Researchers release TelecomGPT-R1, a family of open-source telecom LLMs post-trained with SFT and DAPO reinforcement learning across four task axes (protocol, knowledge, modeling, fault). The 27B model scores 89.64% mean on seven GSMA Open Telco Leaderboard benchmarks, outperforming proprietary models including GPT-5, Claude, and Gemini. Models and training recipe are released openly.
An arXiv paper presents a structured reasoning framework applying LLMs to root cause analysis in telecom network operations. It targets 5G and emerging 6G networks, where complex cross-layer dependencies make diagnosing performance degradations difficult. The paper argues vanilla LLMs hallucinate and reason unstably on structured network evidence, motivating an evidence-grounded approach.
An arXiv paper presents a structured reasoning framework applying LLMs to root cause analysis in telecom network operations. It targets 5G and emerging 6G networks, where complex cross-layer dependencies make diagnosing performance degradations difficult. The paper argues vanilla LLMs hallucinate and reason unstably on structured network evidence, motivating an evidence-grounded approach.
An arXiv paper introduces Rufus-Air, an open and reproducible post-training recipe built on GLM-4.5-Air-Base (106B-A12B). It details an eight-stage pipeline spanning SFT, multiple RL stages, agent training, and RLHF. The recipe improves over the official GLM-4.5-Air release and is competitive with similarly sized open models, making it a practical reference for practitioners reproducing full post-training pipelines.
An arXiv paper introduces Rufus-Air, an open and reproducible post-training recipe built on GLM-4.5-Air-Base (106B-A12B). It details an eight-stage pipeline spanning SFT, multiple RL stages, agent training, and RLHF. The recipe improves over the official GLM-4.5-Air release and is competitive with similarly sized open models, making it a practical reference for practitioners reproducing full post-training pipelines.
A new tutorial-and-survey paper formalizes how LLM-based agentic AI can autonomously control 5G and 6G networks, covering reasoning, planning, tool use, multi-agent coordination, and standardization alignment. The work addresses the gap between AI and network protocol domains.
A new tutorial-and-survey paper formalizes how LLM-based agentic AI can autonomously control 5G and 6G networks, covering reasoning, planning, tool use, multi-agent coordination, and standardization alignment. The work addresses the gap between AI and network protocol domains.
An arXiv paper introduces PAC (Progress-Augmented Advantage Curriculum), a method for multi-task reinforcement learning of LLMs. It dynamically allocates rollout budget across heterogeneous tasks instead of relying on fixed task mixtures, aiming to improve reasoning post-training efficiency.
An arXiv paper introduces PAC (Progress-Augmented Advantage Curriculum), a method for multi-task reinforcement learning of LLMs. It dynamically allocates rollout budget across heterogeneous tasks instead of relying on fixed task mixtures, aiming to improve reasoning post-training efficiency.
A new report introduces SignalReasoner, a Qwen2.5-3B-Base model post-trained with reinforcement learning from verifiable rewards (RLVR) to tackle graduate-level signal processing math problems. It uses the WirelessMATHBench-XL benchmark and compares direct RL with other paradigms, probing the upper bound of 3B-scale models in this niche domain.
A new report introduces SignalReasoner, a Qwen2.5-3B-Base model post-trained with reinforcement learning from verifiable rewards (RLVR) to tackle graduate-level signal processing math problems. It uses the WirelessMATHBench-XL benchmark and compares direct RL with other paradigms, probing the upper bound of 3B-scale models in this niche domain.