Tag
A research post introducing 'reactive optimizers': an agent watches training dynamics and rewrites the optimizer's source code during training, matching or beating the best hand-tuned recipes on image and language tasks while autonomously guarding against loss spikes and overfitting. The method is released via the open-source OPAL package.
TraceForge 提出一种两阶段（监督+对抗）方法，从教师答案中生成匹配学生模型自身风格的推理轨迹，无需访问隐藏思维链，在 HumanEval+ 和 MBPP+ 等编码基准上显著超越使用教师原始思考内容的训练效果。
DeepSeek-V4.1-Flash-1017 is claimed to be able to build complete projects from scratch, enhancing end-to-end project development capabilities through reinforcement learning with self-feedback, covering tasks such as game development and 3D modeling.
Someone gave Opus 5.5 a single task: to build an ancient Greek world as detailed as possible. Everything visible in the scene is generated entirely by code, with no pre-made assets used.
A blog post detailing the development of clx, a compiler that translates Lua 5.5 source code into C++ for native compilation, replacing the Lua VM with a purpose-built runtime for compiled native Lua code.
The author introduces 'Anchor Engineering,' a software development method that encodes human design judgement into 'anchors'—guiding constraints stored in an 'Anchor Book' alongside the code—so AI coding agents can build systems faster without losing shared understanding, quality, or team coordination.
An AI agent called Computer generated Trace's launch video by building a React app from product screenshots, scripted with just six queries and 23 clicks, including spring-physics letterform animations rendered in Canvas with motion blur.
EpochGo is an open-source, local-first multi-agent AutoML engine that uses an offline Ollama LLM (qwen2.5-coder:7b) as an orchestrator over deterministic Scikit-Learn pipelines, with AST-based validation and auto-generated FastAPI serving endpoints, running 100% offline on consumer hardware.
中国研究者发表题为《The End of Harness Engineering》的论文，主张围绕模型编写的程序性 harness 应由自动化循环自动生成与评分迭代，人类只需负责编写评测集，而大模型本身保持不变。
Claude Code Template v1.0 for Spring Boot 发布，包含 6 个重写的专用 agent 和 2 个新 skill，提供生产级 Spring Boot 4.x / Java 21+ 应用开发模板，涵盖 REST API、JPA、安全、测试、云原生部署和 Kubernetes 最佳实践。
A community member reports that Qwen 3.8 Flash Next (UD-Q4_K_XL on 4x RTX 3090s) underperforms the Qwen 3.8 27B Q8 in a Tetris-building coding task, producing bugs and failing to deliver a playable .exe in Strata despite faster generation speeds, challenging the hype around the smaller model.
Reflection 发布开源模型 Beam-1.0-501B-A23B，采用稀疏 MoE 架构（501B 总参数/23B 激活参数），支持 1M token 上下文，主打 token 高效推理、软件工程与智能体工具调用能力，提供 BF16、FP8、NVFP4 多种数值精度版本，采用 Apache 2.0 许可。
作者为 Lombok 辩护，认为在 Java Records 时代它依然不可或缺，因为它在日志注解、构造器注入和 Builder 等方面大幅减少样板代码，而这些 Records 并不能替代。文章以 15 年实际使用经验和项目中 Lombok 注解使用统计为据，反驳了 Lombok 引入奇怪 bug 和"暗黑魔法"的常见批评。
用户让 M3.1 Flash Preview 用 Python 逐帧渲染行星影片，并通过查看自己生成的画面自动发现并修复了遮挡和亮度错误。整个过程仅用一次提示词，19 分钟内完成 3 次完整渲染和 17 次图像检查。
该论文发现 LLM 在生成代码时，中间层会涌现可复用、可组合的抽象表征，这些表征将功能、编程语言和实现策略分解为可独立操纵的成分，通过激活引导可灵活组合控制代码生成，且微调可跨编程语言迁移，验证并扩展了 Factored World Hypothesis 框架。
RSIFlow 提出一种递归自我改进框架，通过诊断任务执行反馈动态选择组件级改进工作流，并演化元智能体以优化编排能力，在六个基准上达到 SOTA 平均性能。
SWE-at-Scale 是一个通过反演真实世界软件修复来自动生成可执行 bug 任务的框架，用于大规模、多样化的编码智能体评估，实验表明其生成的任务比现有合成基线更具多样性和挑战性。
3DCodeVerse introduces a research ecosystem with a 2M-scale dataset of executable 3D code, a multi-agent DAG harness for automated generation and verification of 3D content, and empirical guidance showing fine-tuned open models achieve a 90% win rate against Claude Design (3D Object).
XML Generator 是一款 AI 代理，可从 JSON、CSV、YAML 或自然语言描述生成结构正确、符合词汇表规范的 XML，并支持 Maven POM、SOAP/WSDL、RSS/Atom 等常见格式，还能同步生成 XSD、DTD 或 XSLT。
Fable reviews the AI coding agent muse-wonder-full-1.0, noting it landed fifteen solid PRs in a day with clean mechanisms but made minor process slips like running unnecessary test batteries and leaving stale handoff notes. Compared to Codex (2-3 PRs) and Fable (6-8 PRs), Muse is far more productive in code output.