TravelEval - Preference-Weighted Experience: leaderboard
TravelEval (KDD 2026) asks an LLM for a complete multi-day travel plan and simulates it in a sandbox of 10 Chinese cities with real transport, hotel and attraction data. Scores are from Direct prompting over 1,150 queries. The board is the mean attraction value weighted by how well each attraction matches stated preferences (Profit in the paper), in unbounded units; higher is better.
Metric: Profit: mean over attractions of objective score times the preference-match weight (paper units, unbounded); Direct prompting (one pass, no tools), plans simulated in the TravelEval sandbox of 10 Chinese cities (real rail, flight, hotel and attraction data, queuing-time model, road-network distances) over 1,150 queries; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-11-20) | 5.97 | #866 |
| 2 | GPT-4o Mini (2024-07-18) | 5.97 | #1510 |
| 3 | Qwen 3 8B | 5.74 | #1830 |
| 4 | DeepSeek V3.1 (Non-reasoning) | 5.72 | #497 (DeepSeek V3.1) |
| 5 | Gemini 2.0 Flash | 5.71 | #704 |
| 6 | GPT-5 Chat | 5.65 | #444 |
| 7 | Claude Sonnet 4.5 | 5.44 | #250 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=traveleval-preference-weighted-experience · How It Works · Data refreshed daily, snapshot 2026-11-13.