RepoAtlas: Guiding Coding Agents via Evolving Multimodal Repository Views
Yunxiang Zhang, Haiquan Wang, JiaWei Guo, Hanyang Xia, Yan Chen, Tong Chen, Zhang Zhiwei, Junchen Ye
cs.SE · cs.AI
Quick verdict
Training-free select–project–refresh graph views lift SWE-bench resolve ~2.4 pts at lower cost; solid ablations, but no significance tests or code release.
Summary
RepoAtlas is a training-free module that gives LLM coding agents an evolving multimodal view of a repository code graph through a select–project–refresh loop. An offline typed graph (modules, classes, callables, variables; containment, call, import, and use edges) is built once per snapshot. Selection fuses lexical, semantic, and trajectory evidence via node-wise maximum, propagates relevance with personalized PageRank, and assembles a connected region under a fixed budget (15 nodes/20 edges) using supernodes with representative exposure. Projection pairs an adaptively chosen visual layout with a source-grounded textual index, switching to text-only once editing begins, while a state-aware controller refreshes the view when exploration makes it stale. On all 500 SWE-bench Verified instances with three vision-language models, RepoAtlas reaches 63.1% (Qwen3.6-35B-A3B), 68.0% (MiMo-V2.5), and 72.4% (Kimi-K2.5) resolve rates, averaging +2.4 points over the strongest multimodal baseline (SeeRepo) with 5.8% fewer input tokens and 7.8% fewer model calls. Ablations (Table 3) show every stage contributes, and a LocBench selection study (Table 4) isolates gains from PageRank propagation and representative exposure. Main limitations: no significance testing or variance reporting across the three runs, silently mixed comparators in Section 5.2's efficiency percentages, no stated code release, and a demonstrated blind spot for runtime-only dependencies.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A training-free select–project–refresh loop over a repository code graph yields evolving multimodal views that improve repository-level issue resolution.
- On SWE-bench Verified, RepoAtlas improves resolve rate by 2.4 points on average over the strongest multimodal baseline (SeeRepo) while cutting input tokens by 5.8% and model calls by 7.8%.
- Gains hold across three VLM families and scales: 63.1% (Qwen3.6-35B-A3B), 68.0% (MiMo-V2.5), 72.4% (Kimi-K2.5).
- Both modalities contribute: removing the textual channel costs 3.9 points and the visual channel 1.9 points on Qwen3.6-35B-A3B.
- State-aware refresh is critical: no refresh drops 4.7 points and per-call refresh drops 12.1 points relative to the adaptive controller.
Strengths
Clear problem formalization and positioning
Section 3 models repository context as evolving view maintenance Gt=(Vt,Et,Mt) with explicit budgets; Table 1 differentiates RepoAtlas from five prior interface families along view selection, adaptive projection, and automatic refresh, exposing exactly which capabilities are novel.
Disciplined efficiency accounting
Table 2 counts image tokens and cached tokens in full and compares tokens/calls/cost only within each model, avoiding cross-model cost confounds; per-model margins in prose (e.g., Qwen3.6: 63.1% vs 60.8% SeeRepo, -7.9% tokens) reconcile with the table.
Stage-isolated ablations and selection study
Honest failure analysis
Weaknesses
No statistical inference despite a 3-run protocol
Silently mixed comparators in Section 5.2 percentages
Reproducibility gap
Narrow baseline pool and fairness question on SeeRepo
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
RepoAtlas (arXiv 2609.16936, Beihang University) is a method paper for repository-level issue resolution by LLM agents. Its central claim: a training-free module that maintains an *evolving multimodal repository view* — a select–project–refresh loop over an offline code graph — beats both text-graph and visual-graph interfaces on SWE-bench Verified while consuming fewer resources. Headline numbers: +2.4 resolve points on average over the strongest multimodal baseline (SeeRepo), with 5.8% fewer input tokens and 7.8% fewer model calls, across three vision-language models. Scope is limited to issue resolution; other repository tasks are explicitly deferred to future work.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is RepoAtlas?
RepoAtlas is a training-free module for LLM coding agents that maintains an evolving multimodal view of a repository: it selects a task-relevant region of a pre-built code graph under a fixed budget, renders it as a visual graph plus a text index with exact code locations, and refreshes the view when the agent's exploration makes it outdated. It runs through a select–project–refresh loop and requires no model fine-tuning.
How much does RepoAtlas improve results on SWE-bench Verified?
Across all 500 SWE-bench Verified instances and three vision-language models, it reports resolve rates of 63.1% (Qwen3.6-35B-A3B), 68.0% (MiMo-V2.5), and 72.4% (Kimi-K2.5). That averages +2.4 percentage points over the strongest multimodal baseline (SeeRepo), while using about 5.8% fewer input tokens and 7.8% fewer model calls.
Does RepoAtlas require fine-tuning the underlying model?
No. RepoAtlas is explicitly training-free: it works by constructing a repository code graph, selecting and rendering views, and feeding them to existing vision-language models through prompts. The paper evaluates it on three model families without any architectural changes or fine-tuning.
What are the main limitations of RepoAtlas?
The static code graph cannot capture dependencies that are only established at runtime — in one documented failure case (sympy-17630) the agent never opened the correct file across 8 runs. Also, the paper reports no statistical significance tests across its three evaluation runs, does not state a code release, and evaluates only on issue-resolution tasks.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The appendix (evidence calibration, selection implementation details, refresh state transitions, budget sensitivity, full per-run results) is not included in the extracted text; several design details could not be verified.
- Figures 1-3 appear only as garbled text renderings; visual layout quality and the case-study figures cannot be assessed from this extraction.
- Baseline numbers depend on the authors' reproductions of SeeRepo and LocAgent; equivalence to the original papers' settings is not verifiable from the provided text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.