ACE

ACE (Active learning for Capability Evaluation) is a novel framework that uses active learning and powerful language models to automate fine-grained evaluation of foundation models. It enables scalable, adaptive testing that uncovers strengths and weaknesses beyond static benchmarks.

Installing dependencies

The development environment can be set up using poetry. Hence, make sure it is installed and then run:

python3 -m poetry install
source $(poetry env info --path)/bin/activate

In order to install dependencies for testing (codestyle, unit tests, integration tests), run:

python3 -m poetry install --with test

[Optional] Google Cloud Authentication

The capability evaluation logs (evaluated using Inspect) are stored in a GCP bucket. Use the following command to log in using your GCP account:

gcloud auth application-default login

Run pipeline

Configuration

1. Set environment variables:

• OPENAI_API_KEY
• GOOGLE_API_KEY - To use LLMs provided by Google
• ANTHROPIC_API_KEY - To use LLMs provided by Anthropic
• Rate limit vars (default values given):
  • RATE_LIMIT_CALLS=5
  • RATE_LIMIT_PERIOD=60
• LangSmith tracing vars:
  • LANGSMITH_TRACING=true
  • LANGSMITH_ENDPOINT="https://api.smith.langchain.com"
  • LANGSMITH_API_KEY=<langsmith_api_key>
  • LANGSMITH_PROJECT="automated_capability_evaluation"
• GCP env vars:
  • GOOGLE_CLOUD_PROJECT=<project_id>

2. Modify src/cfg/run_cfg.yaml, if required.

Capability Generation using the scientist LLM

Generates capability names and descriptions in the first step. In the second step, for each capability, it generates tasks, solves them, and verifies the solutions.

python -m src.run_capability_generation

Evaluation of subject LLM on generated capabilities

Evaluates the subject LLM on the generated capabilities and calculates a score for each.

python -m src.run_evaluation

Capability selection/generation using active learning

Utilize the capability and the corresponding subject LLM score to select or generate a new capability.

python -m src.run_lbo

Agentic Generation Scripts

These scripts implement the multi-agent debate workflow for automated generation of areas, capabilities, tasks, and solutions. All configurable parameters are defined in src/cfg/agentic_config.yaml.

Understanding Pipeline Tags

The pipeline uses auto-generated tags to organize outputs from each step. Understanding how tags work is essential for running the pipeline:

• Tag Format: Tags are automatically generated timestamps in the format _YYYYMMDD_HHMMSS (e.g., _20251104_143022)
• Auto-Generation: When you run a step (e.g., Generate Areas), the script automatically creates a tag and includes it in the output path
• Finding Tags: After running a step, check the console output or the output directory to see the generated tag. The tag appears in the file path where outputs are saved
• Using Tags: To run the next step in the pipeline, you need to specify the tag from the previous step's output:
  • Step 2 (Generate Capabilities) needs areas_tag from Step 1
  • Step 3 (Generate Tasks) needs capabilities_tag from Step 2
  • Step 4 (Generate Solutions) needs tasks_tag from Step 3

Example Workflow:

1. Run python -m src.agentic_area_generator → outputs to .../areas/_20251104_143022/areas.json
2. Use the tag _20251104_143022 in the next step:

   python -m src.agentic_capability_generator pipeline_tags.areas_tag=_20251104_143022

3. The capability generator outputs to .../capabilities/_20251104_150315/...
4. Use this new tag for the next step, and so on.

---

1. Generate Areas

Generate domain areas using the scientist–moderator debate system:

python -m src.agentic_area_generator

This step auto-generates a tag (e.g., _20251104_143022) and outputs the results to:

Output location:

~/<output_dir>/<domain>/<exp_id>/areas/<areas_tag>/areas.json

Where:

• <output_dir> comes from global_cfg.output_dir
• <domain> comes from global_cfg.domain (spaces replaced with underscores)
• <exp_id> comes from exp_cfg.exp_id
• <areas_tag> is the auto-generated tag for this run (use this tag in Step 2)

2. Generate Capabilities

Generate capabilities for each area:

# Use the areas_tag from Step 1 (Generate Areas) output
python -m src.agentic_capability_generator pipeline_tags.areas_tag=_YYYYMMDD_HHMMSS pipeline_tags.resume_capabilities_tag=_YYYYMMDD_HHMMSS

Options:

• pipeline_tags.areas_tag specifies which set of areas to use when generating capabilities. This should be the <areas_tag> from the output of Step 1 (Generate Areas).
• pipeline_tags.resume_capabilities_tag (optional) resumes a previous capability generation run.

