PaperlayerView exact revision on arXiv

Source-linked AI summary

Kimi-Dev: Agentless Training as Skill Prior for SWE-Agents

Zonghan Yang, Shengjie Wang, Kelin Fu, Wenyang He, Weimin Xiong, Yibo Liu, Yibo Miao, Bofei Gao, Yejie Wang, Yingwei Ma, Yanhao Li, Yue Liu, Zhenxing Hu, Kaitai Zhang, Shuyi Wang, Huarong Chen, Flood Sung, Yang Liu, Yang Gao, Zhilin Yang, Tianyu Liu

arXiv:2509.23045v3cs.AIcs.CLcs.SE

TL;DR

SWE-bench solutions are divided between multi-turn SWE-Agents and workflow-based Agentless methods, which the paper argues need not be mutually exclusive. It uses Agentless training to induce skill priors for SWE-Agent adaptation; Kimi-Dev reaches 60.4% on SWE-bench Verified, and its SWE-Agent adaptation reaches 48.6% pass@1, comparable to Claude 3.5 Sonnet (241022 version).

  • Problem

    SWE-bench work uses either multi-turn SWE-Agents or single-turn Agentless workflows, while the potential for Agentless training to support SWE-Agent adaptation remains the paper’s central question.

    Source: §1 · pp. 1–2

  • Method

    The paper develops an Agentless training recipe to induce localization, code-repair, and self-reflection skill priors, then uses them to adapt Kimi-Dev to SWE-Agents.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    Kimi-Dev reaches 60.4% on SWE-bench Verified among workflow approaches, while its SWE-Agent adaptation reaches 48.6% pass@1, comparable to Claude 3.5 Sonnet (241022 version).

    Source: §1 · p. 2, §4.1 · p. 8

  • Takeaways & Limitations

    The findings frame Agentless training and agentic frameworks as complementary stages for building coding models with transferable skills.

    Source: §5 · p. 10

    The reflection-skill analysis is coarse-grained because measured gains also capture redo behaviors and SWE-Agent test-writing; isolating the TestWriter prior is left for future work.

    Source: §4.2 · p. 10

Abstract

from arXiv · show

Large Language Models (LLMs) are increasingly applied to software engineering (SWE), with SWE-bench as a key benchmark. Solutions are split into SWE-Agent frameworks with multi-turn interactions and workflow-based Agentless methods with single-turn verifiable steps. We argue these paradigms are not mutually exclusive: reasoning-intensive Agentless training induces skill priors, including localization, code edit, and self-reflection that enable efficient and effective SWE-Agent adaptation. In this work, we first curate the Agentless training recipe and present Kimi-Dev, an open-source SWE LLM achieving 60.4\% on SWE-bench Verified, the best among workflow approaches. With additional SFT adaptation on 5k publicly-available trajectories, Kimi-Dev powers SWE-Agents to 48.6\% pass@1, on par with that of Claude 3.5 Sonnet (241022 version). These results show that structured skill priors from Agentless training can bridge workflow and agentic frameworks for transferable coding agents.

1 INTRODUCTION

SWE-bench solutions are commonly divided between multi-turn SWE-Agents and fixed Agentless workflows, which are often treated as mutually exclusive. The paper argues that Agentless training can instead induce skills for adapting SWE-Agents, and presents Kimi-Dev as evidence for this complementary approach.

1 INTRODUCTION

  • SWE-bench evaluates whether a patch for a real-world GitHub issue passes the corresponding unit tests.The paper identifies the benchmark’s difficulty, outcome reward, and real-world economic value as reasons for its prominence.Source: §1 · p. 1
  • Agent-based solutions such as SWE-Agent and OpenHands use multi-turn interaction, while Agentless workflows decompose tasks into single-turn steps with verifiable rewards (Yang et al., 2024a; Wang et al., 2025a; Xia et al., 2024).The introduction presents these approaches as offering different strengths and limitations, rather than as interchangeable training setups.Source: §1 · pp. 1–2
  • Agentless training is proposed as a source of skill priors, including bug localization, code repair, self-reflection, and verification, for more efficient SWE-Agent adaptation.The paper frames these capabilities as scaffolding for agentic interaction, rather than treating Agentless training as the final deliverable.Source: §1 · p. 2
  • 60.4% on SWE-bench Verified makes Kimi-Dev the state-of-the-art workflow-based solution; its Agentless training also supports adaptation to SWE-Agents.With 5k publicly available trajectories, the adapted model reaches 48.6% pass@1, comparable to Claude 3.5 Sonnet (20241022 version).Source: §1 · p. 2

