Skip to main content
Search

# long-horizon software engineering

4 related articles

Subscribe to updates
  1. arXiv

    Rejected patches stop being wasted: failure feedback enters training, lifting SWE-bench first-attempt resolution from 38.9% to 41.7%

    Repository-level software engineering agents previously treated each attempt's pass or fail as a terminal reward, discarding the failed patch and verifier feedback. FC-SWE restores the repository to its original task state after a patch fails verification, uses the failed patch and verifier diagnostics as context for a recovery trajectory, keeps each trajectory's own verifier outcome, and compares only the initial and recovery trajectories actually executed for the same issue. On all 500 SWE-bench Verified tasks with the same Qwen3.5-4B policy, first-attempt resolution rises from GRPO's 38.9% to 41.7% and two-attempt cumulative resolution from 48.5% to 52.8%; trained with at most two attempts per chain, the same policy reaches 70.7% with eleven test-time attempts.

    Tsinghua University
    Jia Liufu（First author）
    FC-SWEGRPOlong-horizon software engineeringSWE-bench VerifiedCoding agent
  2. Gemini

    Google announces Gemini 4 Argon: 1M output tokens, 77.9% on DeepSWE v1.1, and a claimed 2.7x faster Rust decoder for libgav1

    Google announced Gemini 4 Argon, a frontier model first rolling out to trusted cyber defenders through its Fairwind Program, built for long-horizon deep reasoning, raising the output token limit from 64K to 1M and reporting 77.9% on DeepSWE v1.1, 91.7% on LVBench, 68% on CWE-bench v1, and 51.3% on AutomationBench, alongside internal cases including a 40% improvement over a published baseline in quantum subroutine optimization, over 300 TiB of memory freed in data centers, and a Rust port of libgav1 that replaced 32K lines of SIMD code and runs 2.7x faster than the prior Rust port with identical video output.

    Google
    Gemini 4 Argoncybersecurity defenselong-horizon software engineeringPrompt Injection
  3. arXiv

    WideSWE tests coding agents on 120 cross-repository tasks: best configuration solves only 42.50%, and joint execution beats independent runs on requests

    Drawing on 1,729,171 pull requests across 103 software ecosystems, the authors built WideSWE, 120 real tasks (60 bug fixes, 60 features) that require coordinated changes across multiple repositories, and evaluated seven coding-agent configurations: full task success ranges from 10.83% to 42.50%, with Codex CLI paired with GPT-5.6-sol highest, while joint and per-repository independent execution each help in different ways.

    Zhejiang UniversityTsinghua University
    Baoyi Wang（First author）
    WideSWECoding AgentsAgent evaluationlong-horizon software engineering
  4. arXiv

    QwenGyre pairs elastic GPU scheduling with trajectory-tree processing to lift Qwen 3.8 2.4T from 52.5% to 58.5% on NL2RepoBench in 48 steps

    QwenGyre is an end-to-end framework for extreme-long-horizon (xLong) online agent RL that elastically reallocates GPUs between rollout and training without interrupting live harness executions, while its trajectory processor reconstructs branching executions into shared-prefix trajectory trees, scores partial progress after timeouts, and deduplicates sampled trajectories by role priority; scaled to Qwen 3.8 2.4T with 700K tokens per rollout, it raises NL2RepoBench pass rate from 52.5% to 58.5% in 48 steps and delivers up to 1.85x and 1.78x end-to-end speedups over Colocate and Async.

    QwenAlibaba Group (China)Tsinghua University
    Weiqi Wang（First author）
    QwenGyreAgentic Reinforcement LearningTrajectory ProcessingRL Post-Traininglong-horizon software engineering

You’re all caught up