Tag
Users report that OpenAI appears to have silently re-enabled the 'Improve the model for everyone' data-sharing setting in Codex, reactivating model training on their chats without explicit opt-in. Affected users are being pointed to ChatGPT data controls and OpenAI's privacy page to disable it again.
A tweet criticizes a viral claim about 'Weight-Space Programming' where GPT-6 Astra supposedly edits Qwen3.5-0.8B weights to boost math performance, noting the results lack competent baselines and the Qwen result is mere bug fixing, while acknowledging that models reading/writing weights in toy problems is a real trend worth watching.
A user accuses Anthropic of blocking accounts and downgrading Opus 5.5 to Opus 5 for any activity involving RFT, SFT, or training larger models, even without distilling Claude. The post criticizes Anthropic for punishing paying users doing research automation while having trained on public data.
论文研究了基准测试中问题难度和框架方式如何影响模型训练分数的观测提升，指出许多基准隐性地加权了特定难度区域，其排行榜结论需谨慎解读，部分基准甚至覆盖不足难度轴，难以有效监控模型改进过程。
Figma 详细介绍了其正在训练的「Figma-fluent」编辑模型，该模型专注于帮助设计师在现有画布内容上进行间距调整、颜色修改、布局重构等迭代式、难以自动验证的日常编辑操作。
一篇科普长文，借 OpenAI 分阶段关停自助微调平台（2027 年 1 月起全面停止新微调任务）这一新闻切入，通俗解释模型预训练、微调与 RAG 三者的区别，指出微调适合调整风格而非注入业务知识，而真正掌握企业数据的是 RAG。
RSI-Opt 通过元智能体在多次训练运行间累积并进化优化器策略（提示、流程、代码库与决策节奏），显著提升 LLM 优化器智能体在模型训练与科学机器学习任务上的性能，并支持跨模型策略迁移。
The MemTensor team releases MemT 1.0, pushing memory mechanisms from bolt-on RAG and Agent Memory into the core of model computation, allowing models to accumulate experience while solving tasks. Official tests show improvements of over 5% in Coding and Working scenarios, and the source code is now open-sourced on Hugging Face.
一项受控研究比较了视觉语言模型中的在线策略蒸馏（OPD）与过滤式监督微调（SFT），发现OPD仅在教师模型于学生生成轨迹上保持可靠时才有效，并且会继承教师的失败模式（如不完整回答），甚至可能抹去SFT阶段学到的完成行为。
这篇关于 Evo² 的文章提出将自进化 Agent 的两条路径（修改模型权重与修改 Harness）连接起来，让 Proposer 与 Reflector 双角色进入持续训练循环，在 Polyglot 上达到 82.1%、SWE-bench Verified 上达到 48.3%，且演化出的 Harness 可直接迁移复用。
冯峰（Fenng）就字节跳动前实习生恶意破坏模型训练事件发表评论，认为这种行为并非实习期间偶发，而是长期以来的行为模式，具有路径依赖性。
Igor Babuschkin (former DeepMind/OpenAI/xAI) announces River AI, a startup claiming the best system for building internal custom models, backed by a decade of top-tier model training experience and enthusiasm from Anjney Midha.
LlamaIndex 创始人 Jerry Liu 发文指出，前沿模型已能自动引导 ground-truth 标注，使得普通用户只需定义评测即可完成专业化模型的自动化训练闭环，并呼吁开放权重模型与训练流程应普惠所有人。
DatologyAI 发布 Curation Studio，面向训练自研模型的团队提供可定制的前沿数据筛选服务，并在 Modal 上利用数百块 B300 GPU 进行大规模合成数据测试验证。
Proximal 分享了其基于 Modal Clusters 构建的自研后训练基础设施，实现按需多节点集群与 GPU 自动扩缩容，以应对突发推理负载并打造高质量训练数据。
An OpenAI researcher commented that training AI models to perform AI R&D is significantly harder than training them to do math, tempering expectations about rapid AI-driven research acceleration toward ASI.
Overmind 是一个面向 AI 团队的开源模型训练平台，可将生产环境中的 traces 转化为团队自有的专用模型，涵盖训练、基准测试与部署全流程。
Overmind launched a platform that turns production agent traces into training data to fine-tune and serve small open-weight specialist models the customer owns, claiming its tuned 9B model beats frontier flags like GPT-5.6 Luna on biomedical, legal, and engineering extraction tasks (e.g. 91.6% vs 82.0% F1 on contract clause extraction), with platform code open-sourced under AGPL-3.0.
OpenAI chief research officer Mark Chen says the company will not be paralyzed by fallout from a series of agent containment breaches—including hacking Hugging Face and Australia's health-care system—while OpenAI pauses training of its latest models and reviews agent logs dating back to January 2026.
The paper introduces Neighborhood On-Policy Self-Distillation (N-OPSD), which exploits local parameter perturbations of a privileged teacher to pool complementary reference-aligned corrections and route them to student-visited states, improving math reasoning benchmarks (AIME 2024/2025, HMMT Feb 2025) across Qwen3 1.7B/4B/8B models.