1 of 1

• Travel planning is more than satisfying constraints.
• A feasible itinerary may still be a poor itinerary if it ignores what travelers actually experience.

​

​

​

​

​

​

• Travel planning requires multiple kinds of reasoning at once.
• LLMs are good at understanding preferences, but less reliable at exact scheduling. Small temporal errors: overlapping activities, unrealistic meal times, or missing transit buffers can make an otherwise good itinerary unusable.

​

​

[1] Chaudhuri et al. (2025). TripCraft: A benchmark for spatio-temporally fine-grained travel planning. ACL.

​

[2] Karmakar et al. (2026). TripTide: A benchmark for adaptive travel planning under disruptions. Findings of ACL.

​

[3] Xie et al. (2024). TravelPlanner: A benchmark for real-world planning with language agents. ICML.

​

[4] Choi et al. (2026). ATLAS: Constraints-aware multi-agent collaboration for real-world travel planning. ICLR.

​

TRIPPULSE:

Multi-Agent Travel Planning with Review-Grounded Reasoning

Priyanshu Karmakar¹ , Borru Vijay Sai¹ , Shubhojit Mallick² , Abhik Jana¹ , Shreya Ghosh¹, Manish Gupta²

IIT Bhubaneswar¹, Microsoft²

Why TripPulse

REFERENCES

Structured data tells us what a place offers — reviews tell us what it feels like

Ground travel planning in what people actually experienced.

Our Contributions

Review-augmented Benchmark

111,628

reviews

10,746

entities

140

U.S. cities

Reviews by domain

Accommodation

18,094

Restaurants

42,541

Attractions

50,993

Hybrid Neuro-symbolic Design

1

Semantic agents

Select & rank entities�using persona + reviews

2

Global orchestrator

Coordinates budget,�duration & city transitions

3

Scheduler

LLM flexibility OR�deterministic feasibility

Code & resources

(TripPulse)

​

• TripPulse, a multi-agent framework, combines review-grounded selection with deterministic scheduling to improve itinerary quality and feasibility.

​

• Review-grounded plans achieved 55.41–87.36% win rates under GPT-5 evaluation and were rated equal to or better in 31/40 human-evaluated pairs.

Conclusion

Qwen2.5: Comparison with Baseline Planners