GAIA¶

Overview¶

mcpbr run -c config.yaml --benchmark gaia -n 10

Overview¶

GAIA (General AI Assistants) is a benchmark designed to evaluate AI systems on real-world questions that require fundamental assistant capabilities. Unlike many academic benchmarks that test narrow skills, GAIA questions are intentionally designed to be:

• Easy for humans: A human with access to standard tools (web browser, calculator, etc.) can answer most questions in minutes
• Hard for AI: Questions require combining multiple capabilities that current AI systems struggle to integrate

GAIA covers a broad range of tasks including:

• Factual reasoning: Questions that require multi-step logical deduction
• Multi-modality: Tasks involving images, audio, or documents
• Web browsing: Questions requiring information retrieval from the internet
• Tool use: Tasks that need calculator, code execution, or API access
• File handling: Questions that involve processing attached files

Each question has an unambiguous, fact-based answer that can be verified without subjective judgment.

What It Measures¶

GAIA evaluates the fundamental capabilities of a general AI assistant:

• Information retrieval: Finding specific facts from the web, databases, or attached files
• Multi-step reasoning: Chaining multiple logical steps to arrive at a conclusion
• Tool orchestration: Selecting and combining the right tools (web search, calculator, code execution) for each sub-task
• Multi-modal understanding: Processing and reasoning about images, PDFs, spreadsheets, and other file formats
• Precise answer generation: Producing concise, factually correct answers in the expected format
• Planning and decomposition: Breaking complex questions into manageable sub-problems

GAIA does not test:

• Creative writing or open-ended generation
• Code generation or software engineering
• Long-form document synthesis
• Conversational ability or personality

Difficulty Levels¶

GAIA organizes questions into three difficulty levels:

Task Structure¶

Each GAIA task contains the following fields:

The agent receives the question with its difficulty level and must provide a concise, factual answer.

Example Task (Level 1)¶

Difficulty Level: 1

Question: What is the population of the capital city of the country where
the 2024 Summer Olympics were held? Give your answer to the nearest million.

Expected Answer: 2 million

Example Task (Level 3)¶

Difficulty Level: 3

Question: Download the CSV file at [URL]. Calculate the median value in the
'revenue' column for entries where the 'region' field is 'North America',
then convert this value to EUR using the exchange rate on January 15, 2024.
Round to the nearest hundred.

Expected Answer: 45200

Configuration¶

Basic Configuration¶

# Run GAIA with default settings
mcpbr run -c config.yaml --benchmark gaia

# Run a small sample
mcpbr run -c config.yaml --benchmark gaia -n 10

# Filter by difficulty level using the level parameter
mcpbr run -c config.yaml --benchmark gaia --level 1

# Filter by difficulty using filter-difficulty
mcpbr run -c config.yaml --benchmark gaia --filter-difficulty 1

# Run only Level 3 (hardest) questions
mcpbr run -c config.yaml --benchmark gaia --filter-difficulty 3

# Run with extended timeout for complex tasks
mcpbr run -c config.yaml --benchmark gaia -n 10 -v -o results.json

benchmark: "gaia"
sample_size: 10
timeout_seconds: 300

mcp_server:
  command: "npx"
  args: ["-y", "@modelcontextprotocol/server-filesystem", "{workdir}"]

model: "sonnet"

# Optional: filter by difficulty level
filter_difficulty:
  - "1"
  - "2"

Advanced Options¶

Level Filtering¶

GAIA supports two methods for filtering by difficulty:

1. --level flag: Directly sets the difficulty level (1, 2, or 3). This is applied first during task loading.

2. --filter-difficulty flag: Accepts difficulty level strings. Applied after level filtering, providing additional refinement.

Both methods are case-sensitive on the numeric value. Valid values are 1, 2, and 3.

