acceptodds
Under review as a conference paper at ICLR 2027

Credit Where the Cut Happens: Candidate Contraction for Repository-Level Code Localization

1,882 views

Abstract

Repository-level code localization advances mainly by exclusion: an agent rules out code that looks relevant but lies off the fix path. Existing training signals cannot credit such a move: with a free-form answer, a discarded location never appears. We introduce C3L (Candidate Contraction for Code Localization), which makes exclusions first-class learning events. C3L binds the prediction to an explicit survivor set that the agent contracts monotonically from read-only evidence, so every exclusion becomes an irreversible, observable state transition. Two potentials over this state price each cut on the turn it is made: a logarithmic weighted volume rewards effective contraction, and a reachable localization ceiling charges coverage loss the moment a necessary location is discarded, with no learned verifier and no extra rollouts. Because the survivor set summarizes the remaining search, C3L compares turns across rollouts by survivor-set signature even when their histories share nothing, and an execution-verified label cache credits valid repairs outside the reference patch. With a 4B backbone, C3L reaches 59.61/47.60/55.00 mean F1 on SWE-bench Verified, SWE-bench Pro (Python), and LocBench, exceeding the strongest retrained same-size baseline by 6.51/7.22/6.21 points (every 95% CI excludes zero). It outscores a frozen 30B-A3B agent and resolves 40.40% of Verified issues with a fixed repair backend, against 37.33% for the strongest baseline. The gain has three measured sources: the exclusion-aware potentials add 5.9 points over terminal-reward training, grouping turns by survivor signature beats grouping them by action history by 1.35 points, and delivering credit per turn adds 2.11 points at identical total return.

Primary area:
foundation or frontier models, including LLMs
Keywords:
reinforcement learning for LLM agents, credit assignment, repository-level code localization

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_EMu9OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces C3L, a framework for repository-level code localization that treats candidate exclusion as an explicit learning event. An agent progressively contracts a survivor set using read-only repository evidence. Two potential-based rewards encourage effective contraction while penalizing losses in reachable localization accuracy. Survivor-state grouping supports turn-level credit assignment, while an execution-verified label cache accommodates alternative repairs. Experiments with 4B and 8B models across three Python benchmarks demonstrate improvements over retrained localization baselines. The 4B model achieves 59.61% mean F1 on SWE-bench Verified and enables a 40.40% downstream repair rate. Extensive ablations investigate reward design, credit timing, and computational efficiency.

Critical strengths
  • This paper introduces a clear and compelling formulation that makes exclusions observable and directly creditable, addressing an overlooked aspect of repository search.
  • This paper demonstrates consistent improvements across three benchmarks and two model sizes, with matched training conditions, multiple seeds, confidence intervals, and downstream repair evaluation.
  • This paper provides unusually thorough attribution experiments, including equal-return reward-timing controls, alternative grouping strategies, objective comparisons, and compute-matched evaluations.
Critical weaknesses
  • This paper depends on a fixed candidate universe and irreversible pruning, limiting recovery from retrieval misses or premature exclusions. Initial complete reference coverage is only 58.3% on Pro-Python, and increasing the candidate cap improves LocBench performance by 4.98 points.
  • This paper provides limited evidence for the alternative-label cache: its localization improvement has a confidence interval crossing zero, downstream repair gains are also statistically inconclusive, and seven of 100 audited labels exploit weak tests.
  • This paper evaluates exclusively on Python repositories and two closely related model backbones. Its broader applicability to other languages, model families, and search tasks remains unverified; moreover, survivor signatures omit accumulated evidence, so grouped turns need not represent equivalent decision states.
Questions for authors

NA

Reviewer_dgP9OpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work explores how to perform denser credit assignment for SWE code localization agents by maintaining a candidate set at over the whole trajectory, where each step can decide to delete candidates from as the trajectory proceeds. This design provides a natural mechanism to provide denser rewards by looking at step-level decisions.

Critical strengths
  • The formulation is pretty interesting where each step decides if it wants to prune an initial candidate set. This makes it possible to easily design process rewards/step-level rewards that give credit to useful candidate deletions while penalizing bad deletions.
  • The design of maintaining a candidate set at each step also allows performing more finer-grained turn-level comparison for advantage computation, which the authors combine with typical trajectory-level advantages.
  • The method seems to perform substantially better than the baselines in the author's experiments.
Critical weaknesses
  • I found the writing overall a bit hard to read/parse, likely due to use of LLMs to rephrase writing/make things terse. It'd be worth revisiting writing-style.
Questions for authors

None.

Reviewer_EXQ4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

C3L formulates code localization as an irreversible process of narrowing a candidate set. It combines rewards for reducing the search space, penalties for losing relevant locations, and turn-level credit assignment based on the remaining candidates. A training-time cache adds alternative localization labels from test-passing repairs.

Experiments with Qwen3-4B and 8B on SWE-bench Verified, Pro-Python, and LocBench show improvements over retrained baselines. A fixed repair backend also benefits from the resulting locations.

Critical strengths

The method addresses a clear credit-assignment problem. Making exclusions explicit allows the model to receive feedback when it removes a necessary location. The equal-return experiment, which changes reward timing while preserving total return, provides useful evidence for the proposed turn-level credit assignment.

The core experiments are well controlled. Trained baselines share the backbone, data, tools, and SFT initialization. Multiple seeds, confidence intervals, component ablations, and a fixed downstream repair backend strengthen the evidence. For example, 4B C3L achieves a 40.40% repair rate, compared with 37.33% for the strongest baseline.

Critical weaknesses

Model coverage is limited, leaving practical competitiveness unclear. Experiments mainly use Qwen3-4B/8B, with a frozen Qwen3-Coder-30B-A3B as the larger-model reference. Comparisons with newer small open models and frontier open or closed models are missing. It remains unclear whether stronger small models already achieve similar performance, and whether C3L’s gains carry over to those backbones. Quality and inference-cost comparisons under comparable budgets would help establish the value of the additional training.

Comparisons with existing localization systems remain incomplete. RepoNavigator also trains code localization models through RL, but is not directly evaluated, and its omission is not explained. The retrained CodeScout and FuseSearch variants also change parts of their original tools, initialization, or interaction settings. These shared settings help compare training signals, but leave uncertainty about performance against the original systems. The authors disclose these changes and provide some supporting controls, but their effects deserve further examination.

Questions for authors

1.Can the authors evaluate newer small open models and test whether C3L’s gains persist on stronger backbones? How does it compare with frontier models in localization quality and inference cost under comparable budgets?

2.Why was RepoNavigator [1] omitted? Could it be evaluated using the same backbone, data, and budget?

3.How much do the shared tools and mixed SFT initialization affect CodeScout and FuseSearch? What evidence shows that the adaptations preserve their effectiveness?

Overall, the paper presents a clear method with substantial internal validation, while its competitiveness against broader alternatives needs further evidence. I would consider raising my score if additional experiments address these concerns.

[1] One Tool Is Enough: Reinforcement Learning for Repository-Level LLM Agents. arXiv:2512.20957.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.