Hennara, Khalil, Hreden, Muhammad, Aldallal, Zeina, Chrouf, Sara, AlModhayan, Safwan · Nov 9, 2026
arXiv cs.AI: Let the Fixer Decide: Training a Code Localizer from the Fixer's Likelihood
Markdown · JSON · Original source
The paper’s authors propose Fixer-Guided Localization (FGL), which trains a code localizer to select files based on whether a frozen fixer considers them useful for producing a verified patch. They argue that a fixer may need to read code it does not edit, unlike a localizer trained only to predict which files a reference patch changes. Their pipeline scores candidate localizations without running tests or using , then trains the localizer from those scores. In an experiment using Gemma 4 E4B and trajectories from a coding-agent competition, FGL’s target helped the fixer find bugs as often as the teacher’s localization while using less code. On the development set, the authors report that one EM round improved several task counts against a matched control; the single agent solved more tasks than the base model. They also report more exploration and fewer wrong patches. Separately, the authors say their analysis of a 4B agent under an eight-minute budget identifies , repetition loops and edit quality as constraints. The abstract does not establish how well the approach transfers to other coding tasks or settings. The authors say data, code and adapters are publicly available.
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 9, 2026 (2026-11-09).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T09:20:08.944Z. First observed: 2026-11-10T08:36:26.850Z. Retrieved: 2026-11-12T09:18:36.476Z.
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 results are from a specific setup: Gemma 4 E4B, trajectories for the Kaggle Gemma-4 Developer Agent competition, and a development-set comparison. The abstract does not establish whether the results transfer to other coding tasks or settings. The evidence provided is the paper’s abstract, not its full methods or results; it does not provide further detail for assessing the experiments or reported comparisons.
Open question
Does Fixer-Guided Localization also help code agents on tasks beyond the Kaggle competition described in the paper?
Source-stated
The authors propose training a code localizer using a frozen fixer's scores of candidate localizations, based on how likely each is to produce a verified patch.
the localizer proposes several localizations, a frozen fixer scores each by how likely it makes a verified patch
Source-stated
The authors distinguish the code a fixer needs to read from the code it changes.
the code a fixer needs to read is not the same as the code it changes.
Source-stated
Their method uses weighted supervised fine-tuning and does not run tests or use reinforcement learning.
the localizer is trained on these scores with weighted supervised fine-tuning, without running tests and without reinforcement learning.
Source-stated
In the reported setup, FGL's target helped the fixer find the bug as often as the teacher's localization while using 27.5% less code.
The FGL target helps the fixer find the bug as often as the teacher's localization while using 27.5% less code.
Source-stated
On the development set, the authors report that one EM round moved the agent in the expected direction compared with a matched control, including more solved tasks.
one EM round moves the agent in the expected direction compared with a matched control: more tasks reach a patch (53 vs. 46), fewer tasks get stuck reading files that do not exist (43 vs. 54), and more tasks are solved (8 vs. 6).
Source-stated
The authors report that the trained agents explored more and wrote fewer wrong patches, and that one agent solved more tasks than the base model.
trained agents explore about three times more than the base model and write far fewer wrong patches, and the single agent solves up to 11 of 78 tasks, against 8 for the base model.
Source-stated
The authors identify time, repetition loops and edit quality as constraints for a 4B agent under an 8-minute budget.
identifies what limits a 4B agent under an 8-minute budget (time, repetition loops and edit quality)
Source and contributor context
Hennara, Khalil, Hreden, Muhammad, Aldallal, Zeina, Chrouf, Sara, AlModhayan, Safwan · 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.
Reinforcement learning
Training by trial and error, in which a system is rewarded for the outcomes its designers want.
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: d7989f09-982b-403f-819f-4fee02187e56.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T09:20:08.944Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 35dd3e30bf8e891a11ee350822b57f1cc7a84fd2f4939aed650f0dab885486ec. Source text hash: e05ddefd4e22b40927dd38f0d6f0ce39e69ff5798b6488e0a8c6f1ef1c213a0a.
Revision and source-check history
Version 1 · 2026-11-12T09:20:08.944Z
arXiv cs.AI: Let the Fixer Decide: Training a Code Localizer from the Fixer's Likelihood
The paper’s authors propose Fixer-Guided Localization (FGL), which trains a code localizer to select files based on whether a frozen fixer considers them useful for producing a verified patch. They argue that a fixer may need to read code it does not edit, unlike a localizer trained only to predict which files a reference patch changes. Their pipeline scores candidate localizations without running tests or using , then trains the localizer from those scores. In an experiment using Gemma 4 E4B and trajectories from a coding-agent competition, FGL’s target helped the fixer find bugs as often as the teacher’s localization while using less code. On the development set, the authors report that one EM round improved several task counts against a matched control; the single agent solved more tasks than the base model. They also report more exploration and fewer wrong patches. Separately, the authors say their analysis of a 4B agent under an eight-minute budget identifies , repetition loops and edit quality as constraints. The abstract does not establish how well the approach transfers to other coding tasks or settings. The authors say data, code and adapters are publicly available.
