feat(benchmark): add support for evaluation on futurex

[JubSteven]

Describe this PR

Overview

Integrates Futurex-Online prediction dataset into MiroFlow's benchmark system with majority voting for improved prediction accuracy (adapted from MiroThinker).

Key Changes

🆕 New Files

• utils/prepare_benchmark/gen_futurex.py - Dataset generator
• config/benchmark/futurex.yaml - Benchmark configuration
• scripts/run_evaluate_multiple_runs_futurex.sh - Multi-run evaluation script
• docs/mkdocs/docs/futurex.md - Complete documentation
• docs/mkdocs/mkdocs.yml - Add documentation link for futurex
• utils/extract_futurex_results.py - Extract results from logging and implement majority voting from MiroThinker for multiple runs.

🔧 Modified Files

• utils/prepare_benchmark/main.py - Added Futurex support

Features

• 61 prediction tasks (political events, sports, legal proceedings)
• Majority voting across multiple runs with tie-breaking
• Standardized format using MiroFlow's Task class
• No ground truth (prediction-based evaluation)

Usage

# Download dataset
uv run main.py prepare-benchmark get futurex

# Single run
uv run main.py common-benchmark --config_file_name=agent_quickstart_1 benchmark=futurex output_dir="logs/futurex"

# Multiple runs with voting
./scripts/run_evaluate_multiple_runs_futurex.sh

Checklist for PR

Must Do

• Write a good PR title and description, i.e. feat(agent): add pdf tool via mcp, perf: make llm client async and fix(utils): load custom config via importlib etc. CI job check-pr-title enforces Angular commit message format to PR title.
• Run make precommit locally. CI job lint enforce ruff default format/lint rules on all new codes.
• Run make pytest. Check test summary (located at report.html) and coverage report (located at htmlcov/index.html) on new codes.

Nice To Have

• (Optional) Write/update tests under /tests for feat and test PR.
• (Optional) Write/update docs under /docs for docs and ci PR.

[BinWang28]

This is a typo from old days, should be Claude 3.7

[BinWang28]

mention in the documents that this is a quick start for running futurex benchmark and prepare results, not for fully reproduce the results that is submitted

[ntudy]

should be the runned dir?

[ntudy]

error: unrecognized arguments: --log_dir

[JubSteven]

error: unrecognized arguments: --log_dir

My bad, should be the following format

uv run utils/extract_futurex_results.py logs/futurex/$(date +"%Y%m%d_%H%M")

[JubSteven]

I have pushed a commit that should resolve the issues regarding arguments and documentation.