1
How LLMs Learn to Be Helpful (RLHF vs DPO)
一文对比RLHF与DPO两种主流大模型训练方法的核心差异与适用场景
In this article, we will look at how that learning actually happens, starting with why instruction-following alone falls short, then walking through t…
2
内置“任意门”，高德发布通用世界模型工坊 ABot-WorldStudio
高德推出ABot-WorldStudio，内置“时空任意门”可穿越不同3D世界，单次连续运行超1小时，远超同类1分钟上限。
IT之家 7 月 14 日消息，今日，阿里巴巴集团旗下高德正式发布通用世界模型工坊 ABot-WorldStudio，并同步开放测试。该工坊将交互式视频生成与 3DGS 场景生成统一在同一产品中 —— 用户只需输入一段文字或一张图片，即可生成一个可实时交互、任意分享的 AI 世界，输出结果可保存为视…
3
高德发布通用世界模型工坊ABot-World Studio：5090单卡可生成小时级实时交互式视频与3D场景
高德最新世界模型工坊可单卡生成小时级实时交互视频与3D场景，首创“时空任意门”实现全尺度空间无缝穿梭。
4
高德发布通用世界模型工坊ABot-World Studio
高德发布ABot-World Studio，用通用世界模型工坊赋能智能体训练与模拟
36氪获悉，近日，阿里巴巴集团旗下高德正式发布通用世界模型工坊ABot-World Studio，并同步开放测试。该工坊将交互式视频生成与3DGS场景生成统一在同一产品中——用户只需输入一段文字或一张图片，即可生成一个可实时交互、任意分享的AI世界，输出结果可保存为视频与3DGS文件。
5
openclaw 2026.7.1-beta.6
开源AI工具openclaw发布beta版，新增Featherless、Claude Sonnet 5等模型，默认集成GPT-5.6。
2026.7.1 Highlights New models and providers: add Featherless, Claude Sonnet 5 and Mythos 5, Meta Muse Spark 1.1, and ClawRouter; make GPT-5.6 the new…
6
Do not migrate an AI API by changing only the base URL
仅改base URL迁移AI API？小心陷阱！深度解析兼容性迁移的真正要点
Changing the base URL is the easy part of an OpenAI-compatible migration. The real migration starts when the first request has to be explained. If a t…
7
Show HN: Parlan – a new programming language for fun
一个自称“好玩”的新编程语言Parlan开源在GitHub，适合好奇的程序员尝鲜。
Article URL: https://github.com/parlan-lang/parlan Comments URL: https://news.ycombinator.com/item?id=48871263 Points: 1 # Comments: 1
8
When Implausible Tokens Get Reinforced: Tail-Aware Credit Calibration for LLM Reinforcement Learning
解决LLM强化学习中不合理token被错误放大的问题，提出尾部感知信用校准新方法。
arXiv:2607.07976v1 Announce Type: cross Abstract: Reinforcement learning (RL) has achieved remarkable success in enhancing the reasoning capabilities …
9
Thinking Seeds: Leveraging Historical Diversity for Position-Aware RL in LLMs
提出Thinking Seeds方法，利用历史多样性与位置感知强化学习，显著提升LLM推理效率与泛化能力。
arXiv:2601.21476v2 Announce Type: replace Abstract: On-policy reinforcement learning (RL) for language model post-training suffers from a fundamental …
10
Miles: A PyTorch-Native Stack for Large-Scale LLM RL Post-Training
PyTorch原生打造的大规模LLM强化学习后训练框架，高效协调样本生成与策略更新。
Article URL: https://pytorch.org/blog/miles-a-pytorch-native-stack-for-large-scale-llm-rl-post-training/ Comments URL: https://news.ycombinator.com/it…
11
Vercel CEO Guillermo Rauch on the fight to split off models from agents
Vercel云平台让开发者无需管理基础设施即可快速部署前端、Serverless函数及AI代理，日均处理6百万次部署和万亿级token流量。
"The reality is, when you're optimizing for production, you start looking at a price/performance," Guillermo Rauch tells TechCrunch.
12
CARL: Constraint-Aware Reinforcement Learning for Planning with LLMs
引入约束感知强化学习，让LLM规划不再“天马行空”，ACL 2026最新研究。
arXiv:2607.04854v1 Announce Type: new Abstract: Despite their strong reasoning capabilities and extensive world knowledge, Large Language Models (LLMs…
13
Link Preview API
免费获取任意URL的标题、描述和OG图片，集成简单，快速实现链接预览功能。
Free API to get Open Graph data, title & images for any URL Discussion | Link
14
Show HN: TrainSim – a browser train tycoon
浏览器中的火车大亨模拟游戏，在线运营你的铁路帝国
Article URL: https://aashishh15.github.io/3DTrainSim/ Comments URL: https://news.ycombinator.com/item?id=48797128 Points: 2 # Comments: 1
15
DPO vs RLHF: The Alignment Tax You Pay Without Knowing
对比DPO与RLHF的对齐代价，揭示大模型隐藏的哲学回答偏差
Ask yourself one question. When you talk to ChatGPT or Claude, do you feel like you talk to something that thinks — or something that agrees with you …
16
Does a URL just sitting in a prompt steer an LLM's output toward its content?
探究URL直接放入提示能否让LLM输出偏向其内容，可能颠覆上下文嵌入方式。
Article URL: https://aifoc.us/influencing-model-output-with-urls/ Comments URL: https://news.ycombinator.com/item?id=48779535 Points: 2 # Comments: 0
17
Hyperloop Transformers
探索参数高效LLM架构新突破，为边缘部署提供解决方案。
arXiv:2604.21254v3 Announce Type: replace Abstract: LLM architecture research generally aims to maximize model quality subject to fixed compute/latenc…
18
Show HN: AI latent space with overlapping manifolds
探索AI潜在空间中的重叠流形结构，一个引人深思的开源项目。
Article URL: https://github.com/PJHkorea/Egregore/blob/main/integrated_egregore_core_test_v6_4.py Comments URL: https://news.ycombinator.com/item?id=4…
19
36氪首发 | 前大疆科学家创业，半年内连获四轮数亿融资，耀途资本、锦秋基金等押注
前大疆科学家张富创业，半年获四轮数亿融资，用新一代空间理解技术让飞行器「可操作」世界
作者丨欧雪 编辑丨袁斯来 硬氪获悉，聚焦通用空中智能的硅羽科技（SPARO）半年内已连续完成四轮数亿元融资，在种子轮耀途资本之后，锦秋基金、阿里巴巴、弘毅投资、普洛斯隐山资本、云时资本相继入局。资金用途方面，现阶段融资将主要⽤于扩充关键岗位的团队建设、推进产品线的商业化落地与规模化交付，以及加速技术…
20
百亿估值只是起点，跨维智能的物理AGI之路
国内物理AI标杆，具身世界模型突破空间感知难题，百亿估值只是起点。