Configuration for Level-Specific Evaluation¶

# Level 1 only -- fast baseline
benchmark: "gaia"
sample_size: 20
timeout_seconds: 180
filter_difficulty:
  - "1"

model: "sonnet"

# Level 3 only -- hardest tasks, extended timeout
benchmark: "gaia"
sample_size: 10
timeout_seconds: 600
max_iterations: 30
filter_difficulty:
  - "3"

model: "opus"

Configuration with Rich Tool Access¶

# GAIA with web search and filesystem tools
benchmark: "gaia"
sample_size: 15
timeout_seconds: 300

mcp_server:
  command: "npx"
  args: ["-y", "@modelcontextprotocol/server-filesystem", "{workdir}"]

# Additional MCP servers for broader tool coverage
# (configure based on your available servers)

model: "sonnet"

agent_prompt: |
  {problem_statement}

  Provide your final answer as concisely as possible. Match the expected format exactly.
  If the answer is a number, provide just the number. If it is a name, provide just the name.

Evaluation Methodology¶

GAIA uses exact match evaluation with normalization:

1. Answer normalization: Both the ground truth and the model's response are stripped of leading/trailing whitespace and lowercased.

2. Exact match check: The normalized ground truth is compared to the normalized response. A match occurs if:

3. The ground truth exactly equals the response, OR

4. The ground truth is a substring of the response

5. Result determination: The task is resolved if and only if the normalized ground truth matches or is contained within the normalized response.

Scoring Details¶

resolved = (gt_normalized == solution_normalized) OR (gt_normalized in solution_normalized)

Where: - gt_normalized: Ground truth answer, stripped and lowercased - solution_normalized: Model's response, stripped and lowercased

Answer Format¶

GAIA expects concise, factual answers. The substring matching allows for some flexibility:

• "Paris" matches ground truth "paris" (exact match after normalization)
• "The answer is 42" matches ground truth "42" (substring match)
• "Based on my research, the population is approximately 2 million" matches ground truth "2 million" (substring match)

However, overly verbose responses risk matching unintended substrings. Keep answers concise.

Interpreting Results¶

Key Metrics¶

What Good Results Look Like¶

Common Failure Patterns¶

Example Output¶

Successful Evaluation¶

{
  "resolved": true,
  "agent_answer": "The population of Paris is approximately 2 million.",
  "ground_truth": "2 million"
}

Failed Evaluation (Wrong Answer)¶

{
  "resolved": false,
  "agent_answer": "The population is about 11 million in the metropolitan area.",
  "ground_truth": "2 million"
}

Failed Evaluation (No Ground Truth)¶

{
  "resolved": false,
  "error": "No ground truth answer available"
}

Best Practices¶

Recommended Workflow¶

1. Start with Level 1: Begin evaluation with Level 1 questions to establish a baseline and verify configuration
2. Evaluate each level separately: Track accuracy per level for a nuanced view of capabilities
3. Add tool access incrementally: Test with minimal tools first, then add web search, computation tools to measure their impact
4. Compare with human baseline: Use the GAIA paper's human accuracy numbers to contextualize results

Performance Tips¶

• Use generous timeouts: Set timeout_seconds to at least 300 for Level 2 and 600 for Level 3
• Provide tool access: GAIA is designed to test tool use. MCP servers with web browsing, file access, and computation tools produce the best results
• Keep answers concise: The exact-match evaluation penalizes verbose responses. Instruct the model to provide only the final answer without extensive explanation
• Monitor the agent_answer field: The evaluation truncates stored agent answers to 500 characters. For debugging, use verbose output to see full responses

Cost Optimization¶

• Level 1 is cheapest: Fewer reasoning steps and tool calls required per task
• Level 3 is most expensive: Multi-hop reasoning, multiple tool calls, and extended thinking increase token usage significantly
• Start with 10-20 tasks per level: This provides a statistically meaningful sample without excessive cost
• Use sonnet for Level 1-2: Simpler tasks work well with faster, cheaper models
• Reserve opus for Level 3: Complex planning tasks benefit from the strongest reasoning capability
• Filter by level to avoid running expensive Level 3 tasks during initial configuration testing

Common Issues & Solutions¶

Comparison with Similar Benchmarks¶

References¶

• GAIA Paper (arXiv)
• GAIA Dataset on HuggingFace
• GAIA Leaderboard
• MCPToolBench++ -- MCP-specific tool use benchmark
• AgentBench -- multi-environment agent benchmark
• ToolBench -- real-world API tool use benchmark
• Benchmarks Overview
• Configuration Reference
• CLI Reference