FeatLens: Feature-Guided Dynamic Code Graph Construction and Retrieval for Repository-Level Code Generation
Xutian Li, Bo Xiong, Yifeng Zhu, Kunze Li, Xianlin Zhao, Runbang Yan, Yanzhen Zou, Lu Zhang, et al.
cs.SE · cs.AI
Quick verdict
FeatLens: feature index + dynamic seed graph + personalized PageRank for repo-level code retrieval; strong DR@15 gains, but no ablation or significance tests.
Summary
FeatLens is a retrieval method for repository-level code generation that uses software features as anchors between natural-language requirements and reusable code dependencies. Offline, it hierarchically clusters function embeddings and summarizes each cluster into a feature description (adapting RepoSummary). At task time, it retrieves feature clusters aligned with the requirement, builds a task-specific seed graph from the associated functions, expands it along calls and imports, and applies personalized PageRank (following HippoRAG) with semantic-structural scoring to select a compact reasoning graph, requiring no LLM calls during retrieval. On DevEval (90 Python projects, 1,825 tasks) FeatLens reaches DR@15 of 0.501 versus 0.430 for CodexGraph, 0.415 for UniXcoder RAG, 0.296 for BM25, and 0.127 for RepoGraph; on EvoCodeBench (5 projects) it reaches 0.460. For generation on DevEval it attains the highest DIR@1 (52.91% with DeepSeek-V3.2, 53.58% with GPT-5-mini), keeps Pass@1 competitive (CodexGraph slightly higher on GPT-5-mini: 56.35% vs 55.03%), and generates shorter code (101.1 vs 167.7 LOC on GPT-5-mini). Against CodexGraph it cuts graph nodes by 61.0%, edges by 86.2%, and total token overhead by 45.9%. Main limitations: no ablation isolating each stage's contribution (acknowledged in Section 6.1), no significance testing, Python-only evaluation, and no stated code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 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 feature index linking natural-language feature descriptions to function-level entities serves as an effective retrieval anchor for repository-level dependency retrieval.
- FeatLens achieves the highest DR@15 among the compared baselines on DevEval (0.501) and EvoCodeBench (0.460).
- On DevEval generation, FeatLens attains the highest DIR@1 (52.91% with DeepSeek-V3.2, 53.58% with GPT-5-mini) with competitive Pass@1 and shorter generated code.
- Relative to CodexGraph, FeatLens reduces graph nodes by 61.0%, edges by 86.2%, and total token overhead by 45.9%, using zero LLM tokens during retrieval.
Strengths
Consistent retrieval gains across benchmarks and context budgets
Figure 3: FeatLens has the best DR at every Top-N on both datasets; DevEval DR@10 0.461 already nearly matches the best baseline's DR@20 0.466 (CodexGraph), and the diminishing-returns analysis in Section 5.1.1 (+4.0 then +1.6 points) supports the compact-context thesis.
Genuine efficiency win with zero retrieval-time LLM use
Table 4: 15 nodes/13.7 edges vs CodexGraph 38.5/99.3 (−61.0%/−86.2%), 0 vs 3,851.3 retrieval tokens, −45.9% total tokens, while DR@15 rises 0.430→0.501; offline indexing is cheap (269.4–316.6 s, $0.21–0.30 per repo, Table 2).
Retrieval gains translate into dependency reuse in generated code
Concrete failure-mode analysis motivating the design
Weaknesses
No ablation isolating component contributions
No statistical testing or variance reporting
Baseline fairness is under-established
Thin external validity and undiscussed index maintenance
Under-specified method with no released artifacts
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
FeatLens (arXiv 2609.26480, cs.SE/cs.AI) is a method paper on dependency retrieval for repository-level code generation. Its thesis: software features — natural-language functional units — are better retrieval anchors than raw text similarity or repository-wide graphs. The pipeline (Section 3) has three stages: an offline feature index that clusters functions and summarizes each cluster; a task-time seed graph built from feature clusters aligned with the requirement plus the target function location; and semantic-structural reranking with personalized PageRank producing a compact reasoning graph used as LLM context. Three headline claims: (1) best DR@15 among compared baselines on DevEval (0.501) and EvoCodeBench (0.460); (2) highest DIR@1 on DevEval generation (52.91% with DeepSeek-V3.2, 53.58% with GPT-5-mini) with competitive Pass@1 and shorter code; (3) versus CodexGraph, 61.0% fewer nodes, 86.2% fewer edges, 45.9% lower total token overhead, and zero LLM tokens during retrieval.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 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 FeatLens and what problem does it solve?
FeatLens is a retrieval method for repository-level code generation. It links natural-language feature descriptions to functions through an offline feature index, then builds a small task-specific code graph and filters it with personalized PageRank, so the LLM receives only the dependencies it should reuse. It targets the problem that text-similarity retrieval and whole-repository graph methods return noisy or incomplete context at high construction and token cost.
How much better is FeatLens than existing retrieval baselines?
On DevEval it reports DR@15 of 0.501 versus 0.430 for CodexGraph, 0.415 for UniXcoder RAG, 0.296 for BM25, and 0.127 for RepoGraph; on EvoCodeBench 0.460 versus 0.362 for the best baseline. These are point estimates from single runs — the paper reports no significance tests.
Does FeatLens use an LLM during retrieval?
No. Retrieval is deterministic (feature matching, graph expansion, personalized PageRank), consuming zero LLM tokens per task. LLMs are used offline to build the feature index (about $0.21–0.30 per repository) and afterwards to generate the code patch from the selected context.
Does FeatLens improve code correctness (Pass@1)?
Not over the strongest baseline. On DevEval with GPT-5-mini, Pass@1 is 55.03% versus CodexGraph's 56.35%. FeatLens does achieve the highest DIR@1 (dependency reuse, 53.58% vs 46.27%) and much shorter code (101.1 vs 167.7 lines), so its benefit is reuse and conciseness rather than raw pass rate.
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.
- Figures 1–3 were extracted as garbled text; Figure 3's bar values were recovered from embedded numerals and cross-checked against the prose, but Figure 2's architecture internals (e.g., 'Weighted Fusion') cannot be analyzed from the extraction.
- The supplied metadata lists 8 authors while the extracted text includes a ninth (Bing Xie) — a minor metadata inconsistency, not affecting the assessment.
- No appendix or supplementary material was available; hyperparameters, prompts, and scoring formulas may exist outside the extracted 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.