Latest
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。
The separation of search and reading tasks in KnowForge demonstrates a 1.9 F1 improvement over unified approaches, showing that specialized agents outperform generalist models in multimodal QA. Researchers and developers can access the system through the GitHub repository and three indexed benchmarks, making it a practical tool for document analysis tasks that require both information retrieval and comprehension.
GraphTok transforms tokenization from a bottleneck into a GPU-accelerated stage, dramatically increasing serving capacity for agentic workloads that resend long contexts. For developers managing LLM infrastructure, this means handling up to 518 requests/s of 121k-token traffic with a P99 target of 2 seconds, sustaining 1.75-2.5x the load of previous solutions.
arXiv 上这篇论文做安全强化学习,29 自由度人形机器人跑到 99.95% 安全率还保留 97% 回报,真机 Unitree G1 也验证了,做机器人方向可以看看。
Marformer直接预测条件边际分布,无需建模完整联合分布,在缺失数据处理任务上表现优异且速度快。
三家公司真实翻车案例都拆给你看了,还给出 PASAC 框架教你怎么在智能体运行中持续验证边界,做安全评估的必看。
METR那个时间跨度指标被重新算了一遍:2到30分钟区间难度曲线几乎是平的,同样是10倍跨度,难度差别很大,附诊断图。
一篇挺有意思的论文,教深度学习处理屈曲、相分离这类一对多解的分岔问题,单次推理就能出多个解分支,物理仿真方向的朋友可以看看。
想用 LLM 智能体做黑盒优化的可以看看,五个领域统一评测,还拆了哪些设计选择真有用、哪些是伪收益。
用模型内部激活训练探针抓欺骗,SHADE-Arena 上 AUC 98.8%,比直接读文本监控的 Opus 5.5 还准,数据集也开源了。
把一份文档直接编译成 LoRA 权重,284B 的 DeepSeek v4 Flash 也能用,top-1 准确率从 63.4% 拉到 84.9%
拿 Minecraft 重建曼哈顿、白金汉宫这些真实地点来考模型找路,GPT-6 Astra 考了 85.6 分,开源的 DeepSeek-V4.1-Flash 才 23.9,差距很直观。
QUILT addresses a fundamental inefficiency in sparse attention kernels by eliminating redundant memory operations through cross-query reuse. This approach significantly improves performance for long-context models without compromising accuracy, making it particularly valuable for developers working with large language models like GLM-5.3 and DeepSeek-3.2 that process extensive sequences.
StoreBench introduces a dynamic, production-grade environment that reveals significant gaps between current LLM capabilities and human performance in e-commerce operations. The benchmark's realistic challenges - including supplier failures and market shocks - provide a more rigorous test than static evaluations. Researchers and developers can access five training tasks and sample trajectories to improve agent performance in complex, time-sensitive commercial scenarios.
Only the latest 2000 items are listed; for older ones see the daily archive