Multi-Turn Constraint Adaptation on FlexTravelBench 3-Turn Global-to-Local
100Delivery RateReAct (GPT-4)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| ReAct (GPT-4)Backbone=GPT-42026.03 | 100 | 84.06 | 32.8 | 59.43 | 36.24 | 17.2 | |
| ATLAS (Gemini-2.5-Pro)Backbone=Gemini-2.5-Pro2026.03 | 100 | 86.81 | 47.09 | 71.38 | 52.12 | 31.75 | |
| HiMAP-Travel (4B)Model Size=4B2026.03 | 100 | 87.5 | 48.2 | 72.6 | 53.35 | 33.15 | |
| HiMAP-Travel (8B)Model Size=8B2026.03 | 100 | 88.2 | 49.8 | 73.65 | 54.4 | 37.32 |