This step auto-generates a new tag for the capabilities output.

Output location:

~/<output_dir>/<domain>/<exp_id>/capabilities/<capabilities_tag>/<area>/capabilities.json

Where:

• <capabilities_tag> is the auto-generated tag for this run (use this tag in Step 3)

3. Generate Tasks

Generate evaluation tasks for a specific capabilities tag:

# Use the capabilities_tag from Step 2 (Generate Capabilities) output
python -m src.agentic_task_generator pipeline_tags.capabilities_tag=_YYYYMMDD_HHMMSS pipeline_tags.resume_tasks_tag=_YYYYMMDD_HHMMSS

Options:

• pipeline_tags.capabilities_tag specifies which set of capabilities to use when generating tasks. This should be the <capabilities_tag> from the output of Step 2 (Generate Capabilities).
• pipeline_tags.resume_tasks_tag (optional) resumes a previous task generation run.

This step auto-generates a new tag for the tasks output.

Output location:

~/<output_dir>/<domain>/<exp_id>/tasks/<tasks_tag>/[<area>]-[<capability>]/tasks.json

Where:

• <tasks_tag> is the auto-generated tag for this run (use this tag in Step 4)

4. Generate Solutions

Solve generated tasks using the multi-agent debate system:

# Use the tasks_tag from Step 3 (Generate Tasks) output
python -m src.agentic_task_solver pipeline_tags.tasks_tag=_YYYYMMDD_HHMMSS pipeline_tags.resume_solutions_tag=_YYYYMMDD_HHMMSS

Options:

• pipeline_tags.tasks_tag specifies which set of tasks to solve. This should be the <tasks_tag> from the output of Step 3 (Generate Tasks).
• pipeline_tags.resume_solutions_tag (optional) resumes a previous solution generation run.

This step auto-generates a new tag for the solutions output.

Output location:

~/<output_dir>/<domain>/<exp_id>/task_solutions/<solutions_tag>/[<area>]-[<capability>]/<task_id>_solution.json

Where:

• <solutions_tag> is the auto-generated tag for this run

Wikipedia-Based Analysis Tools

Tools for extracting, processing, and matching mathematical capabilities from Wikipedia. All prompts are centralized in wikipedia/prompts.py.

Wikipedia Glossary Scraper

Scrapes Wikipedia's "Glossary of areas of mathematics", extracts capability descriptions, and generates summaries with LLM-powered categorization.

cd wikipedia
python wikipedia_scraper.py

Outputs JSON files to wikipedia/pages/ containing capability_name, description, summary, area, url, and timestamp.

Wikipedia-Generated Capability Matcher

Matches Wikipedia capabilities with generated capabilities using LLM-based similarity analysis. Supports bidirectional matching.

Configure wikipedia/cfg/wiki_vs_generated.yaml:

• data_cfg.wikipedia_pages_dir: Wikipedia pages directory
• data_cfg.generated_dir: Generated capabilities directory
• processing_cfg.match_direction: generated_to_wikipedia or wikipedia_to_generated

cd wikipedia
python wiki_vs_generated.py

Dataset Question Categorizer

Categorizes questions from GSM8K or MATH datasets into mathematical areas using generated or Wikipedia taxonomies. Supports checkpoint-based resume.

Configure wikipedia/cfg/static_vs_generated.yaml:

• data_cfg.dataset_name: gsm8k or math
• data_cfg.dataset_path: Dataset file (GSM8K) or directory (MATH)
• categorization_cfg.extraction_method: generated or wikipedia

cd wikipedia
python static_vs_generated.py