2 BACKGROUND

The background contrasts modular Agentless workflows with adaptable but demanding multi-turn SWE-Agents, then motivates Agentless training as a way to build priors for agentic training. It highlights trajectory collection, long-horizon learning, and model initialization as challenges for SWE-Agent training.

2 BACKGROUND

  • The background introduces the framework dichotomy and training challenges as context for reinterpreting Agentless training as skill priors for SWE-Agents.The paper develops this theme in later sections.Source: §2 · p. 2

2.1 FRAMEWORK DICHOTOMY

  • Agentless workflows modularize localization, repair, and test generation into steps that can be optimized separately with verifiable rewards, but struggle with multi-round updates (Xia et al., 2024; Wei et al., 2025; Ma et al., 2025a;b; Xie et al., 2025).Their rigidity trades flexibility for modularity and stability.Source: §2.1 · p. 2
  • SWE-Agents iteratively plan, act, and reflect, enabling greater adaptability but requiring the model to manage exploration, reasoning, tools, and potentially lengthy interaction histories (Yang et al., 2024a; Wang et al., 2025a).Trajectories can extend tens or hundreds of steps, placing demands on context windows.Source: §2.1 · p. 2
  • The paper challenges the view that Agentless and agentic frameworks are mutually exclusive, arguing that Agentless-induced skill priors make SWE-Agent training more stable and efficient.The two approaches are framed as complementary despite their different trade-offs.Source: §2.1 · pp. 2–3

2.2 TRAINING SWE-AGENTS

  • Collecting high-quality SWE-Agent trajectories requires executable environments, substantial human labor, and costly frontier-model calls, making data collection slow and resource-demanding (Pan et al., 2024).Some studies synthesize bugs to scale environment construction (Jain et al., 2025; Yang et al., 2025c).Source: §2.2 · p. 3
  • Sparse outcome rewards, instability in long-horizon reinforcement learning, and sensitivity to initialization make SWE-Agent training challenging (Pan et al., 2024; Yang et al., 2025c).Generic pretrained models can exhibit brittle tool use or get stuck in repeated action patterns; the paper proposes Agentless training to induce priors for localization, repair, test composition, and self-reflection.Source: §2.2 · p. 3

3 AGENTLESS TRAINING RECIPE

Kimi-Dev’s Agentless recipe trains BugFixer and TestWriter skills through mid-training, cold-start, reinforcement learning, and test-time self-play. The resulting model reaches 60.4% on SWE-bench Verified, while its skills also support adaptation to agentic workflows.

3.1 FRAMEWORK: THE DUO OF BUGFIXER AND TESTWRITER

  • BugFixer repairs software logic, while TestWriter creates tests that reproduce reported bugs; both depend on file localization and code editing.A successful resolution requires the patch to pass the issue’s tests, while a high-quality test fails before the fix and passes afterward.Source: §3.1 · p. 3

3.2 MID-TRAINING & COLD START

  • Mid-training uses real-world GitHub issues and commits, followed by cold-start SFT on reasoning trajectories to develop analysis, self-refinement, and solution exploration.The mid-training recipe includes Agentless-derived natural-diff patches, curated PR commit packs, and synthetic reasoning and interaction data.Source: §3.2 · p. 4

3.3 REINFORCEMENT LEARNING

  • RL focuses on code editing after localization is strong, using executable environments and a REINFORCE-based policy optimization approach.The recipe uses execution outcomes as rewards and gradually reintroduces harder prompts as performance improves.Source: §3.3 · p. 4

