Yeung, Henry · Nov 3, 2026
KG-AST for Cheaper CI/CD: Budgeted Test Selection and AI Bug Review
Markdown · JSON · Original source
Henry Yeung’s paper studies KG-AST, which uses git history and code-graph features to rank commits by risk before heavy tests or AI review run. It also gives the AI reviewer a graph slice around a change, rather than only the diff. In a study of 155 open-source projects across 11 languages, heavy testing of the riskiest 20% of commits covered 49–64% of bug-introducing commits, depending on language; random selection covered 20%, and testing the biggest change first covered 40–55%. In 20 Java projects, the reported coverage was 64.6% versus 49.5% for the biggest-change baseline, with the model ahead in 19 projects. For AI review, exact matches to later-fixed defects rose from 26% with only the diff to 32% with graph context; about half of unmatched findings were judged real by independent AI judges. A static call-graph selector ran about 8% of a test suite while including a test written for the fix in 85% of 67 fix commits. The paper also reports that Poincare embedding added nothing measurable. These results describe the study’s measures, not measured CI minutes: the budget metric assumes a heavy test catches the bug it covers, and the authors note noisy labels.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 3, 2026 (2026-11-03).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T23:03:46.242Z. First observed: 2026-11-06T05:10:39.352Z. Retrieved: 2026-11-08T23:02:47.470Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The authors say the labels are noisy and that the budget metric assumes a heavy test catches the bug it covers. The authors report that no CI minutes were measured, so the budget results are not measurements of actual CI . The abstract does not establish how the reported results would transfer to projects outside the studied 155 open-source projects.
Open question
How would KG-AST’s budgeted results compare when measured in actual CI minutes, which the paper says it did not measure?
Source-stated
KG-AST ranks commits using git history and code-graph features, then allocates heavy tests or AI review to higher-risk commits; its AI reviewer receives a graph slice around the change.
scores each commit risk from git history and code-graph features before heavy tests or AI review run
Source-stated
Across 155 open-source projects in 11 languages, testing the riskiest 20% covered 49–64% of bug-introducing commits, compared with 20% for random selection and 40–55% for testing the biggest change first.
On 155 open-source projects in 11 languages, a single cross-language model trained on other projects ranks commits so that heavy testing of the riskiest 20% covers 49-64% of bug-introducing commits depending on the language, against 20% for random selection and 40-55% for testing the biggest change
Source-stated
In 20 Java projects, the reported gain at a 20% budget was 15.1 points, and the model won in 19 projects.
In 20 Java projects the gain at a 20% budget is 15.1 points (64.6% vs 49.5%, 95% interval +11.6 to +18.4), and the model wins in 19 of the 20 projects.
Source-stated
AI review matched 26% of later-fixed defects exactly using only a commit’s diff and 32% with graph context; about half of unmatched findings were judged real by independent AI judges.
An AI reviewer that sees only a commit's diff finds 26% of later-fixed defects exactly (32% with graph context), and about half of its findings that did not match a later fix were judged real by independent AI judges.
Source-stated
The authors report that Poincare embedding added nothing measurable in the evaluated tasks.
The Poincare embedding used in our retrieval work adds nothing measurable here: commit-risk AUC, bug review and severity ranking are unchanged or worse.
Source-stated
The study’s budget metric assumes a heavy test catches the bug it covers, and the authors did not measure CI minutes.
The labels are noisy and the budget metric assumes a heavy test catches the bug it covers; no CI minutes were measured.
Source and contributor context
Yeung, Henry · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 9d54815c-40e5-4e29-8708-e924b8a3d13f.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T23:03:46.242Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b225a1f13b1861b8733e4aaf6589078c5f5bcf1a467270840ef84933191d3f5d. Source text hash: 250589ea364dc117d5fe11f6d4b5745615594fdbd1a6b95feb64dd0ccbc6399d.
Revision and source-check history
Version 1 · 2026-11-08T23:03:46.242Z
KG-AST for Cheaper CI/CD: Budgeted Test Selection and AI Bug Review
Henry Yeung’s paper studies KG-AST, which uses git history and code-graph features to rank commits by risk before heavy tests or AI review run. It also gives the AI reviewer a graph slice around a change, rather than only the diff. In a study of 155 open-source projects across 11 languages, heavy testing of the riskiest 20% of commits covered 49–64% of bug-introducing commits, depending on language; random selection covered 20%, and testing the biggest change first covered 40–55%. In 20 Java projects, the reported coverage was 64.6% versus 49.5% for the biggest-change baseline, with the model ahead in 19 projects. For AI review, exact matches to later-fixed defects rose from 26% with only the diff to 32% with graph context; about half of unmatched findings were judged real by independent AI judges. A static call-graph selector ran about 8% of a test suite while including a test written for the fix in 85% of 67 fix commits. The paper also reports that Poincare embedding added nothing measurable. These results describe the study’s measures, not measured CI minutes: the budget metric assumes a heavy test catches the bug it covers, and the authors note noisy labels.
