Improve dataset card: add metadata and link to paper

README.md

@@ -1,22 +1,51 @@
-# DisasterBench
+---
+license: mit
+task_categories:
+- text-generation
+tags:
+- planning
+- multi-agent
+- disaster-response
+---
 
-DisasterBench is a benchmark for evaluating LLM structured planning over typed multi-agent disaster-response workflows.
+# DisasterBench: Benchmarking LLM Planning under Typed Tool Interface Constraints
+
+DisasterBench is a benchmark for evaluating whether Large Language Models (LLMs) can generate executable structured workflows in disaster-management scenarios. It focuses on workflow grounding under typed interface constraints, requiring models to compose specialized agents into valid multi-step plans with correct agent selection, parameter grounding, and dependency propagation.
+
+- **Paper:** [DisasterBench: Benchmarking LLM Planning under Typed Tool Interface Constraints](https://huggingface.co/papers/2605.27957)
+- **Repository:** [TamuChen18/DisasterBench_Open](https://github.com/TamuChen18/DisasterBench_Open)
 
 ## Benchmark Statistics
 
-- 233 planning tasks
-- 26 task-interface agents
-- Typed compatibility graph
-- Structured executable workflow plans
+- **233** expert-verified planning tasks
+- **26** task-interface agents across 13 functional categories
+- **81** typed compatibility edges
+- **Workflow Structures:** Node, Chain, and DAG
+- **Planning Paradigms Evaluated:** Direct Planning (DP), Chain-of-Thought (CoT), Tree-of-Thought (ToT), RAP (MCTS), and ReAct.
 
 ## Dataset Structure
 
-Each example contains:
+Each example in the dataset contains the following fields:
+
+- `task_id`: Unique identifier for the planning task.
+- `task_desc`: Natural language description of the disaster-response goal.
+- `structured_plan`: The ground-truth executable workflow plan including agents, parameters, and dependency structures.
+
+## Evaluation Metrics
 
-- `task_id`
-- `task_desc`
-- `structured_plan`
+The benchmark evaluates models using several metrics:
+- **Overall**: Exact structured-plan match.
+- **Tools**: Correct tool selection.
+- **Parameters**: Correct parameter grounding.
+- **Dependencies**: Correct dependency structure.
+- **FPoF (First-Point-of-Failure)**: Localization of the earliest root cause failure in a predicted workflow.
 
-## Links
+## Citation
 
-- Code: https://github.com/TamuChen18/DisasterBench_Open
+```bibtex
+@misc{chen2026disasterbench,
+  title={DisasterBench: Benchmarking LLM Planning under Typed Tool Interface Constraints},
+  author={Chen, Zhitong and Yin, Kai and Zhang, Weifeng and Wang, Zhiyuan and Dong, Xiangjue and Liu, Chengkai and Liu, Zhewei and Xiao, Yiming and Mostafavi, Ali and Caverlee, James},
+  year={2026}
+}
+```