Instruction Following on IFBench
82.9Pass@1 (Strict)Nemotron-Cascade-2 30B-A3B
Evaluation results
| Method | Links | |
|---|---|---|
| Nemotron-Cascade-2 30B-A3B2026.03 | 82.9 | |
| Nemotron-3-Super 120B-A12B2026.03 | 72.6 | |
| Nemotron-3-Nano 30B-A3B2026.03 | 71.5 | |
| Qwen3.5 35B-A3B2026.03 | 70.2 | |
| Olmo 3.1 Think 32BTraining Stage=Final Think 3.1, Model Family=Olmo 3.1, Parameter Count=32B, Thinking Capability=true2025.12 | 68.1 | |
| Qwen 3 VL 32B ThinkModel Family=Qwen 3, Parameter Count=32B, Thinking Capability=true2025.12 | 55.1 | |
| Olmo 3 Think (Final 3.0)Training Stage=Final Think 3.0, Model Family=Olmo 3, Parameter Count=32B, Thinking Capability=true2025.12 | 47.6 | |
| DeepSeek-V3.2Mode=Non-Reasoning, Architecture=MoE, Total Params=671B, Activated Params=37B2026.01 | 47 | |
| K2-V2 70B InstructModel Family=K2, Parameter Count=70B, Thinking Capability=false2025.12 | 46.3 | |
| A.X K1Parameters=519B-A33B, Non-Thinking Mode=true2026.01 | 44.3 | |
| Qwen3-235B-A22B Instruct-2507Architecture=MoE, Total Params=235B, Activated Params=22B2026.01 | 43.2 | |
| K-EXAONEMode=Non-Reasoning, Architecture=MoE, Total Params=236B, Activated Params=23B2026.01 | 40.5 | |
| DeepSeek-V3.1Parameters=685B-A37B, Non-Thinking Mode=true2026.01 | 37.8 | |
| Qwen 3 32BModel Family=Qwen 3, Parameter Count=32B, Thinking Capability=false2025.12 | 37.3 | |
| Olmo 3 Think (SFT)Training Stage=SFT, Model Family=Olmo 3, Parameter Count=32B, Thinking Capability=true2025.12 | 37 | |
| GLM-4.6Parameters=357B-A32B, Non-Thinking Mode=true2026.01 | 36.7 | |
| EXAONE 4.0Mode=Non-Reasoning, Architecture=Dense, Total Params=32B, Activated Params=32B2026.01 | 34.8 | |
| Olmo 3 Think (DPO)Training Stage=DPO, Model Family=Olmo 3, Parameter Count=32B, Thinking Capability=true2025.12 | 34.4 | |
| Qwen 3 VL 8B Inststage=Instruct2025.12 | 34 | |
| Olmo 3 7B Instructstage=Final Instruct2025.12 | 32.3 | |
| Olmo 3 7B Instructstage=DPO2025.12 | 29.3 | |
| Qwen 3 8Bstage=Instruct2025.12 | 29.3 | |
| Qwen 2.5 7Bstage=Instruct2025.12 | 28.4 | |
| Olmo 3 7B Instructstage=SFT2025.12 | 27.4 | |
| OLMo 2 7B Inststage=Instruct2025.12 | 26.7 | |
| DS-R1 32BModel Family=DeepSeek-R1, Parameter Count=32B, Thinking Capability=true2025.12 | 23.8 | |
| Granite 3.3 8B Inststage=Instruct2025.12 | 22.3 | |
| Apertus 8B Inststage=Instruct2025.12 | 22.1 | |
| LLM ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-based mitigation (S1)2026.02 | 20.8 | |
| LLM ROUTERMalicious Setting=M4-RL, Mitigation Strategy=supervisor-free mitigation2026.02 | 20.8 | |
| TEXT DEBATEMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-based mitigation (S1)2026.02 | 20.8 | |
| LLM ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-free mitigation2026.02 | 20.4 | |
| LLM ROUTERMalicious Setting=M4-RL, Mitigation Strategy=supervisor-based mitigation (S1)2026.02 | 20.4 | |
| GRAPH ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-based mitigation (S2)2026.02 | 20.4 | |
| GRAPH ROUTERMalicious Setting=M4-RL, Mitigation Strategy=supervisor-based mitigation (S2)2026.02 | 20.4 | |
| TEXT DEBATEMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-based mitigation (S2)2026.02 | 20.4 | |