Original source
cs.AI · Article
Source details
Original source date: Nov 3, 2026 (2026-11-03).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T23:03:46.242Z. First observed: 2026-11-06T05:10:39.352Z. Retrieved: 2026-11-08T23:02:47.470Z.
Limits
The authors say the labels are noisy and that the budget metric assumes a heavy test catches the bug it covers. The authors report that no CI minutes were measured, so the budget results are not measurements of actual CI . The abstract does not establish how the reported results would transfer to projects outside the studied 155 open-source projects.
Open question
How would KG-AST’s budgeted results compare when measured in actual CI minutes, which the paper says it did not measure?
Source-stated
KG-AST ranks commits using git history and code-graph features, then allocates heavy tests or AI review to higher-risk commits; its AI reviewer receives a graph slice around the change.
scores each commit risk from git history and code-graph features before heavy tests or AI review run
Source-stated
Across 155 open-source projects in 11 languages, testing the riskiest 20% covered 49–64% of bug-introducing commits, compared with 20% for random selection and 40–55% for testing the biggest change first.
On 155 open-source projects in 11 languages, a single cross-language model trained on other projects ranks commits so that heavy testing of the riskiest 20% covers 49-64% of bug-introducing commits depending on the language, against 20% for random selection and 40-55% for testing the biggest change
Source-stated
In 20 Java projects, the reported gain at a 20% budget was 15.1 points, and the model won in 19 projects.
In 20 Java projects the gain at a 20% budget is 15.1 points (64.6% vs 49.5%, 95% interval +11.6 to +18.4), and the model wins in 19 of the 20 projects.
Source-stated
AI review matched 26% of later-fixed defects exactly using only a commit’s diff and 32% with graph context; about half of unmatched findings were judged real by independent AI judges.
An AI reviewer that sees only a commit's diff finds 26% of later-fixed defects exactly (32% with graph context), and about half of its findings that did not match a later fix were judged real by independent AI judges.
Source-stated
The authors report that Poincare embedding added nothing measurable in the evaluated tasks.
The Poincare embedding used in our retrieval work adds nothing measurable here: commit-risk AUC, bug review and severity ranking are unchanged or worse.
Source-stated
The study’s budget metric assumes a heavy test catches the bug it covers, and the authors did not measure CI minutes.
The labels are noisy and the budget metric assumes a heavy test catches the bug it covers; no CI minutes were measured.
Source and contributor context
Yeung, Henry · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 9d54815c-40e5-4e29-8708-e924b8a3d13f.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T23:03:46.242Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: b225a1f13b1861b8733e4aaf6589078c5f5bcf1a467270840ef84933191d3f5d. Source text hash: 250589ea364dc117d5fe11f6d4b5745615594fdbd1a6b95feb64dd0ccbc6399d.
uncertain · 2026-11-08T23:03:46.242Z · anthropic/claude-opus-5.5
Publication: 9d54815c-40e5-4e29-8708-e924b8a3d13f
attribution — pass: The paper is correctly attributed to Henry Yeung on arXiv, and its results are kept as reported findings. Minor fix: the source lists a single author, so 'the authors note noisy labels' and the plural in the limitations should read 'the author'.
Authors:Henry Yeung
summary — uncertain: The numbers are faithful, but a non-specialist would struggle with central terms left unexplained: CI/CD, commit, diff, git history, code-graph features, call-graph selector, Poincare embedding and AUC. Tooltips or brief glosses would fix this. The reading also lists results without saying why they matter: spending scarce testing and review effort on the commits most likely to add bugs. It drops a key condition: a single cross-language model trained on other projects. The phrase 'rather than only the diff' slightly reframes the source's 'only a graph slice', which describes limiting the reviewer's context.
a single cross-language model trained on other projects ranks commits
date — pass: The originalPublished date of 2026-11-03 matches the arXiv v1 submission date. The later feed date is not used.
[Submitted on 3 Nov 2026]
support — pass: Every claim and figure matches the abstract: 49-64% coverage, 20% for random selection, 40-55% for the biggest-change baseline, 64.6% vs 49.5% in Java with wins in 19 of 20 projects, 26%/32% for AI review, about 8% of the suite with 85% of 67 fix commits, and no gain from Poincare. The limitations and the CI-minutes question follow from the source's own caveat and add no unsupported benefits.
The labels are noisy and the budget metric assumes a heavy test catches the bug it covers; no CI minutes were measured.
media — pass: The selected media is the verified direct link to the arXiv abstract page, which is the original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIME