SLM Trustworthiness - Robustness: leaderboard
SLM Trustworthiness (Benchmarking Trustworthiness of SLMs, arXiv 2608.11981v1, Tables I, II and IV): TrustLLM evaluation of uncompressed instruction-tuned models from 360M to 8B parameters, the baselines of a pruning, quantization and distillation study. Robustness: mean score on AdvGLUE and AdvInstruction natural-noise tasks plus out-of-distribution detection and generalization, percent. Higher is better.
Metric: TrustLLM robustness accuracy (%; sub-task mean). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.1 8B Instruct | 71.62 | #1017 |
| 2 | Qwen 2.5 1.5B Instruct | 69.38 | #1393 |
| 3 | Gemma 1.1 7B (IT) | 68.9 | #1348 |
| 4 | Qwen 2.5 3B Instruct | 68.79 | #1159 |
| 5 | Qwen 2.5 7B Instruct | 67.05 | #863 |
| 6 | Llama 3.2 3B Instruct | 66.42 | #1327 |
| 7 | Qwen 2.5 0.5B Instruct | 62.42 | #1510 |
| 8 | SmolLM2-360M-Instruct | 60.16 | #1594 |
| 9 | Llama 3.2 1B Instruct | 59.43 | #1518 |
| 10 | h2o-danube3-500m-chat | 56.85 | #1574 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=slm-trustworthiness-robustness · How It Works · Data refreshed daily, snapshot 2026-11-05.