| TEXT DEBATEMalicious Setting=M2-Activation Steering, Mitigation Strategy=None (Malicious)2026.02 | 20 | |
| GRAPH ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=None (Malicious)2026.02 | 19.6 | |
| GRAPH ROUTERMalicious Setting=M4-RL, Mitigation Strategy=None (Malicious)2026.02 | 19.6 | |
| GRAPH ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=None (Benign)2026.02 | 18.8 | |
| TEXT DEBATEMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-free mitigation2026.02 | 18.8 | |
| TEXT DEBATEMalicious Setting=M4-RL, Mitigation Strategy=supervisor-based mitigation (S2)2026.02 | 18.8 | |
| TEXT FEEDBACKMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-based mitigation (S2)2026.02 | 18.8 | |
| LLM ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-based mitigation (S2)2026.02 | 18.4 | |
| GRAPH ROUTERMalicious Setting=M4-RL, Mitigation Strategy=supervisor-free mitigation2026.02 | 18.4 | |
| TEXT DEBATEMalicious Setting=M4-RL, Mitigation Strategy=supervisor-based mitigation (S1)2026.02 | 18.4 | |
| LLM ROUTERMalicious Setting=M4-RL, Mitigation Strategy=supervisor-based mitigation (S2)2026.02 | 18 | |
| GRAPH ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-based mitigation (S1)2026.02 | 18 | |
| TEXT DEBATEMalicious Setting=M4-RL, Mitigation Strategy=supervisor-free mitigation2026.02 | 18 | |
| TEXT FEEDBACKMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-free mitigation2026.02 | 18 | |
| TEXT FEEDBACKMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-based mitigation (S1)2026.02 | 18 | |
| LLM ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=None (Benign)2026.02 | 17.6 | |
| LLM ROUTERMalicious Setting=M4-RL, Mitigation Strategy=None (Malicious)2026.02 | 17.6 | |
| GRAPH ROUTERMalicious Setting=M4-RL, Mitigation Strategy=supervisor-based mitigation (S1)2026.02 | 17.6 | |
| LLM ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=None (Malicious)2026.02 | 16.8 | |
| GRAPH ROUTERMalicious Setting=M2-Activation Steering, Mitigation Strategy=supervisor-free mitigation2026.02 | 15.2 | |
| TEXT DEBATEMalicious Setting=M4-RL, Mitigation Strategy=None (Malicious)2026.02 | 14 | |
| TEXT FEEDBACKMalicious Setting=M2-Activation Steering, Mitigation Strategy=None (Benign)2026.02 | 14 | |
| SCEModels=Fuse2026.02 | 12.24 | |
| TEXT FEEDBACKMalicious Setting=M4-RL, Mitigation Strategy=supervisor-based mitigation (S2)2026.02 | 12 | |
| TEXT DEBATEMalicious Setting=M2-Activation Steering, Mitigation Strategy=None (Benign)2026.02 | 11.6 | |
| TEXT FEEDBACKMalicious Setting=M2-Activation Steering, Mitigation Strategy=None (Malicious)2026.02 | 11.6 | |
| TEXT FEEDBACKMalicious Setting=M4-RL, Mitigation Strategy=supervisor-free mitigation2026.02 | 11.6 | |
| TEXT FEEDBACKMalicious Setting=M4-RL, Mitigation Strategy=supervisor-based mitigation (S1)2026.02 | 11.6 | |
| Dare Task ArithmeticModels=Fuse2026.02 | 11.34 | |
| TEXT FEEDBACKMalicious Setting=M4-RL, Mitigation Strategy=None (Malicious)2026.02 | 11.2 | |
| Task ArithmeticModels=Fuse2026.02 | 10.48 | |
| Dare Ties MergingModels=Fuse2026.02 | 10.48 | |
| Ties MergingModels=Fuse2026.02 | 9.55 | |
| w/o MergingModels=Math2026.02 | 8.36 | |
| SCF-RKLModels=Fuse2026.02 | 7.46 | |
| w/o MergingModels=Code2026.02 | 5.07 |