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
Satya Nadella has issued a shocking warning to companies using AI
微软CEO警告企业AI使用风险：你的核心数据可能正被模型制造商学习利用。
In a surprising blog post on Monday, Microsoft CEO is warning enterprises of the dangers of using proprietary models like Anthropic's and OpenAI's.
3
OpenAI’s Head of Safety Is Leaving the Company
OpenAI安全负责人离职，内部安全挑战随模型训练加速而加剧，引发行业关注
Johannes Heidecke’s departure comes as OpenAI tries to further integrate its research and safety teams.
4
GIFT: Geometry-Informed Low-precision Gradient Communication for LLM Pretraining
提出GIFT方法，利用梯度几何信息实现低精度通信，在不牺牲模型精度的前提下显著降低LLM预训练通信开销。
arXiv:2607.07494v1 Announce Type: cross Abstract: Gradient communication is a primary scaling bottleneck in large language model (LLM) pretraining. Co…
5
Not All Refusals Are Equal: How Safety Alignment Fails Cybersecurity at Scale
安全对齐虽关键，但一刀切的拒绝机制在网络安全等高风险领域适得其反，这篇论文揭示了系统性缺陷。
arXiv:2607.02714v1 Announce Type: cross Abstract: There is no doubt that safety alignment is an essential step in LLM training. However, conceptually …
6
Jackrong LLM Fine-Tuning Guide
LLM微调实战指南，从入门到进阶的完整开源教程，适合AI开发者快速上手
Article URL: https://github.com/R6410418/Jackrong-llm-finetuning-guide Comments URL: https://news.ycombinator.com/item?id=48812171 Points: 2 # Comment…
7
Neuron-Aware Data Selection for Annotation-Free LLM Self-Distillation
无需人工标注，通过神经元激活模式筛选数据，实现LLM高效自蒸馏训练。
arXiv:2607.02460v1 Announce Type: cross Abstract: Post-training large language models (LLMs) without real-world interaction feedback or human-labeled …
8
SCAPE: Accurate and Efficient LLM Training with Extreme Sparse Communication
新论文提出SCAPE方法，通过极端稀疏通信显著降低大模型预训练通信开销，兼顾准确与效率。
arXiv:2607.01678v1 Announce Type: new Abstract: Communication increasingly dominates the cost of Large Language Model (LLM) pre-training, especially u…
9
Show HN: I trained a 1B LLM from scratch for $315 and open-sourced weights+data
仅花315美元从零训练10亿参数大语言模型，并开源全部权重与数据，堪称极低成本LLM训练范本。
Article URL: https://huggingface.co/AIIT-Threshold/Tessera-1B Comments URL: https://news.ycombinator.com/item?id=48758380 Points: 2 # Comments: 0
10
CausalMix: Data Mixture as Causal Inference for Language Model Training
用因果推断视角优化数据混合，提升语言模型训练效率与效果的创新方法。
arXiv:2607.01104v1 Announce Type: cross Abstract: In Large Language Model (LLM) training, data mixing plays a pivotal role in determining model perfor…
11
Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care
当LLM学会在癫痫护理中“推荐”与“延迟”，医疗AI的公平性难题有了新解法
arXiv:2606.31036v1 Announce Type: new Abstract: Specialist epilepsy expertise is scarce in resource-constrained settings, making LLM-based decision su…
12
From Search to Synthesis: Training LLMs as Zero-Shot Workflow Generators
零样本工作流生成新范式：训练LLM从搜索到合成，无需示例即可自动编排流程
arXiv:2606.30704v1 Announce Type: cross Abstract: Large language models (LLMs) excel across a wide range of tasks, yet their instance-specific solutio…
13
Building Multi-Task Agentic LLMs via Two-Phase Distillation
两阶段蒸馏法让LLM成为多任务智能代理，训练效率与推理能力双提升
arXiv:2606.30044v1 Announce Type: new Abstract: A key step toward artificial general intelligence is to train models that can perform multiple tasks. …
14
Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents
LLM Agent 如何应对交互中的事实变化？论文诊断并训练记忆更新缺口，确保 Agent 使用当前有效信息而非过期值。
arXiv:2606.27472v1 Announce Type: cross Abstract: Large language model (LLM) agents operate over long, multi-session interactions in which facts chang…
15
The Interplay of Harness Design and Post-Training in LLM Agents
揭秘工具架构设计如何影响LLM Agent后训练效果，为智能体优化提供新视角
arXiv:2606.25447v1 Announce Type: new Abstract: Tool-integrated LLM agents are often wrapped within a harness: the scaffolding that determines which t…
16
Beyond Safe Data: Pretraining-Stage Alignment with Regular Safety Reflection
提出预训练阶段对齐新方法，用“安全反射”机制超越单纯安全数据，提升大模型本质安全性。
arXiv:2606.19168v1 Announce Type: cross Abstract: To achieve deeper safety alignment for large language models (LLMs), recent efforts have studied how…
17
Small Initialization Matters for Large Language Models
这项研究揭示了参数初始化对大型语言模型训练效果的关键影响，为模型优化提供了新视角。
arXiv:2606.17945v1 Announce Type: new Abstract: Large language models provide a tractable system for asking how intelligence itself emerges, rather th…
18
The Curse and Blessing of Mean Bias in FP4-Quantized LLM Training
FP4训练中均值偏差的双重效应，揭示极端激活值如何破坏长尾信号，提出关键改进方向
arXiv:2603.10444v2 Announce Type: replace-cross Abstract: FP4 training promises substantial memory and compute savings for large language models, but …
19
遭社区强烈反对后，Anthropic 调整 Claude Fable 5 面向研究员的“降智”措施
社区强烈反对迫使Anthropic调整Claude Fable 5的“降智”措施，AI模型安全与开放性的博弈再起波澜。
IT之家 6 月 11 日消息，此前，Anthropic 被曝会在用户不知情的情况下， 限制竞争对手使用新模型 Claude Fable 5 开发其他 AI 模型 。AI 研究社区强烈反对后，Anthropic 决定做出点改变。 Anthropic 在给《连线》的声明中致歉称：“我们正在调整 Fab…
20
When RL Fails after SFT: Rejuvenating Model Plasticity for Robust SFT-to-RL Handoff
揭秘SFT后强化学习失效的成因，提出恢复模型可塑性的新方法。
arXiv:2606.09932v1 Announce Type: cross Abstract: Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has become a standard pipeline …