3.4 TEST-TIME SELF-PLAY

  • BugFixer patches are scored using reproduction and regression tests, and the highest-scoring patch is selected as the answer.Self-play evaluates each candidate patch against generated tests both before and after applying the patch.Source: §3.4 · p. 5

3.5.1 MAIN RESULTS

  • 60.4% on SWE-bench Verified makes Kimi-Dev state-of-the-art among open-source models under Agentless-like frameworks.Table 1 compares models under the standard 40-patch, 40-test setting, except Llama3-SWE-RL, which uses 500 patches and 30 tests.Source: §3.5.1 · p. 5, Table 1 · p. 5

3.5.2 MID-TRAINING

  • More mid-training tokens consistently improve BugFixer pass@1 after cold-start activation.The evaluation compares models trained on 50B, 100B, and approximately 150B tokens.Source: Figure 2 · p. 5, §3.5.2 · pp. 5–6

3.5.3 REINFORCEMENT LEARNING

  • During RL, BugFixer pass rate and TestWriter reproduced rate rise alongside response length, with similar scaling curves observed on Qwen2.5-14B-Instruct.The authors report occasional TestWriter false positives due to limited reproduction coverage.Source: §3.5.3 · p. 6, Figure 3 · p. 6

3.5.4 TEST-TIME SELF-PLAY

  • 60.4% on SWE-bench Verified is reached as generated patch-test pairs increase from 1×1 to 40×40, consistently exceeding BugFixer majority voting.Three patches and three tests per instance already surpass majority voting over 40 BugFixer patches, while self-play remains below pass@N using ground-truth tests.Source: Figure 4 · p. 6, §3.5.4 · pp. 6–7

4 INITIALIZING SWE-AGENTS FROM AGENTLESS TRAINING

The paper tests whether Agentless-trained Kimi-Dev can serve as a prior for multi-turn SWE-Agents, using lightweight SWE-Agent fine-tuning and further transfer analyses. It reports strong adaptation performance and evidence of data efficiency, longer-turn progress, skill transfer, and end-to-end RL potential.

4.1 PERFORMANCE AFTER SWE-AGENT FINE-TUNING

  • The study fine-tunes Kimi-Dev on 5,016 SWE-Agent trajectories and compares Base, mid-trained, cold-start SFT, and RL priors across trajectory-token budgets.The trajectories were collected with Claude 3.7 Sonnet in synthetic environments; the sweep includes zero-shot and one-step settings as well as nested data subsets.Source: §4.1 · pp. 7–8, §4.2 · p. 8, Figure 5 · p. 8
  • 48.6% pass@1 on SWE-bench Verified is achieved without additional test-time scaling, realistic-environment trajectories, or multi-turn agentic RL, comparable to Claude 3.5 Sonnet’s 49%.The adapted model also reaches 74.0% pass@10, above its Agentless pass@30 of 73.8%.Source: §4.1 · p. 8

4.2 SKILL TRANSFER AND GENERALIZATION

  • After fine-tuning, the RL prior continues making progress beyond 70 turns, while SFT, MT, and Base show diminishing returns around 70, 60, and 50 turns.The comparison uses pass@3 under varying turn limits.Source: Figure 6 · p. 9, §4.2 · p. 9
  • The RL prior leads in nearly all SWE-Agent SFT settings; it needs 223 tokens to reach Base’s top pass@1, while Base uses 1.5 × 228 tokens.The paper interprets this as stronger adaptation efficiency from the Agentless training recipe.Source: §4.2 · p. 9
  • After 200 trajectories, the MT prior becomes comparable to SFT and RL, despite lagging in zero-shot and one-step settings.These settings correspond to 0 and 221 SFT tokens, respectively.Source: §4.2 · p. 9
  • The SFT prior beats RL zero-shot but degrades with 200 trajectories; the authors suggest overfitting to a collapsed data mode as a possible explanation.They contrast this with RL’s stronger transferable skills and generalization in that setting.Source: §4.2 · p. 9
  • The RL prior has the strongest combined BugFixer and reflection results, with Stage-3 resolutions rising from 484 for Base to 605 for RL and reflection gains from +94 to +113.The reflection measure is coarse-grained: its gains also include intermediate test-writing, so it does not isolate reflection and redo behaviors.Source: §4.2 · pp. 9–10
  • With the same 221 SWE-Agent SFT tokens, end-to-end SWE-Agent RL slightly favors the RL prior over SFT at pass@1, pass@3, and pass@5.The authors conclude that Agentless training provides strong priors for end-to-end SWE-Agent learning with minimal proprietary end-to-end trajectory supervision.Source: Figure 7 · p. 10, §4.2 · p. 10

