PaperlayerView exact revision on arXiv

Source-linked AI summary

TelecomGPT-R1: A Unified Open-Source Reasoner for the Telecom Stack

Bohao Wang, Chenwei Wu, Haoyu Li, Hang Zou, Yu Tian, Lina Bariah, Li Wei, Chongwen Huang, Yongliang Shen, Zhaoyang Zhang, Merouane Debbah

arXiv:2608.26126v1cs.CLcs.IT

TL;DR

TelecomGPT-R1-9B targets the gap between generic reasoners without telecom grounding and telecom models with limited unified multi-step reasoning. It builds a four-axis, 67,427-example corpus and applies LoRA-SFT followed by DAPO-stabilized GRPO with binary verifier rewards. Across seven public benchmarks, it ranks first among open-source telecom LLMs and reaches a seven-axis mean comparable to closed-source frontier reasoners.

  • Problem

    Generic reasoners lack telecom-specific grounding, while telecom-domain LLMs remain limited in structured reasoning across heterogeneous standards, telemetry, formulas, and fault evidence.

  • Method

    The paper curates a 67,427-example corpus across protocol, knowledge, modeling, and fault axes, then post-trains Qwen3.5-9B with LoRA-SFT followed by DAPO-stabilized GRPO and axis-aligned binary verifier rewards.

  • Results

    TelecomGPT-R1-9B ranks first among open-source telecom LLMs across seven benchmarks, achieving an 82.1% seven-axis mean.

  • Takeaways & Limitations

    The model provides a unified open-source telecom reasoner that grounds general reasoning in telecom-specific evidence across heterogeneous tasks and data sources.

Abstract

from arXiv · show

Telecommunications is a high-leverage domain for large language model (LLM)-based reasoning because routine engineering workflows require joint grounding in normative specifications, operational telemetry, vendor-specific fault evidence, and exact RF/network calculations. However, current LLM integration in telecom remains bottlenecked by a two-sided capability gap: generic reasoners often lack telecom-specific grounding, while domain-specific telecom LLMs remain limited in structured, multi-step reasoning. To bridge this gap, we release TelecomGPT-R1-9B, a unified open-source telecom reasoner that ranks top-performing on the GSMA open telco leaderboard. Specifically, we curate a 67,427-example supervised fine-tuning (SFT) corpus organized around four complementary reasoning axes: protocol, knowledge, modeling, and fault. The corpus is built from axis-matched public web sources and enhanced through axis-specific chain-of-thought (CoT) generation and prefix-continuation self-validation. Starting from Qwen3.5-9B, we further develop a two-stage post-training recipe. First, multi-teacher low-rank adaptation (LoRA)-based SFT injects telecom knowledge and induces axis-specific reasoning formats. Second, group relative policy optimization (GRPO), stabilized by decoupled clip and dynamic sampling policy optimization (DAPO), optimizes the policy using four axis-aligned binary verifier rewards. Across seven public telecom benchmarks, TelecomGPT-R1-9B ranks first among open-source telecom LLMs and achieves a seven-axis mean comparable to state-of-the-art closed-source frontier reasoners.

I. INTRODUCTION

TelecomGPT-R1-9B addresses the gap between generic reasoners lacking telecom grounding and telecom models limited in unified multi-step reasoning. It combines a four-axis corpus with two-stage post-training and ranks first among open-source telecom LLMs.

  • Motivation: Telecom engineering requires joint reasoning across standards, protocols, logs, tables, formulas, code, and configuration evidence.Failure diagnosis may connect 3GPP procedures, O-RAN configurations, logs, KPI behavior, and standard constraints.
  • Capability gap: Generic reasoners lack reliable grounding in telecom standards, counters, protocol procedures, and operational evidence.Existing telecom models instead often target static QA, retrieval-assisted answering, or source-specific tasks.
  • Research question: The paper studies how to build one telecom reasoning model that generalizes across heterogeneous tasks and data sources.This is framed as a unified policy spanning heterogeneous evidence and source-specific correctness criteria.
  • Approach: The authors curate a 67,427-example corpus across protocol, knowledge, modeling, and fault reasoning, then apply multi-teacher LoRA-SFT and DAPO-stabilized GRPO.Training uses axis-specific CoT generation, self-validation, and axis-aligned binary verifier rewards.
  • Contributions: TelecomGPT-R1-9B ranks #1 among open-source telecom LLMs across seven public benchmarks.The model weights are publicly released, alongside practical findings about SFT, CoT source matching, dynamic sampling, verifier design, and teacher diversity.

