Agentic (multi-turn) evaluation on NitiBench
78.02AccuracyInK-GRPO
Evaluation results
| Method | Links | |
|---|---|---|
| InK-GRPOagentic=true2026.01 | 78.02 | |
| GPT-5agent=true2026.01 | 75.34 | |
| GRPOagentic=true2026.01 | 73.73 | |
| Qwen3-4B-Instruct-2507agent=true2026.01 | 46.11 | |
| GPT-5search=true2026.01 | 38.07 |