Latest
AssetForge用可重用资产库替代重复基础设施工作,在AutomationBench上准确率提升4.4倍,比传统方法高35.7%。
做剪枝的可以看看:CACTUS 比 LLM-Surgeon 快 28 倍,还能保留影响函数做数据选择。
给文本批量打标签太贵?SLKR 让旗舰模型离线写好标签规则,Qwen3.5-4B 一次调用就能分类。
免训练就能让 Qwen3.5-35B-A3B 看小时级长视频,三个基准最高涨近 13 个点,推理时直接插上就能用。
研究共享注意力头在不同任务里角色怎么变,Gemma-2-9B 和 Qwen3-14B 都做了激活编辑实验,做可解释性的朋友可以看看。
训练智能体只在终点给反馈、中间决策没法打分?这篇 MCGPO 直接复用已有轨迹做加权偏好,Qwen3-8B 在 WebShop 上提了 5 个点,还不额外加 critic。
清华团队推出ProgramGym框架,能构建1648个训练环境,让Qwen3.8-27B模型性能提升近8个百分点,甚至超过一些旗舰模型。
论文教你看 agent 失败原因来决定改运行时框架还是微调模型,Qwen3.5 上从 0.16 拉到 0.30,做 agent 的都该看看这套诊断方法。
一篇 MoE 训练论文:把固定 top-k 换成在 k 附近随机采样,Qwen3-30B-A3B 微调能多拿 2 分,推理成本不变。
给 TTS 加了个情绪旋钮,不用重新训练模型,注入转向向量就能调情绪强度,Qwen3-TTS 上效果最高提升 7 倍。
这篇论文把 RAG 从'检索'升级成'主动算证据',RouterLM 用 Qwen3.5-9B 就打赢了 Gemini 3.5 Flash,做 RAG 的可以看看思路。
Serving stacks have so far been split between LLM servers optimized for autoregressive decode and diffusion stacks built for denoising, forcing teams to glue disjoint runtimes together in production. vLLM-Omni replaces that with one control plane supporting session-shaped, long-lived workloads like robot loops and world models. Engineers building multimodal assistants on Qwen3-Omni or MiniCPM-o should evaluate it against their current inference setups.
不用 RL 训练,只在开头加个 "Okay" 就能让 Olmo-3-7B 数学成绩翻近一倍,这篇论文把原因挖得很透,做模型训练的值得看看。
给医学问答做了个按问题动态分配 LoRA 秩的微调方法,在 Qwen3-8B 上比 LoRA r16 高了 1 个多百分点,专科越多优势越大,做医疗方向微调的可以看看。
有人测了 TypeSafe 那个号称便宜快的 JEV,结果在 OpenAI 批量价格下根本不省钱,校准也没明显赢 Qwen3.8-27B,想做文本标注的可以看看。
GUI 智能体缺训练数据的老问题,这篇用图像编辑模型自己造数据,Qwen3-VL 微调后 AndroidWorld 成绩还涨了不少,代码开源可以直接试。
孟加拉语这种低资源语言上,Llama 4 Maverick 不微调就拿下 0.931 的 F1,比 BanglaBERT+XGBoost 还高,做低资源语言分类的可以看看。
用中子仪器设计测智能体会不会物理,7个模型最多做对7/16题,RL把Qwen3-8B从11%练到77%,比32B还强。
长链推理时 KV 缓存会爆炸,这篇提出用增量 SVD 压到 4-5 倍还能保住准确率,比删 token 的老办法稳,做推理部署的可以看看。
This research significantly reduces the computational cost of implementing MTP heads, making advanced inference optimization more accessible. The findings could enable faster deployment of large language models without requiring massive pretraining resources. Developers and researchers working on model optimization can implement these techniques to improve throughput with minimal computational overhead.
SCAPO算法通过半事实稳定性改进了令牌级信用分配,在数学和分布外基准上表现优异。
This research addresses a critical gap in VLM distillation where students can match teacher outputs without relying on the same visual evidence. The CW-OPD method provides a corrective signal that makes visual evidence reliance explicit rather than implicit. Developers working with smaller models can leverage this approach to achieve performance comparable to much larger models while maintaining better visual grounding.
ACSD provides a novel approach to token-level supervision throughout long responses by leveraging activation contrasts, offering a more stable alternative to existing methods. Researchers and developers working on reasoning models can implement this technique to improve mathematical accuracy without requiring reference solutions, potentially enhancing model performance across diverse training datasets.