Multi-Turn Constraint Adaptation on FlexTravelBench 3-Turn Local-to-Global
100Delivery Success RateReAct (GPT-4)
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| ReAct (GPT-4)Backbone=GPT-42026.03 | 100 | 83.7 | 33.07 | 59.59 | 36.51 | 15.34 | |
| ATLAS (Gemini-2.5-Pro)Backbone=Gemini-2.5-Pro2026.03 | 100 | 87.96 | 49.21 | 73.58 | 53.97 | 33.6 | |
| HiMAP-Travel (4B)Model Size=4B2026.03 | 100 | 88.65 | 50.35 | 74.8 | 55.15 | 35.2 | |
| HiMAP-Travel (8B)Model Size=8B2026.03 | 100 | 89.35 | 51.6 | 75.8 | 56.25 | 37.52 |