Tag
The paper introduces HybridAgents, an LLM agent architecture that dynamically switches between latent-space reasoning and text-based action via a lightweight HybridLink, combined with HybridSFT and HybridRL post-training, achieving 10.4% performance gains with large token, latency, and cost reductions across 11 benchmarks.
Stanford、Together AI、UIUC 与 Meta 联合提出 Hybrid Reasoning Agents，让模型主要通过隐藏的内部推理进行思考（更快、更省），仅在调用工具或给出最终答案时使用显式文本，在数学、科学、医疗等测试上提升约 10%，同时 token 减半、速度翻倍且成本更低。
When2Think is a post-training framework that dynamically allocates computation in large reasoning models based on problem difficulty, improving accuracy-efficiency trade-offs on mathematical benchmarks.
Ant Group releases Ling-3.0-flash, a hybrid-reasoning MoE model with 124B total parameters and only 5.1B active per token, matching the performance of its 1T flagship model on most benchmarks.
AntLing-3.0-flash is a hybrid-reasoning mixture-of-experts model designed for production-scale agent applications, as shown by benchmarks.
AntLingAGI releases Ling-3.0-flash, a 124B-parameter MoE model with 5.1B active parameters, now free for a week on Nous Portal. It matches or beats their 1T flagship on many benchmarks, designed for agent workloads like coding and tool use.
OpenBMB releases MiniCPM5-1B, a dense 1B Transformer model achieving SOTA among open-source 1B-class models, designed for on-device deployment with hybrid reasoning and long-context support.