Original source
cs.AI · Article
Source details
Original source date: Nov 9, 2026 (2026-11-09).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-12T09:20:08.944Z. First observed: 2026-11-10T08:36:26.850Z. Retrieved: 2026-11-12T09:18:36.476Z.
Limits
The reported results are from a specific setup: Gemma 4 E4B, trajectories for the Kaggle Gemma-4 Developer Agent competition, and a development-set comparison. The abstract does not establish whether the results transfer to other coding tasks or settings. The evidence provided is the paper’s abstract, not its full methods or results; it does not provide further detail for assessing the experiments or reported comparisons.
Open question
Does Fixer-Guided Localization also help code agents on tasks beyond the Kaggle competition described in the paper?
Source-stated
The authors propose training a code localizer using a frozen fixer's scores of candidate localizations, based on how likely each is to produce a verified patch.
the localizer proposes several localizations, a frozen fixer scores each by how likely it makes a verified patch
Source-stated
The authors distinguish the code a fixer needs to read from the code it changes.
the code a fixer needs to read is not the same as the code it changes.
Source-stated
Their method uses weighted supervised fine-tuning and does not run tests or use reinforcement learning.
the localizer is trained on these scores with weighted supervised fine-tuning, without running tests and without reinforcement learning.
Source-stated
In the reported setup, FGL's target helped the fixer find the bug as often as the teacher's localization while using 27.5% less code.
The FGL target helps the fixer find the bug as often as the teacher's localization while using 27.5% less code.
Source-stated
On the development set, the authors report that one EM round moved the agent in the expected direction compared with a matched control, including more solved tasks.
one EM round moves the agent in the expected direction compared with a matched control: more tasks reach a patch (53 vs. 46), fewer tasks get stuck reading files that do not exist (43 vs. 54), and more tasks are solved (8 vs. 6).
Source-stated
The authors report that the trained agents explored more and wrote fewer wrong patches, and that one agent solved more tasks than the base model.
trained agents explore about three times more than the base model and write far fewer wrong patches, and the single agent solves up to 11 of 78 tasks, against 8 for the base model.
Source-stated
The authors identify time, repetition loops and edit quality as constraints for a 4B agent under an 8-minute budget.
identifies what limits a 4B agent under an 8-minute budget (time, repetition loops and edit quality)
Source and contributor context
Hennara, Khalil, Hreden, Muhammad, Aldallal, Zeina, Chrouf, Sara, AlModhayan, Safwan · 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.
Reinforcement learning
Training by trial and error, in which a system is rewarded for the outcomes its designers want.
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: d7989f09-982b-403f-819f-4fee02187e56.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-12T09:20:08.944Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 35dd3e30bf8e891a11ee350822b57f1cc7a84fd2f4939aed650f0dab885486ec. Source text hash: e05ddefd4e22b40927dd38f0d6f0ce39e69ff5798b6488e0a8c6f1ef1c213a0a.
uncertain · 2026-11-12T09:20:08.944Z · anthropic/claude-opus-5.5
Publication: d7989f09-982b-403f-819f-4fee02187e56
attribution — pass: The reading attributes the proposal and results to the paper's authors on arXiv cs.AI and keeps the reported outcomes framed as their claims.
Authors:Khalil Hennara, Muhammad Hreden, Zeina Aldallal, Sara Chrouf, Safwan AlModhayan
summary — uncertain: Central terms are not explained for a general reader: code localizer, fixer, frozen, patch, and especially 'one EM round', which appears with no gloss. Smallest repair: add plain explanations or tooltip entries. 'Trajectories from a coding-agent competition' also blurs the source: the trajectories are 1,629 test-verified Kimi K3 runs, and the pipeline was built for the Kaggle competition. Repair: say 'test-verified Kimi K3 trajectories, for the Kaggle Gemma-4 Developer Agent competition'. The reading also drops the 27.5% figure and the 11-of-78 versus 8 counts, which weakens how concrete the result is.
We build the full pipeline on Gemma 4 E4B for the Kaggle Gemma-4 Developer Agent competition, from 1,629 test-verified Kimi K3 trajectories.
date — pass: originalPublished 2026-11-09 matches the v1 submission date shown on the arXiv page.
[Submitted on 9 Nov 2026]
support — pass: Each claim matches its verbatim excerpt. The limitations about setup specificity and abstract-only evidence are appropriately conditional. The question does not assume an answer, and the reading adds no unsupported benefits.
the single agent solves up to 11 of 78 tasks, against 8 for the base model
media — pass: The selected media is the verified direct link to the arXiv abstract page for this paper.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Training by trial and error, in which a system is rewarded for the outcomes its designers want.
Definition referenceTIME 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 arXivTraining by trial and error, in which a system is rewarded for the outcomes its designers want.
Definition referenceTIME 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 arXiv