5 CONCLUSION AND FUTURE WORK

The conclusion presents Agentless and agentic SWE paradigms as complementary. It attributes transferable skills and strong benchmark performance to structured Agentless training, which also supports efficient SWE-Agent adaptation.

5 CONCLUSION AND FUTURE WORK

  • The paper reframes Agentless and agentic paradigms for automated software engineering as complementary rather than competing.This positions staged training as a route toward more generalizable coding agents.Source: §5 · p. 10
  • Kimi-Dev’s structured Agentless training is reported to induce transferable bug-localization, code-repair, and self-reflection skill priors.Source: §5 · p. 10
  • Kimi-Dev achieves state-of-the-art SWE-bench Verified results among workflow approaches and enables efficient SWE-Agent adaptation.Source: §5 · p. 10

ETHICS AND REPRODUCIBILITY STATEMENTS

The ethics and reproducibility statement reports ethics-code compliance, privacy treatment, and locations of detailed experimental setups.

ETHICS AND REPRODUCIBILITY STATEMENTS

A DETAILS OF MID-TRAINING

This part describes the mid-training data recipe, combining filtered GitHub pull-request data with synthetic localization reasoning and simulated agent interactions, including editing-stage trajectories derived from successful localization rollouts and ground-truth PR changes. The data is structured to support both direct patch generation and agentic reasoning-and-action learning.

A DETAILS OF MID-TRAINING

  • GitHub PR data is drawn from repositories with at least five stars, excludes SWE-bench repositories, and retains merged PRs with pre-change codebase snapshots.Abandoned, superseded, and still-under-review PRs are discarded.Source: A DETAILS OF MID-TRAINING · p. 15
  • Natural diff patches consolidate PR commits into final SEARCH–REPLACE edits, while PR commit packs preserve the sequence of commit messages and corresponding code changes.Natural diffs are restricted to Python, Markdown, and reStructuredText-related PRs, with Python edits limited to three files and five SEARCH–REPLACE blocks; commit packs allow at most five Python files and remove developer signatures and GitHub IDs.Source: A DETAILS OF MID-TRAINING · pp. 15, 18
  • The filtered recipe yields ∼50B tokens of natural diff patches and ∼20B tokens of PR commit packs.These are complementary data forms: final SEARCH–REPLACE patches and sequences pairing human commit messages with code modifications.Source: A DETAILS OF MID-TRAINING · pp. 15, 18
  • Approximately ∼10B tokens of synthetic localization reasoning are generated by retaining only exactly correct file-localization rollouts.A Qwen-2.5-72B-Instruct model is lightly SFT-trained on 2,000 R1 trajectories before generating rollouts for BugFixer and TestWriter.Source: A DETAILS OF MID-TRAINING · p. 19
  • Synthetic agentic rollouts use file-viewing and keyword-search tools without shell execution, training the model on actions and observations along the trajectory.Because repository snapshots may lack executable Docker environments, the system prompt is loss-masked while the model learns both policy and world modeling.Source: A DETAILS OF MID-TRAINING · p. 19
  • Stage 2 retains Stage 1 rollouts with at least one correctly localized file and injects false-positive contexts to strengthen self-reflection about files that need no modification.The agent is allowed to open incorrectly localized files before encountering reasoning such as deciding not to edit them.Source: A DETAILS OF MID-TRAINING · p. 22
  • Mid-training starts from Qwen2.5-72B-Base with 4× upsampling of synthetic data, a batch size of 256, 32K-token sequences, and a 2e-5 learning rate decaying to 2e-6 by approximately 150B tokens.Training uses next-token prediction with cosine decay and approximately 3B tokens of warm-up.Source: A DETAILS OF MID-TRAINING · p. 22

B DOCKER ENVIRONMENTS

The authors construct executable Docker environments to validate generated patches and support cold-start and reinforcement-learning training. Their configuration process fixes dependencies, selects Python versions, installs requirements, resolves missing modules, and checks test transitions.

