[ACL 2026 Main] CostBench

This is the official repository for paper "CostBench: Evaluating Multi-Turn Cost-Optimal Planning and Adaptation in Dynamic Environments for LLM Tool-Use Agents" accepted in ACL 2026 Main.

🎯 Project Overview

CostBench is a comprehensive benchmark for evaluating multi-turn cost-optimal planning and adaptation capabilities of large language models (LLMs) in tool-using scenarios.

The benchmark systematically assesses how LLM agents navigate complex tool-calling environments by testing their ability to:

• 📋 Cost-Optimal Planning: Plan cost-optimal multi-step tool invocation sequences in static environments
• 🔄 Dynamic Adaptation: Dynamically adapt their strategies when tool costs, availability, or preferences change during execution in dynamic environments

✨ Core Features

• Hierarchical Tool System: Supports atomic and composite tools, each with clear input/output types and costs
• Flexible Cost Assignment: Supports configurable cost ranges for atomic tools and composite tools with component-based cost calculation plus Gaussian noise, enabling customizable cost distributions for evaluation scenarios
• Dynamic Blocking: Supports multiple blocking modes (cost changes, preference changes, tool disabling, etc.) to test model adaptation capabilities
• Adjustable Difficulties: Supports different levels of task sequence to control task complexity
• Reproducible Random System: Features a seed-controlled pseudo-random system that ensures reproducibility across runs while preventing data leakage through deterministic randomization

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/JiayuJeff/CostBench.git
cd CostBench

# Install the package in editable mode
pip install -e .

# Or install dependencies only
pip install -r requirements.txt

Search Database Setup

python env/domains/travel/generate_search_space.py \
    --generate_test \
    --test_config_path env/domains/travel/test_travel_config.yaml \
    --num-times 5000 \
    --num-time-combinations 5000 \
    --seed 42

This script will generate the search database for the search tools in our environment. Note: The seed value 42 is the random seed used in our experiments. Users can choose other random seeds as needed.

Environment Configuration

1. Local Model Deployment

   Replace <your_base_url> in the model.endpoints field of env/config/travel_config.yaml with your own port.

2. API Calls

   Replace <your_api_key_env> and <your_base_url> in the endpoints.base_url field of env/config/travel_config.yaml with your own API key environment variable name and port.

   Then, configure your API key in env/.env:

# Example configuration
OPENAI_API_KEY=your_api_key_here
OPENAI_BASE_URL=https://api.openai.com/v1

3. Check and modify the configuration in env/config/travel_config.yaml (if needed).

Running Examples

Main Results

Basic run command (Table 4 results):

python env/run.py \
    --refinement_level 2 \
    --ban_longest_tool \
    --model_name gpt-5 \
    --num_threads 10 \
    --output_dir outputs/

Run with dynamic blockings (could change cost_change to other types):

python env/run.py \
    --refinement_level 2 \
    --ban_longest_tool \
    --model_name gpt-5 \
    --num_threads 10 \
    --use_blocker \
    --block_num 1 \
    --block_mode cost_change \
    --output_dir outputs/

📖 Advanced Usage

For advanced configuration, please modify the configuration file to customize tool parameters, blocking behavior, model endpoints, and other settings.

Tool-Related Parameters

• --tool_creation_seed: Random seed for tool generation, controlling the tool creation seed and cost change batches for each query
• --refinement_level: Tool refinement level, controlling task complexity (defaults to maximum depth). For advanced usage, `task_sequence equals = refinement_level + 3'.
• --max_tool_steps: Maximum number of tool-calling steps
• --min_atomic_cost / --max_atomic_cost: Cost range for atomic tools
• --noise_std: Noise scaling factor for composite tools
• --ban_longest_tool: Whether to disable tools that complete the task in one step

Blocking-Related Parameters

• --use_blocker: Enable dynamic blocking functionality
• --block_mode: Blocking mode (preference_change, cost_change, steplen_change, ban_tool)
• --block_num: Number of dynamic blocking events per query

Query-Related Parameters

• --query_path: Path to query file (JSON format)
• --start_index / --end_index: Index range of queries to process (-1 means process all remaining queries)

Model-Related Parameters

• --model_name: Model name (must be defined in configuration file)
• --temperature: Sampling temperature
• --max_tokens: Maximum number of generated tokens

Runtime Parameters

• --num_threads: Number of concurrent threads
• --output_dir: Output directory for results
• --require_goal_state: Still in construction. Unexpected behavior may happen if set to true.

Simulation-Related Parameters

• --use_stimulation: Enable random strategy simulation
• --stimulation_num: Number of simulation runs per query
• --greedy: Use greedy selection strategy in simulation. Would use random policy if set to False.

You can specify a custom configuration file path through the COSTBENCH_TRAVEL_CONFIG environment variable.

🔧 Extending to New Domains or Larger Task Sequences

Still in construction.

🤝 Contributing

Contributions are welcome! Please feel free to submit Issues and Pull Requests.

📚 Citing this work

@article{liu2025costbench,
  title={CostBench: Evaluating Multi-Turn Cost-Optimal Planning and Adaptation in Dynamic Environments for LLM Tool-Use Agents},
  author={Liu, Jiayu and Qian, Cheng and Su, Zhaochen and Zong, Qing and Huang, Shijue and He, Bingxiang and Fung, Yi R},
  journal={arXiv preprint arXiv:2511.02734},
  year={2025}
}