揭秘大模型rollout强化学习中的子空间几何，提出GCPO约束方法，助你理解训练背后的数学本质。
arXiv:2608.11674v1 Announce Type: cross Abstract: On-policy rollout methods such as GRPO are central to post-training of large language models, yet th…
用MoE代理模型低成本复现LLM强化学习后训练故障，大幅降低排查成本，直击调试痛点。
arXiv:2608.10823v1 Announce Type: new Abstract: Reinforcement learning (RL) post-training of large language models (LLMs) is computationally intensive…
用离线Top-K logits和融合分块KL损失，大幅降低大模型蒸馏成本，兼顾效率与质量。
arXiv:2608.03796v1 Announce Type: cross Abstract: Small language models are often the only option for deployment under tight latency, cost, and on-pre…
把 Transformer 参数与激活约束到超球面，nGPT 带来表示学习新范式，一文看懂归一化训练的关键设计。
arXiv:2608.01284v1 Announce Type: cross Abstract: The normalized Transformer (nGPT) realizes hyperspherical representation learning by constraining mo…
揭秘大模型性能增长的底层规律，一文读懂 Scaling Laws 如何指导 LLM 训练与架构设计。
Article URL: https://aidoses.substack.com/p/scaling-laws-the-law-behind-every Comments URL: https://news.ycombinator.com/item?id=49111433 Points: 2 # …
无需训练的dLLMs推理加速方法，用草稿引导连续跳跃解码，提升效率且保持质量
arXiv:2607.20467v1 Announce Type: new Abstract: While parallel decoding is central to the efficiency of Diffusion Large Language Models (dLLMs), curre…
低秩预训练大模型遭遇不稳定困境？这项ICML 2026研究提出原生低秩LLM预训练的稳定化方法，兼顾效率与质量。
arXiv:2602.12429v2 Announce Type: replace Abstract: Foundation models have achieved remarkable success, yet their growing parameter counts pose signif…
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 …
用推测解码的思路让训练加速：小模型打草稿、大模型只做校验，改下训练流程就能省下大量算力。
A new paper introduces a method to speed up reward-based fine-tuning by having the model generate a cheap, compressed copy of itself to draft text, wh…
用严谨实验揭示扩散模型时间步编码冗余，为你省下多余参数、提速训练
arXiv:2606.20416v1 Announce Type: new Abstract: Diffusion models rely heavily on explicit timestep embeddings to modulate the denoising process across…
未开微调先知性能？新方法TuneAhead实现提前预测，节省训练成本。
arXiv:2606.17660v1 Announce Type: cross Abstract: Fine-tuning large language models (LLMs) is compute-intensive and error-prone: model performance dep…
无需额外训练，通过未来价值引导解码就能让基础大模型找到正确推理路径，这篇论文提出了一种高效的采样方法。
arXiv:2605.02427v3 Announce Type: replace Abstract: A recurring pattern in "reasoning without training" is that base LLMs already assign non-trivial p…
利用Audio-LLMs智能过滤低质量数据，大幅提升语音到语音翻译模型的训练效率与效果，INTERSPEECH 2026录用论文。
arXiv:2606.13507v1 Announce Type: new Abstract: Large-scale mined corpora provide abundant training data for end-to-end speech-to-speech translation (…
用奖励校准解决LLM路由难题，RL方法让大模型调度更高效精准。
arXiv:2606.12479v1 Announce Type: cross Abstract: Large language model (LLM) routing has emerged as an effective paradigm for leveraging the complemen…
强化学习训练智能体搜索时，巧妙回收零方差查询，效率提升有新思路。
arXiv:2606.10709v1 Announce Type: cross Abstract: The use of GRPO-style algorithms has become the standard strategy for training LLM search agents und…
提出在线验证提示选择方法，大幅提升大推理模型RL训练效率
arXiv:2603.25184v2 Announce Type: replace Abstract: Reinforcement learning (RL) has become essential for post-training large language models (LLMs) in…
提出通过分区拟阵约束梯度匹配优化语言模型小批量选择，提升训练效率与收敛质量。
arXiv:2606.07954v1 Announce Type: new Abstract: Training large language models (LLMs) on heterogeneous data requires selecting minibatches that balanc…
LLM继续预训练中，超参数配置可预测的缩放规律，告别启发式搜索与高昂成本
arXiv:2606.05610v1 Announce Type: new Abstract: The efficacy of continued pre-training for Large Language Models (LLMs) hinges upon hyperparameter con…
重新审视张量分解在LLM压缩中的核心作用，为后训练压缩提供新思路。
arXiv:2606.03465v1 Announce Type: cross Abstract: Post-training compression is essential for deploying large language models (LLMs) under tight resour…
改进激活预言机训练，解决幻觉和模糊性，可提升残差流激活的可解释性研究价值。
arXiv:2606.02609v1 Announce Type: cross Abstract: Activation Oracles (AOs) are promising methods for interpreting residual stream activations. However…