A. Data Curation

The corpus is organized around four telecom reasoning axes and built from sources that preserve each axis’s native evidence structure. Questions and verified reasoning traces are then normalized, filtered, stratified, and tagged for training.

  • Reasoning axes: The corpus defines protocol, knowledge, modeling, and fault as complementary telecom reasoning axes.They target normative specifications, factual telecom knowledge, RF and mathematical computation, and root-cause analysis from RAN evidence, respectively.
  • Source collection: Protocol sources include 3GPP and O-RAN specifications and account for 50.7% of the corpus.Knowledge sources account for 21.5% and draw from standards, research papers, and operator documentation.
  • Question synthesis: Questions are synthesized from specification structure, document and working-group labels, and implementation symbols such as srsRAN abstract syntax trees and doxygen entries.This aligns question construction with clause-level, procedure-level, and hierarchy-aware evidence.
  • Pipeline: The two-frame recipe first curates 67,427 examples across four axes and then post-trains Qwen3.5-9B with LoRA-SFT and DAPO-stabilized GRPO.The second frame uses per-axis binary verifier rewards with asymmetric clipping, dynamic sampling, token-level loss, and a KL anchor.
  • CoT generation and verification: Source-matched CoT generation uses self-validation, Python re-execution with symbolic-equivalence checks, and deterministic rule replay.The generation method differs by axis: protocol and knowledge use teacher rationales, modeling uses executable checks, and fault analysis uses rule replay.
  • Normalization: All examples undergo normalization, multi-pass verification, augmentation, leakage filtering, difficulty stratification, and style mixing before receiving axis and source tags.The final representation uses a standard {system, user, assistant} schema.

B. Post-Training: SFT and DAPO

Post-training uses interleaved four-axis LoRA-SFT to install telecom knowledge and reasoning formats, followed by DAPO-style GRPO with axis-specific binary rewards. Dynamic sampling and stabilization mechanisms focus learning on informative rollouts while preserving structured outputs.

  • Stage 1: LoRA-SFT: Stage 1 uses LoRA-SFT to install telecom knowledge, answer discipline, and axis-specific CoT formats into one interleaved policy.The curated trace set spans all four reasoning axes.
  • Stage 2: DAPO-GRPO: Stage 2 samples rollout groups from the current policy and optimizes them with DAPO-style GRPO and KL regularization.The objective uses per-token importance ratios and a group-relative advantage.
  • Verifier rewards: Per-axis binary verifier rewards are selected according to each prompt’s protocol, knowledge, modeling, or fault label.For each rollout, the reward is produced by the verifier associated with the prompt’s axis.
  • Dynamic sampling: Dynamic sampling removes uniformly correct and uniformly incorrect groups before policy updates, retaining prompts with recoverable uncertainty.Binary rewards otherwise leave many rollout groups with little policy-gradient signal.
  • Stabilization: Asymmetric clipping, token-level loss aggregation, and an SFT-anchored KL regularizer stabilize training and preserve structured response formats.The clipping values are εℓ=0.20 and εh=0.28, while the KL coefficient is β=0.001.

III. EXPERIMENTAL EVALUATION

TelecomGPT-R1-9B is evaluated across seven GSMA telecom benchmarks against open-source and frontier baselines. It reaches the highest open-source seven-axis mean and remains within the closed-source frontier tier.

  • Overall evaluation: +22.8 pp over DeepSeek-V3 demonstrates the gap between TelecomGPT-R1-9B and the strongest open-source generalist baseline.DeepSeek-V3 reaches a 59.3% seven-axis mean with 685B parameters, whereas TelecomGPT-R1-9B uses 9B.
  • Baseline comparison: 71.9%, 73.3%, and 75.6% are the reported seven-axis means for GPT-5, Claude-Opus-4.6, and Gemini-3.1-Pro, respectively.TelecomGPT-R1-9B stays within the closed-source frontier tier on the aggregate.
  • Evaluation design: Table I compares per-axis accuracy across SFT, SFT+GRPO, and SFT+DAPO checkpoints from a shared 9B Qwen3.5 LoRA-SFT initialization.The released TelecomGPT-R1-9B checkpoint is the SFT+DAPO model.

A. Domain knowledge is the bottleneck

The results identify domain knowledge and verifier-aligned reasoning structure as prerequisites for effective reinforcement learning. On TeleLogs and 3GPP-TSG, DAPO provides the largest gains after SFT installs the relevant telecom rules.