B DOCKER ENVIRONMENTS

  • Table 3 lists the sources of the Docker environments used to develop Kimi-Dev.Source: Table 3 · p. 22
  • For datasets without Docker support, the authors configure environments by fixing dependencies, selecting Python by commit year, installing requirements, and resolving missing modules.They used publicly available Docker environments for SWE-Gym and R2E-Gym-Lite, and automated configuration for SWE-Bench-Extra.Source: B DOCKER ENVIRONMENTS · p. 23
  • 3,846 of 6.38k SWE-bench-extra instances received successfully constructed environments for cold-start and RL training.Construction success was validated using FAIL TO PASS tests that transition from failing without the ground-truth patch to passing with it.Source: B DOCKER ENVIRONMENTS · p. 23

C MORE DETAILS OF RL TRAINING

RL training uses an adaptive prompt set that begins with problems the model can partially solve, then progressively adds problems it cannot yet solve. This schedule forms a curriculum that increases difficulty.

C.1 PROMPT SET SELECTION

  • The initial RL prompt set contains 1,200 problems with pass@16 > 0, selected from SWE-Gym, SWE-bench-extra, and R2E-gym.Source: §C.1 · p. 23
  • Every 100 training steps, the prompt set expands with 500 randomly sampled problems filtered from those where the current model has pass@16 = 0.Source: §C.1 · p. 23
  • Adding currently unsolved problems progressively increases difficulty and forms a curriculum.Source: §C.1 · p. 23

C.2 RL EXPERIMENT ABLATION

This section covers RL experiments on scaling behavior in BugFixer and TestWriter and positive-example reinforcement for BugFixer, alongside a partial BugFixer trace showing column insertion.

C.2 RL EXPERIMENT ABLATION

  • Both BugFixer and TestWriter exhibit clear scaling-law behavior in Qwen2.5-14B RL experiments.Figure 9 presents the scaling experiments.Source: §C.2 · p. 23, Figure 9 · p. 23
  • Positive-example reinforcement improves 72B BugFixer RL performance over training without the technique.The improvement was incorporated during a later training stage.Source: §C.2 · p. 23, Figure 8 · p. 23
  • Positive examples reinforce correct reasoning and accelerate convergence, especially as exploration ability diminishes later in training; a similar trick was used for Seed1.5-thinking (Seed et al., 2025).Source: §C.2 · p. 24

E.1 BUGFIXER TRACES

  • The column-insertion helper expands the column count and maps indices before, within, and after the insertion point to the original or inserted matrix.The supplied code fragments do not report ablation measurements.Source: §E.1 · p. 28

E.2 TESTWRITER FALSE POSITIVE CASES

TestWriter can produce incomplete tests that pass its reward check despite failing to reliably judge BugFixer patches. Two examples show distinct coverage gaps in Django and SymPy.

E.2 TESTWRITER FALSE POSITIVE CASES

  • TestWriter false positives arise when generated tests have coverage gaps, so reward-based performance improvements may not reflect actual performance.The authors leave further TestWriter enhancement for future work.Source: §E.2 · p. 28
  • In Django django-16877, the generated test checked only one of two locations and omitted mark safe behavior, yet passed the reward check.This produced a false positive.Source: §E.2 · p. 28
  • In SymPy sympy-13974, the generated test omitted associativity coverage, making it an unreliable judge for BugFixer.The passage gives tensor-product powers as an example of the untested property.Source: §E.2 · p. 28

F EMERGENT PARALLEL SCALING IN KIMI-DEV

This section covers Kimi-Dev’s parallel scaling through comparison and aggregation of independent rollout patches, alongside implementation snippets and examples concerning gabbi test-data path resolution for YAML files and stdin.

