1
Train LLM from Scratch
从零开始训练大模型，详解预训练、SFT与PPO全流程。
Article URL: https://FareedKhan-dev.github.io/train-llm-from-scratch/ Comments URL: https://news.ycombinator.com/item?id=48615416 Points: 2 # Comments…
2
SFT Overtraining Predicts Rank Inversion via Entropy Collapse Under RLVR
ICML 2026 研究揭示：监督微调过度训练会通过熵塌陷导致强化学习下模型输出排名反转，对齐策略需要警惕。
arXiv:2606.18487v1 Announce Type: new Abstract: The standard heuristic of selecting the SFT checkpoint with the highest pass@1 for GRPO can fail when …
3
Show HN: Conduit, a local-first SSH, Mosh and SFTP client for Android
Android上的本地优先SSH/Mosh/SFTP客户端，开源、离线可用，远程管理利器。
Article URL: https://github.com/gwitko/Conduit Comments URL: https://news.ycombinator.com/item?id=48487447 Points: 1 # Comments: 0
4
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 …
5
Beyond Two-Stage Training: Cooperative SFT and RL for LLM Reasoning
打破两阶段训练范式，提出SFT与RL协同强化大模型推理能力的新方法。
arXiv:2509.06948v3 Announce Type: replace Abstract: Supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) are two wid…
6
Good SFT Optimizes for SFT, Better SFT Prepares for Reinforcement Learning
探讨SFT如何优化自身并为强化学习做准备，揭示大模型训练策略的关键演进方向
arXiv:2602.01058v2 Announce Type: replace-cross Abstract: Post-training of reasoning LLMs is a holistic process that typically consists of an offline …
7
Fixing LLM Writing with Distribution Fine Tuning
用分布微调技术让LLM写作告别公式化，创造力提升164%，效果显著
Article URL: https://rosmine.ai/2026/05/18/fixing-llm-writing-with-distribution-fine-tuning/ Comments URL: https://news.ycombinator.com/item?id=481847…
8
Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling
结合有监督与强化微调的创新方法，通过前缀采样平衡模仿学习与探索，提升LLM后训练效果。
arXiv:2507.01679v3 Announce Type: replace-cross Abstract: Existing LLMs-post-training techniques are broadly categorized into supervised fine-tuning (…