Add paper, project page, and GitHub links
#2
by nielsr HF Staff - opened
README.md
CHANGED
|
@@ -1,27 +1,51 @@
|
|
| 1 |
-
---
|
| 2 |
-
configs:
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
configs:
|
| 3 |
+
- config_name: attraction
|
| 4 |
+
data_files:
|
| 5 |
+
- split: test
|
| 6 |
+
path: attraction/test*
|
| 7 |
+
- config_name: restaurant
|
| 8 |
+
data_files:
|
| 9 |
+
- split: test
|
| 10 |
+
path: restaurant/test*
|
| 11 |
+
- config_name: hotel
|
| 12 |
+
data_files:
|
| 13 |
+
- split: test
|
| 14 |
+
path: hotel/test*
|
| 15 |
+
- config_name: data all
|
| 16 |
+
data_files:
|
| 17 |
+
- split: test
|
| 18 |
+
path: data all/test*
|
| 19 |
+
- config_name: evals
|
| 20 |
+
data_files:
|
| 21 |
+
- split: test
|
| 22 |
+
path: evals/test*
|
| 23 |
+
- config_name: human query
|
| 24 |
+
data_files:
|
| 25 |
+
- split: test
|
| 26 |
+
path: human query/test*
|
| 27 |
+
task_categories:
|
| 28 |
+
- text-generation
|
| 29 |
+
---
|
| 30 |
+
|
| 31 |
+
# ItinBench: Benchmarking Planning Across Multiple Cognitive Dimensions with Large Language Models
|
| 32 |
+
|
| 33 |
+
[**Project Page**](https://ethanwtl.github.io/IBweb/) | [**Paper**](https://huggingface.co/papers/2603.19515) | [**GitHub**](https://github.com/EthanWTL/ItinBenchV1)
|
| 34 |
+
|
| 35 |
+
ItinBench is a benchmark designed to evaluate the cognitive capabilities of Large Language Models (LLMs) in real-world planning scenarios. Specifically, it focuses on trip itinerary planning by integrating spatial reasoning (route optimization) with traditional verbal reasoning tasks.
|
| 36 |
+
|
| 37 |
+
## Dataset Summary
|
| 38 |
+
ItinBench evaluates models across diverse tasks simultaneously, providing insights into how LLMs handle multiple cognitive dimensions concurrently. The dataset includes configurations for:
|
| 39 |
+
- **Attractions**
|
| 40 |
+
- **Restaurants**
|
| 41 |
+
- **Hotels**
|
| 42 |
+
- **Human Queries**
|
| 43 |
+
- **Evaluation Sets**
|
| 44 |
+
|
| 45 |
+
## Supported Tasks
|
| 46 |
+
- **Spatial Reasoning**: Route optimization within a trip.
|
| 47 |
+
- **Verbal Reasoning**: Constraint-based itinerary planning.
|
| 48 |
+
- **Tool Use**: Evaluating the ability of models to use external tools for planning.
|
| 49 |
+
|
| 50 |
+
## Evaluation
|
| 51 |
+
The benchmark has been used to evaluate various models, including Llama 3.1 8B, Mistral Large, Gemini 1.5 Pro, and the GPT family. Results indicate that LLMs often struggle to maintain consistent performance when faced with tasks from multiple cognitive domains.
|