  • Domain knowledge bottleneck: SFT installs domain facts and source conventions that verifier rewards require before reinforcement learning can produce a useful signal.Without the underlying telecom fact, fluent rollouts can be wrong and receive R = 0 across the group.
  • TeleLogs diagnosis: TeleLogs errors are dominated by missing per-axis RAN rules, whereas the verifier-aligned model uses rule replay to recover the correct diagnosis.In the illustrated case, the model computes m07 = −30.69 Mbps and m10 = 23◦, hits rule S5, and recovers C1.
  • Stage attribution: 75.2% is the SFT-only mean before DAPO adds 6.9 pp to reach 82.1%.The reported attribution assigns the knowledge load to SFT and later trajectory-selection and robustness improvements to RL.
  • RL gains: +33 pp lifts TeleLogs from 42.0% to 75.0%, while +8 pp lifts 3GPP-TSG from 71.0% to 79.0% under DAPO.These are the largest DAPO gains on the under-saturated axes.

B. Source-matched CoT generators matter

Source-matched chain-of-thought generators improve reasoning because telecom axes use different evidence structures. Answer-only and poorly structured CoT targets leave less useful structure for verifier-driven optimization.

  • Generator design: Source-matched CoT pairs each telecom axis with a verifier-aligned generator instead of one generic teacher prompt.This targets arithmetic mistakes, hallucinated table evidence, and skipped diagnostic thresholds before SFT.
  • Ablation: 72.0 / 59.0 is the BadCoT SFT-only score on TeleMath / TeleLogs, compared with 68.0 / 42.0 for multi-source SFT.BadCoT initially appears competitive despite using a less aligned reasoning structure.
  • Ablation: 65.2 / 72.9 is BadCoT’s DAPO score on TeleMath / TeleLogs, below the multi-source DAPO checkpoint’s 75.0 / 75.0.The paper attributes the difference to whether SFT leaves a verifier-aligned scaffold for DAPO to refine.

C. DAPO outperforms GRPO under heterogeneous source training

DAPO is most effective when training spans heterogeneous sources with different saturation speeds. Its dynamic sampling reallocates rollouts toward difficult, non-saturated axes rather than producing uniform gains.

  • Heterogeneous optimization: DAPO outperforms GRPO most strongly on TeleLogs and 3GPP-TSG, the axes that remain weak after SFT.Per-axis accuracy is tracked over 100 update steps; srsRANBench and TeleTables are omitted because they overlap with saturated axes.
  • Saturation effects: ∼1–2 pp separates DAPO and GRPO on MCQ-shaped axes near the SFT ceiling.The binary letter-match verifier provides near-degenerate gradient signal there, limiting sampling-policy gains.
  • Sampling strategy: Dynamic sampling reallocates rollouts to non-saturated TeleLogs and 3GPP-TSG prompts instead of uniformly lifting every axis.Static sampling would spend rollouts on easy prompts while undertraining difficult ones.

D. Teacher diversity boosts SFT quality

Teacher diversity improves the SFT corpus by avoiding narrow explanation styles and better preparing the policy for verifier-guided RL. Multi-source, multi-teacher initialization therefore outperforms single-teacher BadCoT after DAPO on TeleMath and TeleLogs.

  • Teacher diversity reduces overfitting to a narrow explanation style in the SFT corpus.
  • 75.0 / 75.0 on TeleMath / TeleLogs is achieved by multi-source SFT initialization with DAPO, versus 65.2 / 72.9 from BadCoT-init DAPO.The comparison holds corpus size and target axes fixed, attributing the gap to teacher diversity rather than additional rollouts or verifier feedback.

IV. CONCLUSION

TelecomGPT-R1-9B unifies telecom-specific evidence with general reasoning through structured data curation and verifier-coupled post-training. It ranks first among open-source telecom LLMs and remains within the closed-source frontier tier on the seven-axis mean.

  • TelecomGPT-R1-9B grounds general reasoning in telecom-specific evidence through an axis-indexed corpus and unified policy optimization.The corpus covers protocol, knowledge, modeling, and fault reasoning.
  • TelecomGPT-R1-9B ranks first among open-source telecom LLMs and remains within the closed-source frontier tier on the seven-axis mean.
  • Future work will scale the backbone, broaden the corpus across additional telecom evidence sources, and refine post-training for heterogeneous tasks.
Loading 2608.26126v1…