Daxerer, Christoph, Glock, Johann, Macho, Christian, Ramler, Rudolf, Pinzger, Martin, Pichler, Josef · Oct 26, 2026
Do Coding Agents Still Need a Debugger? Frontier-Model-Based Fault Localization at the Accuracy Ceiling
Markdown · JSON · Original source
The paper’s authors examine whether coding agents need debugger assistance to locate faults in short programs. They report that agents based on Sol 6.1, Luna 6, and Qwen 3.8 localized nearly all seeded faults from source code alone, including in behavior-preserving transformed programs. Encouraging debugger assistance removed some remaining errors for Luna 6 and Qwen 3.8, but used substantially more tokens; with optional assistance, agents used the interactive debugger in few runs. The results suggest debugger assistance can still help on these tasks, though the study does not establish how well the findings apply to larger or less precisely specified software. How would debugger assistance affect fault localization in programs beyond the study’s short, precisely specified examples?
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: Oct 26, 2026 (2026-10-26).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T05:18:16.806Z. First observed: 2026-10-28T08:08:36.476Z. Retrieved: 2026-10-30T04:02:14.651Z.
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 reported findings concern seeded faults in short, precisely specified programs; the evidence does not establish how well they generalize to larger or less precisely specified software.
Open question
How would debugger assistance affect fault localization in programs beyond the study’s short, precisely specified examples?
Source-stated
The authors seeded single-line faults into 144 LeetCodeDataset solutions and rewrote each faulty program using five behavior-preserving transformation recipes.
We seed 610 single-line faults into 144 LeetCodeDataset solutions and rewrite each faulty program with five recipes of behavior-preserving transformations.
Source-stated
Without tools, the tested agents localized 98.2 to 99.9% of faults in original programs and 96.7 to 99.9% in transformed programs.
Without tools, the agents localized 98.2 to 99.9% of faults in original and 96.7 to 99.9% in transformed programs
Source-stated
Encouraged debugger assistance removed some remaining transformed-program errors for Luna 6 and Qwen 3.8, with higher token use; Sol 6.1 remained at 99.9%.
Encouraged assistance removed 63% and 61% of Luna 6's and Qwen 3.8's remaining transformed errors (+2.1 and +1.3 percentage points) at 14.7 to 27.0 times the token use, while Sol 6.1 stayed at 99.9%.
Source-stated
With optional assistance, the agents gained accuracy while using the interactive debugger in a small share of runs.
With optional assistance, Luna 6 and Qwen 3.8 still gained +1.5 and +1.3 points, though the agents used the interactive debugger in only 0.2 to 6.2% of runs.
Source and contributor context
Daxerer, Christoph, Glock, Johann, Macho, Christian, Ramler, Rudolf, Pinzger, Martin, Pichler, Josef · 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.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 161d5461-67f9-4c1c-8718-97a7b8a8c3cf.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T05:18:16.806Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 5494383e003781ddd1c646d0d9731ce91c651620e0e38fe2bae9f2dd4cfdce87. Source text hash: 817590a0c330450f7b843d7221c42b316cea0a6fd3b03858a61a3ec8680499d8.
Revision and source-check history
Version 1 · 2026-11-14T05:18:16.806Z
Do Coding Agents Still Need a Debugger? Frontier-Model-Based Fault Localization at the Accuracy Ceiling
The paper’s authors examine whether coding agents need debugger assistance to locate faults in short programs. They report that agents based on Sol 6.1, Luna 6, and Qwen 3.8 localized nearly all seeded faults from source code alone, including in behavior-preserving transformed programs. Encouraging debugger assistance removed some remaining errors for Luna 6 and Qwen 3.8, but used substantially more tokens; with optional assistance, agents used the interactive debugger in few runs. The results suggest debugger assistance can still help on these tasks, though the study does not establish how well the findings apply to larger or less precisely specified software. How would debugger assistance affect fault localization in programs beyond the study’s short, precisely specified examples?
