TeleResilienceBench - ORANBench: leaderboard

TeleResilienceBench (2026) on its ORANBench subset (O-RAN knowledge): the model sees the question, the options and the first half of a flawed reasoning trace that led a weak generator to a wrong answer, and must continue it to the correct option. Score is the percent of the 63 instances recovered (correct flip rate); higher is better.

Metric: Correct flip rate (CFR, %): share of the 63 ORANBench instances (O-RAN knowledge, GSMA Open-Telco ot-lite) where the model, given the question, the options and the first half of a flawed reasoning trace from qwen3.5:2b that ended in a wrong answer, continues the reasoning in its thinking stream and returns the correct option; one continuation prompt, Ollama builds on one RTX 4090; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1nemotron-3-nano:4B (Ollama)39.7
2gemma4:e4B (Ollama)38.1
3gemma4:e2B (Ollama)36.5
4gemma4:26B (Ollama)33.3
5gemma4:31B (Ollama)31.7
6qwen3.5:4B (Ollama)28.6
7qwen3.5:27B (Ollama)23.8
8qwen3.5:9B (Ollama)23.8

Interactive version: theaggregate.ai/benchmark?slug=teleresiliencebench-oranbench · How It Works · Data refreshed daily, snapshot 2026-11-21.