lexically anchored structural localization in repository-level coding agents and reports empirical benchmark gains (e.g., +13.9 and +11.8 Acc@5 on LocBench). No equations, formal derivations, fitted parameters, or predictions are defined anywhere in the provided text. The central claims rest on experimental results against external benchmarks rather than any self-referential construction, self-citation load-bearing step, or renaming of known results. The method description emphasizes integration into existing agent loops but contains no mathematical reduction that could be inspected for circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, the paper introduces no explicit free parameters, mathematical axioms, or new invented entities. It relies on standard concepts from information retrieval and graph traversal in the domain of code repositories.

pith-pipeline@v0.9.0 · 5790 in / 1222 out tokens · 46233 ms · 2026-05-20T23:31:41.058346+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages

  1. [1]

    Agentless: Demystifying

    Xia, Chunqiu Steven and Deng, Yinlin and Dunn, Soren and Zhang, Lingming , journal=. Agentless: Demystifying

    work page

  2. [2]

    arXiv preprint arXiv:2510.01003 , year=

    Improving code localization with repository memory , author=. arXiv preprint arXiv:2510.01003 , year=

    work page arXiv

  3. [3]

    Hu, Yuntong and Lei, Zhihan and Zhang, Zheng and Pan, Bo and Ling, Chen and Zhao, Liang , journal=

    work page

  4. [4]

    Yang, John and Jimenez, Carlos E and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , journal=

    work page

  5. [5]

    and Shet, Shashank , journal=

    Bairi, Ramakrishna and Sonwane, Atharv and Kanade, Aditya and D C, Vageesh and Iyer, Arun Shankar and Parthasarathy, Suresh and Rajamani, Sriram and Ashok, B. and Shet, Shashank , journal=. 2024 , doi=

    work page 2024

  6. [6]

    Jimenez, Carlos E and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik R , booktitle=

    work page

  7. [7]

    arXiv preprint arXiv:2509.25242 , year=

    A Benchmark for Localizing Code and Non-Code Issues in Software Projects , author=. arXiv preprint arXiv:2509.25242 , year=

    work page arXiv

  8. [8]

    and Tang, Xiangru and Zhuge, Mingchen and others , journal=

    Wang, Xingyao and Li, Boxuan and Song, Yufan and Xu, Frank F. and Tang, Xiangru and Zhuge, Mingchen and others , journal=

    work page

  9. [9]

    2024 , url=

    Liu, Xiangyan and Lan, Bo and Hu, Zhiyuan and Liu, Yang and Zhang, Zhicheng and Wang, Fei and Shieh, Michael Qizhe and Zhou, Wenmeng , journal=. 2024 , url=

    work page 2024

  10. [10]

    2025 , url=

    Yu, Zhongming and Zhang, Hejia and Zhao, Yujie and Huang, Hanxian and Yao, Matrix and Ding, Ke and Zhao, Jishen , journal=. 2025 , url=

    work page 2025

  11. [11]

    Traag, Vincent A and Waltman, Ludo and van Eck, Nees Jan , journal=. From. 2019 , publisher=

    work page 2019

  12. [12]

    Chen, Zhaoling and Tang, Robert and Deng, Gangda and Wu, Fang and Wu, Jialong and Jiang, Zhiwei and Prasanna, Viktor and Cohan, Arman and Wang, Xingyao , booktitle=

    work page

  13. [13]

    2024 , url=

    Ouyang, Siru and Yu, Wenhao and Ma, Kaixin and Xiao, Zilin and Zhang, Zhihan and Jia, Mengzhao and Han, Jiawei and Zhang, Hongming and Yu, Dong , journal=. 2024 , url=

    work page 2024

  14. [14]

    He, Xiaoxin and Tian, Yijun and Sun, Yifei and Chawla, Nitesh V and Laurent, Thomas and LeCun, Yann and Bresson, Xavier and Hooi, Bryan , booktitle=

    work page

  15. [15]

    Learning from Labeled and Unlabeled Data with Label Propagation , author=

    work page

  16. [16]

    2025 , eprint=

    Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG , author=. 2025 , eprint=

    work page 2025

  17. [17]

