Instruction Following on IFBench (Prompt-level and Instruction-level Accuracy)
34.69Prompt-level AccuracyRTT-AON (Ours)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| RTT-AON (Ours)Backbone=Qwen3-4B-Instruct2026.04 | 34.69 | 37.61 | |
| RTT-CSR (Ours)Backbone=Qwen3-4B-Instruct2026.04 | 34.01 | 36.12 | |
| RL-AONBackbone=Qwen3-4B-Instruct2026.04 | 32.99 | 35.52 | |
| RL-CSRBackbone=Qwen3-4B-Instruct2026.04 | 32.31 | 33.43 | |
| RTT-CSR (Ours)Backbone=Qwen2.5-7B-Instruct2026.04 | 32.31 | 34.63 | |
| RTT-AON (Ours)Backbone=Qwen2.5-7B-Instruct2026.04 | 32.31 | 33.43 | |
| RL-AONBackbone=Qwen2.5-7B-Instruct2026.04 | 31.97 | 33.13 | |
| Qwen3-4B-Instruct†Backbone=Qwen3-4B-Instruct2026.04 | 30.95 | 32.83 | |
| DPOBackbone=Qwen3-4B-Instruct2026.04 | 29.93 | 32.84 | |
| SFTBackbone=Qwen3-4B-Instruct2026.04 | 29.59 | 31.64 | |
| SFTBackbone=Qwen2.5-7B-Instruct2026.04 | 28.23 | 30.15 | |
| RL-CSRBackbone=Qwen2.5-7B-Instruct2026.04 | 28.23 | 29.85 | |
| DPOBackbone=Qwen2.5-7B-Instruct2026.04 | 27.55 | 28.96 | |
| RTT-AON (Ours)Backbone=Llama3.2-3B-Instruct2026.04 | 26.19 | 27.46 | |
| Qwen2.5-7B-Instruct†Backbone=Qwen2.5-7B-Instruct2026.04 | 25.85 | 27.16 | |
| RTT-CSR (Ours)Backbone=Llama3.2-3B-Instruct2026.04 | 25.51 | 26.87 | |
| DPOBackbone=Llama3.2-3B-Instruct2026.04 | 24.15 | 26.27 | |
| RL-AONBackbone=Llama3.2-3B-Instruct2026.04 | 24.15 | 25.37 | |
| RL-CSRBackbone=Llama3.2-3B-Instruct2026.04 | 23.81 | 25.97 | |
| SFTBackbone=Llama3.2-3B-Instruct2026.04 | 22.79 | 24.78 | |
| Llama3.2-3B-Instruct⋆Backbone=Llama3.2-3B-Instruct2026.04 | 21.77 | 24.18 |