F EMERGENT PARALLEL SCALING IN KIMI-DEV

  • Parallel scaling emerges without specialized training: BugFixer rollout patches are supplied as candidates for the model to compare and combine into a new bugfix.The model is prompted to summarize differences among the candidates and compose a patch.Source: F EMERGENT PARALLEL SCALING IN KIMI-DEV · p. 31
  • Performance improves as more patch candidates are included, and model-based aggregation surpasses majority voting by analyzing candidate differences rather than answer frequency.The reported results are on SWE-bench Verified.Source: Figure 10 · p. 31, F EMERGENT PARALLEL SCALING IN KIMI-DEV · p. 32
  • The runner adds a test_directory argument and passes it into test-suite construction.The snippets also show callers supplying the directory derived from each input file’s absolute path.Source: F EMERGENT PARALLEL SCALING IN KIMI-DEV · p. 42
  • The proposed fix sets the test directory from the input file’s location and joins it with the full data filename.The model’s patch changes runner.py to pass the input file’s directory and case.py to preserve subdirectories in the filename.Source: F EMERGENT PARALLEL SCALING IN KIMI-DEV · pp. 47–48
  • The model recognizes that data paths should be resolved relative to the YAML file, not the current working directory.It identifies stripping directories with os.path.basename as the cause of incorrect resolution for paths containing subdirectories.Source: F EMERGENT PARALLEL SCALING IN KIMI-DEV · p. 47
  • When YAML is read from stdin, the model leaves the test directory as the current directory because the YAML file’s location is unavailable.The passage notes that redirected input does not expose the original filename, so relative-path resolution cannot use its directory.Source: F EMERGENT PARALLEL SCALING IN KIMI-DEV · p. 48
  • YAML-referenced data files now resolve relative to the YAML file’s directory, supporting flat and nested structures, while stdin runs retain current-working-directory behavior.The change passes a test directory through `run_suite` and preserves directory structure when resolving data-file paths.Source: F EMERGENT PARALLEL SCALING IN KIMI-DEV · p. 50

G ANALYSIS FOR SWE-AGENTS

This section describes how an LLM annotates SWE-Agent interaction turns by stage, using accumulated trajectory context because stage boundaries are not strict. The stages can recur as agents test, revise, and reflect.

G.1 STAGE ANNOTATION FOR SWE-AGENT TRAJECTORIES

  • The suggested workflow includes locating relevant code, reproducing the error, editing the source, rerunning the reproducer, and checking edge cases.The prompt asks the agent to make minimal changes to non-test files and not modify tests.Source: §G.1 · pp. 50–52
  • Agents can move backward between stages, returning to code repair when rerun tests or edge-case tests still fail.The examples include transitions from Stage 4 to Stage 3 and from Stage 5 to Stage 3.Source: §G.1 · p. 51
  • The annotator classifies each interaction turn into one of five SWE-Agent stages, using the agent’s prior actions, observations, and earlier classifications as context.Annotation proceeds turn by turn, with each observation–action pair appended as input.Source: §G.1 · pp. 50–53
  • Because stages have no strict boundaries or transition markers, the study uses an LLM to label trajectory turns.The annotation prompt incorporates most SWE-Agent tool descriptions and instructions.Source: §G.1 · pp. 50–51

G.2 COMPARATIVE STUDY

The comparative study examines Kimi-Dev on sympy-20590, where finding the correct parent class is difficult without issue-discussion hints. Agentless localization misses the target, while one RL-prior SWE-Agent trajectory reaches the fix through reflection.

