TeleResilienceBench: Quantifying Resilience for LLM Reasoning in Telecommunications

2026-05-10

AI Summary

This work addresses the limited ability of current large language models to continue and correct erroneous reasoning chains in telecommunications scenarios—a capability not captured by existing evaluation frameworks. To quantify this “reasoning resilience,” the study introduces the concept of reasoning resilience along with a novel metric, Correct Flip Rate (CFR), and presents TeleResilienceBench, a benchmark constructed from seven telecom subdomains derived from the GSMA Open-Telco LLM suite. The benchmark truncates erroneous reasoning chains generated by weaker models at intermediate points and tasks target models with continuing and correcting them. Experimental results reveal that model scale exhibits no strong correlation with resilience, and prevailing evaluations emphasize knowledge coverage over reasoning depth. Among evaluated models, Nemotron-3-nano 4b achieves the best resilience-to-cost trade-off, attaining a macro-average CFR of 29.1% and leading performance on TeleMath.

Abstract

Deploying large language models in telecommunications requires more than task accuracy. In realistic workflows, a model may inherit partially completed reasoning from a prior step, an upstream agent, or its own earlier generation, and must continue that reasoning even when it is already going wrong. We introduce TeleResilienceBench, a benchmark that quantifies this capability, which we term reasoning resilience, across seven telecom sub-domains drawn from the GSMA Open-Telco LLM suite. Instances are constructed by collecting failures from a weak generator model, truncating the flawed reasoning trace at its midpoint, and asking a target model to continue and correct it. We propose the Correct Flip Rate (CFR) as a direct measure of successful recovery and evaluate eight models spanning the Qwen3.5, Gemma4, and Nemotron-3 families. Our results show that even the strongest model achieves a macro-average CFR of only 29.1%, and scale does not reliably improve resilience within families. Nemotron-3-nano 4b outperforms all Qwen3.5 variants including the 27b model and leads the auxiliary TeleMath numerical evaluation at 23.4% CR%, offering the best resilience-to-cost ratio in the set. A difficulty-stratified analysis further reveals that existing telecom benchmark difficulty labels reflect factual specificity rather than reasoning depth, suggesting that current evaluations measure knowledge coverage more than reasoning ability.

Problem

• reasoning resilience
• telecommunications
• large language models
• benchmark
• error recovery

Innovation

• reasoning resilience
• TeleResilienceBench
• Correct Flip Rate
• error recovery
• telecom LLM evaluation