Tag
MemScope 是一个由中国科学院提出的框架，通过工具使用的 Patch Agent 检查原始执行轨迹、运行时内存状态和现有架构代码，实现对智能体记忆架构的证据驱动、局部化原位进化，在 GAIA、xBench 和 ALFWorld 上均取得显著提升。
PreciseCoder 提出在单元测试奖励之外加入语义验证的局部编辑奖励信号，显著减少模型在调试时的多余改动并防止幻觉式 bug 修改；在 4B/9B/27B 模型上将平均编辑行数从 7-15 行降至 3 行，并能泛化到仓库级程序修复，其推理轨迹还能提升 GPT-5.6 Sol 等前沿模型的编辑精度。
RSI-Anything 提出一个基于 JEPA-Anything 的领域通用递归自我改进框架，利用预测性世界模型指导行动与学习数据选择，在代码修复、GUI 任务、实验室规划和反应优化等多个领域均取得显著性能提升。
Rules to Tools (R2T) studies whether giving LLM coding agents ready-to-run executable checks of written scientific requirements improves program repair on SciCode-style tasks. Results are mixed and task-dependent: the tool group slightly outperforms text-based checks in some cohorts (29/30 vs 26/30 complete repair), while others tie, and checks can lower agent-side model output costs but raise CPU use.
Failure Map is an open archive of 20,168 compact Python debugging tasks across 254 categories, with contracts, broken implementations, and executable boundary checks to benchmark local coding models on code repair. The author provides sample cases, a standardized prompt template, program baselines, and evaluation guidelines, noting no local-model scores have been measured yet.
论文提出 Rules to Tools (R2T)，为科学计算中的 LLM 代码修复代理提供可执行的公开需求检查工具，并在 SciCode 基准上对比纯文本说明与可调用检查的修复效果与成本。
The paper investigates whether Large Language Models fix bugs by making minimal edits or reimplementing solutions from scratch in competitive programming, finding that LLMs often modify more than necessary and solve better when generating new solutions, which has implications for debugging tools.
ScienceIDE introduces infrastructure for converting scientific code repositories into programmable environments for scientific agents, enabling task generation, execution, and verification, with trained models showing improvements in scientific code repair and general capabilities.
A repair tool that infers fixes from a single worked example and successfully repairs 5 out of 5 bugs in unseen code using only 4 CPU instructions and zero tokens, with exhaustive verification.
Proposes Anchored Self-Play (ASP), a method for scaling code repair supervision via generator–fixer self-play with an embedding-similarity reward and reference bug mixing, achieving +24% relative improvement in fix rates over standard self-play on a new benchmark BugSourceBench.
PatchWorld introduces a gradient-free framework that transforms offline trajectories into executable Python world models via counterexample-guided code repair, enabling interpretable and inspectable belief-state programs for planning in partially observable environments.
This paper proposes a signal reshaping method for Group Relative Policy Optimization (GRPO) to improve weak-feedback agentic code repair, showing significant gains in compile and semantic accuracy.