TelcoAgent-Bench - Intent Recognition (Arabic): leaderboard

TelcoAgent-Bench (2026) tests small language models as telecom troubleshooting agents: from the first message of an engineer the agent recognizes the intent, calls network tools along a reference path among distractor tools and writes a resolution summary. Intent recognition without an intent list, as cosine similarity times 100 between the stated and gold intent in Arabic dialogues; higher is better.

Metric: Intent recognition accuracy: mean cosine similarity x100 (-100 to 100, sentence embeddings) between the agent's free-text intent and the gold intent label on TelcoAgent-Bench (15 telecom troubleshooting intents, 49 blueprints with 30 sampled dialogues each; the agent starts from the engineer's first message and calls core and distractor network tools), Arabic dialogues, without the intent list; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 8B74.2#1873
2Granite 3.3 8B Instruct72.1#3866
3Mistral 7B Instruct71.2#7568
4Gemma 3 4B (IT)70.9#4139
5Llama 3 8B Instruct70.7#4852
6granite-3.1-3B-a800m-instruct70.3#8600
7Qwen 2.5 7B Instruct70#2723
8Qwen-7B Chat67.7#7138

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=telcoagent-bench-intent-recognition-arabic · How It Works · Data refreshed daily, snapshot 2026-11-21.