Wikibench

Multi-IF

reasoning

Multi-IF benchmarks LLMs on multi-turn and multilingual instruction following. It expands upon IFEval by incorporating multi-turn sequences and translating English prompts into 7 other languages, resulting in 4,501 multilingual conversations with three turns each. The benchmark reveals that current leading LLMs struggle with maintaining accuracy in multi-turn instructions and shows higher error rates for non-Latin script languages.

Leaderboard

Showing 20 of 20 results

  1. Qwen3-235B-A22B-Thinking-2507
    80.6%
    i
  2. o3-mini
    79.5%
    i
  3. Qwen3 VL 235B A22B Thinking
    79.1%
    i
  4. Qwen3 VL 32B Thinking
    78.0%
    i
  5. Qwen3-Next-80B-A3B-Thinking
    77.8%
    i
  6. Qwen3-235B-A22B-Instruct-2507
    77.5%
    i
  7. Qwen3 VL 235B A22B Instruct
    76.3%
    i
  8. Qwen3-Next-80B-A3B-Instruct
    75.8%
    i
  9. Qwen3 VL 8B Instruct
    75.1%
    i
  10. Qwen3 VL 8B Thinking
    75.1%
    i
  11. Qwen3 VL 4B Thinking
    73.6%
    i
  12. Qwen3 VL 30B A3B Thinking
    73.0%
    i
  13. Qwen3 30B A3B
    72.2%
    i
  14. Qwen3 VL 32B Instruct
    72.0%
    i
  15. GPT-4.1
    70.8%
    i
  16. GPT-4.5
    70.8%
    i
  17. GPT-4.1 mini
    67.0%
    i
  18. Qwen3 VL 30B A3B Instruct
    66.1%
    i
  19. GPT-4o
    60.9%
    i
  20. GPT-4.1 nano
    57.2%
    i