G.2 COMPARATIVE STUDY

  • Developer discussion hints pointed to adding empty slots to DefaultPrinting, but those hints were unavailable during problem solving.The correct repair led to the printer implementation file, which was also changed by the ground-truth patch.Source: §G.2 · pp. 53–54
  • None of 40 Agentless runs localized the correct file; most chose sympy/core/symbol.py, while a few chose other incorrect files.The problem’s symptom involved Symbol, making symbol.py plausible, but the correct repair concerned a parent class.Source: §G.2 · pp. 53–54
  • Every adapted SWE-Agent trajectory included a test-reproducing phase, and all agents’ initial code repairs failed before rerunning the script exposed the unresolved issue.The study describes the script as providing an unambiguous signal of whether the issue was fixed.Source: §G.2 · p. 54
  • Only the RL-prior SWE-Agent adaptation succeeded on this instance, using reflection to identify Printable as the class missing __slots__.The agent checked the class hierarchy, then reran the test script and observed that it passed.Source: §G.2 · pp. 54, 55, 57

G.3 PATTERN COMPARISON BETWEEN SFT WITH EXPERT DATA AND END-TO-END RL

This section qualitatively compares interaction patterns from heavy expert-data SFT and light SFT plus end-to-end RL on an instance both methods solve. It describes differences in reasoning regulation, exploration, and code-edit reasoning, and discusses the persistence of Agentless skill priors in the agentic setting.

G.3 PATTERN COMPARISON BETWEEN SFT WITH EXPERT DATA AND END-TO-END RL

  • In the compared instance, the light-SFT-plus-RL trajectory repeatedly inspects bin-calculation code and considers whether density=True overrides the specified range.The model revisits the relevant code and its reasoning shifts as it examines how bins and range are handled.Source: §G.3 · pp. 57, 60, 61
  • Heavy SFT identifies that density=True drops range from hist_kwargs, fixes the bug, and verifies the fix across the listed histogram edge cases.Tests cover specified and absent ranges, density on and off, explicit bin sequences, and ranges extending beyond the data.Source: §G.3 · pp. 58–59
  • Heavy SFT produces regulated reasoning with explicitly labeled steps 1 to 5 and generally short reasoning content.The passage attributes this pattern to training on proprietary-model data.Source: §G.3 · p. 67
  • Light SFT plus end-to-end RL produces less-regulated interactions, with more repo exploration and more tokens at code-edit steps before action.The passage contrasts this pattern with heavy SFT.Source: §G.3 · p. 67
  • The authors interpret the light-SFT-plus-RL pattern as reflecting end-to-end RL and retaining reasoning-intensive skill priors in the agentic setting.They say this behavior is less impacted by expert data.Source: §G.3 · p. 67

G.4 GENERALIZATION TO OTHER TASKS

The analysis tests whether Agentless-induced skill priors transfer to SWE-bench-Live and SWE-bench Multilingual. SFT- and RL-enhanced priors show stronger generalization than Base, especially when adaptation data is scarce.

G.4 GENERALIZATION TO OTHER TASKS

  • SWE-bench-Live uses a monthly updated curation pipeline; this study selected its default 300 tasks collected between October 2024 and March 2025.The authors describe it as having greater distributional shift than SWE-bench Verified.Source: §G.4 · p. 67
  • SWE-bench Multilingual contains 300 curated tasks from 42 repositories across 9 programming languages, with greater linguistic and domain diversity than Python-only SWE-bench Verified.Its diversity poses additional challenges for cross-language generalization and transferability.Source: §G.4 · p. 67
  • The study compares Base, mid-trained, SFT, and RL priors, using open-source SWE-Smith trajectories to activate agentic capabilities across varied adaptation amounts.Each adaptation uses lightweight supervised finetuning, taking from several minutes to at most two hours.Source: §G.4 · p. 67
  • SFT- and RL-enhanced priors generalize better than Base, especially with scarce adaptation data, while Base and MT approach their performance as more trajectories are used.The authors attribute the narrowing gap potentially to differences among SWE tasks and leave stronger out-of-distribution, task-agnostic generalization recipes for future work.Source: §G.4 · p. 67, Figure 11 · p. 68, Figure 12 · p. 68

H USE OF LARGE LANGUAGE MODELS

  • The authors report that the initial draft was written by them and that GPT-5 assisted only with language polishing.They attribute the conceptual contributions, experimental designs, analyses, and conclusions to the authors.Source: H USE OF LARGE LANGUAGE MODELS · p. 68
Loading 2509.23045v3…