Original source
cs.AI · Article
Source details
Original source date: Oct 26, 2026 (2026-10-26).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-14T05:18:16.806Z. First observed: 2026-10-28T08:08:36.476Z. Retrieved: 2026-10-30T04:02:14.651Z.
Limits
The reported findings concern seeded faults in short, precisely specified programs; the evidence does not establish how well they generalize to larger or less precisely specified software.
Open question
How would debugger assistance affect fault localization in programs beyond the study’s short, precisely specified examples?
Source-stated
The authors seeded single-line faults into 144 LeetCodeDataset solutions and rewrote each faulty program using five behavior-preserving transformation recipes.
We seed 610 single-line faults into 144 LeetCodeDataset solutions and rewrite each faulty program with five recipes of behavior-preserving transformations.
Source-stated
Without tools, the tested agents localized 98.2 to 99.9% of faults in original programs and 96.7 to 99.9% in transformed programs.
Without tools, the agents localized 98.2 to 99.9% of faults in original and 96.7 to 99.9% in transformed programs
Source-stated
Encouraged debugger assistance removed some remaining transformed-program errors for Luna 6 and Qwen 3.8, with higher token use; Sol 6.1 remained at 99.9%.
Encouraged assistance removed 63% and 61% of Luna 6's and Qwen 3.8's remaining transformed errors (+2.1 and +1.3 percentage points) at 14.7 to 27.0 times the token use, while Sol 6.1 stayed at 99.9%.
Source-stated
With optional assistance, the agents gained accuracy while using the interactive debugger in a small share of runs.
With optional assistance, Luna 6 and Qwen 3.8 still gained +1.5 and +1.3 points, though the agents used the interactive debugger in only 0.2 to 6.2% of runs.
Source and contributor context
Daxerer, Christoph, Glock, Johann, Macho, Christian, Ramler, Rudolf, Pinzger, Martin, Pichler, Josef · 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.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 161d5461-67f9-4c1c-8718-97a7b8a8c3cf.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T05:18:16.806Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 5494383e003781ddd1c646d0d9731ce91c651620e0e38fe2bae9f2dd4cfdce87. Source text hash: 817590a0c330450f7b843d7221c42b316cea0a6fd3b03858a61a3ec8680499d8.
uncertain · 2026-11-14T05:18:16.806Z · anthropic/claude-opus-5.5
Publication: 161d5461-67f9-4c1c-8718-97a7b8a8c3cf
attribution — pass: The findings are attributed to the paper's authors. The tested systems (Sol 6.1, Luna 6, Qwen 3.8) and the results are presented as the authors' reported findings, not as established fact.
Authors:Christoph Daxerer, Johann Glock, Christian Macho, Rudolf Ramler, Martin Pinzger, Josef Pichler
summary — uncertain: The summary flattens the source's force. Repair: say encouraged assistance removed most (61-63%) remaining errors. Also add the authors' closing point that agents ran the program far more than they stepped through it yet kept most of the benefit; the optional-assistance gain is missing from the summary. Central terms (fault localization, seeded faults, behavior-preserving transformations, tokens as compute cost) are not explained in plain language. A non-specialist may miss what was tested and why token use matters.
debugger assistance removes most errors that remain. Given the choice, agents executed the program far more often than they stepped through it, yet kept most of the benefit.
date — pass: originalPublished 2026-10-26 matches the arXiv v1 submission date. The feed date of Oct 29 is later and is correctly not used.
[Submitted on 26 Oct 2026]
support — uncertain: Claim 4 says the agents gained accuracy with optional assistance, but the source reports gains only for Luna 6 and Qwen 3.8. Repair: name those two models. Claim 3's 'some remaining errors' understates the cited 63% and 61%. The limitation and the question are supported by the source's own scoping to short, precisely specified programs.
With optional assistance, Luna 6 and Qwen 3.8 still gained +1.5 and +1.3 points
media — pass: The selected media is the verified arXiv abstract page for this paper, which is the direct 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 arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv