arxiv arXiv cs.CL · 7h ago · src: 2d ago · research

TelecomGPT-R1: Unified Post-Training for Reasoning Across Heterogeneous Telecom Tasks

from English

Researchers introduce TelecomGPT-R1, a family of open-source unified telecom reasoning models designed to automate engineering tasks by reasoning over standards, configurations, and logs. The model addresses the limitations of existing general-purpose and specialized LLMs through a structured approach covering protocol, knowledge, modeling, and fault axes.

An axis-aware data generation framework refines public telecom artifacts into 104,880 verified question-answer pairs with chain-of-thought reasoning.
Supervised fine-tuning instills telecom knowledge to overcome the cold start barrier for reinforcement learning.
Dynamic sampling policy optimization (DAPO) with task-routed rubric rewards stabilizes RL updates across heterogeneous tasks.
The TelecomGPT-R1-27B model achieves an 89.64% mean score on seven GSMA Open Telco Leaderboard benchmarks, outperforming GPT-5, Claude, and Gemini.

The authors release the models and a reproducible training recipe to support community development in telecom automation.

Importance 2/3 arXiv cs.CL Zhipu AI Reasoning models Research paper

Related articles

arxiv arXiv cs.CL · 75d ago · 16 views

Probing Eternis-Forecaster 8B reveals internal representations improve calibration and expose CoT faithfulness issues

Researchers investigated whether internal representations in large language models provide a more direct window into calibration and faithfulness than chain-of-thought (CoT) reasoning. Working with Eternis-Forecaster 8B on OpenForesight, they trained representation-pooling probes on intermediate activations that achieved substantially better calibration compared to the model's standard outputs.

arxiv Moonshot AI (arXiv) · 595d ago · 26 views

GAIR-NLP releases LIMO, achieving strong math reasoning with 817 samples

GAIR-NLP introduces LIMO, a model demonstrating that complex mathematical reasoning can be effectively elicited using only 817 curated training samples. This approach challenges the conventional assumption that supervised fine-tuning requires massive datasets to avoid memorization.

arxiv arXiv cs.LG · 1d ago · 9 views

Fathom-Vaidya improves medical reasoning with rubric-based rewards

The authors introduce Fathom-Vaidya, a 30B parameter model that uses synthetic data and rubric-based reinforcement learning to enhance diagnostic and clinical healthcare reasoning. The training framework first applies rule-guided RL to MedBullets-derived questions for diagnosis, then utilizes 5.3k synthetic multi-turn scenarios with multi-dimensional rubrics for interactive clinical tasks.

media Hugging Face Forums · 2d ago · 15 views

Proposal for a proposition decomposition and contradiction-driven reasoning system

An author proposes a research design for a reasoning system that detects contradictions between two propositions by decomposing them into smaller components and searching for logical collisions, rather than relying on direct LLM judgment.

arxiv arXiv cs.CL · 13d ago · 28 views

Single-direction attack on GLM-5.3-Flash reveals safety alignment fragility in MoE models

Researchers demonstrate that directional ablation, a white-box attack removing refusal by projecting out a single direction from weights, remains effective on frontier mixture-of-experts (MoE) models like GLM-5.3-Flash. The study shows that while the attack survives the architecture, its effects are distributed across attention, dense, and routed-expert writers rather than concentrated in one location.