    2026 , eprint=

    SoK: Agentic Retrieval-Augmented Generation (RAG): Taxonomy, Architectures, Evaluation, and Research Directions , author=. 2026 , eprint=

    work page 2026

  18. [18]

    2024 , eprint=

    Graph Retrieval-Augmented Generation: A Survey , author=. 2024 , eprint=

    work page 2024

  19. [19]

    2025 , eprint=

    When to use Graphs in RAG: A Comprehensive Analysis for Graph Retrieval-Augmented Generation , author=. 2025 , eprint=

    work page 2025

  20. [20]

    2025 , eprint=

    Issue Localization via LLM-Driven Iterative Code Graph Searching , author=. 2025 , eprint=

    work page 2025

  21. [21]

    2025 , eprint=

    GraphCodeAgent: Dual Graph-Guided LLM Agent for Retrieval-Augmented Repo-Level Code Generation , author=. 2025 , eprint=

    work page 2025

  22. [22]

    2025 , eprint=

    Code Graph Model (CGM): A Graph-Integrated Large Language Model for Repository-Level Software Engineering Tasks , author=. 2025 , eprint=

    work page 2025

  23. [23]

    2025 , eprint=

    RANGER -- Repository-Level Agent for Graph-Enhanced Retrieval , author=. 2025 , eprint=

    work page 2025

  24. [24]

    2026 , eprint=

    Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP , author=. 2026 , eprint=

    work page 2026

  25. [25]

    2601.21162 , archivePrefix=

    Liu, Jiate and Chen, Zebin and Qiao, Shaobo and Ju, Mingchen and Zhang, Danting and Han, Bocheng and Yu, Shuyue and Shu, Xin and Wu, Jingling and Wen, Dong and Cao, Xin and Liu, Guanfeng and Yang, Zhengyi , year=. 2601.21162 , archivePrefix=

    work page arXiv

  26. [26]

    2509.22009 , archivePrefix=

    Yang, Cehao and Wu, Xiaojun and Lin, Xueyuan and Xu, Chengjin and Jiang, Xuhui and Sun, Yuanliang and Li, Jia and Xiong, Hui and Guo, Jian , year=. 2509.22009 , archivePrefix=

    work page arXiv

  27. [27]

    One Tool Is Enough: Reinforcement Learning for Repository-Level

    Zhang, Zhaoxi and Duan, Yitong and Zhang, Yanzhi and Xu, Yiming and Wang, Zhixiang and Liang, Kun and Li, Yang and Liang, Jiahui and Xia, Deguo and Huang, Jizhou and He, Jiyan and Wu, Yunfang , year=. One Tool Is Enough: Reinforcement Learning for Repository-Level. 2512.20957 , archivePrefix=

    work page arXiv

  28. [28]

    2512.20482 , archivePrefix=

    Reddy, Revanth Gangi and Liu, Ye and Zhao, Wenting and Doo, JaeHyeok and Suresh, Tarun and Lee, Daniel and Xiong, Caiming and Zhou, Yingbo and Yavuz, Semih and Joty, Shafiq , year=. 2512.20482 , archivePrefix=

    work page arXiv

  29. [29]

    2023 , eprint=

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author=. 2023 , eprint=

    work page 2023

  30. [30]

    2024 , eprint=

    Theoretical Guarantees on the Best-of-n Alignment Policy , author=. 2024 , eprint=

    work page 2024

  31. [31]

    2023 , eprint=

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author=. 2023 , eprint=

    work page 2023

  32. [32]

    2023 , eprint=

    Reasoning with Language Model is Planning with World Model , author=. 2023 , eprint=

    work page 2023

  33. [33]

    2024 , eprint=

    Rational Metareasoning for Large Language Models , author=. 2024 , eprint=

    work page 2024

  34. [34]

    The Probabilistic Relevance Framework:

    Robertson, Stephen and Zaragoza, Hugo , journal=. The Probabilistic Relevance Framework:. 2009 , doi=

    work page 2009

  35. [35]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , year=

    Hierarchical Reward Modeling for Fault Localization in Large Code Repositories , author=. Findings of the Association for Computational Linguistics: EMNLP 2025 , year=. doi:10.18653/v1/2025.findings-emnlp.966 , url=

    work page doi:10.18653/v1/2025.findings-emnlp.966 2025