Disentangling Language Roles in Multilingual L: leaderboard
Multilingual LLMs are increasingly used when instruction, source content, and required response languages do not coincide. Existing benchmarks have expanded multilingual instruction-following evaluation, but they rarely isolate these three roles within a fully crossed design. We introduce MTM-Bench, a controlled benchmark for language-conditioned task execution in which each instance is defined by
Metric: JointSuccess (self-reported). Source: benchmarklist.com. Saturation forecast: Around June 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 Max | 84 | #208 |
| 2 | DeepSeek R1 | 82.2 | #251 |
| 3 | Qwen 3.5 Plus | 82 | #125 |
| 4 | GPT-5.4 Mini | 81.5 | #212 |
| 5 | GPT-5.4 | 79.9 | #79 |
| 6 | DeepSeek V3.2 | 79.6 | #205 |
| 7 | DeepSeek V3.1 | 78.6 | #264 |
| 8 | GPT-5.3 Codex | 78.1 | #69 |
| 9 | Claude Sonnet 4.6 | 76.7 | #90 |
| 10 | Gemini 2.5 Flash Lite | 76.7 | #422 |
| 11 | Qwen 3.5 Flash | 76.5 | #232 |
| 12 | GPT-4.1 | 75.8 | #248 |
| 13 | Claude Opus 4.1 | 74.1 | #140 |
| 14 | Llama 3.3 70B | 74 | #579 |
| 15 | Gemini 2.5 Flash | 72.7 | #245 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=disentangling-language-roles-in-multilingual-l · How It Works · Data refreshed daily, snapshot 2026-10-20.