ActiveARC

Interactive query → test trials over ARC-style tasks: explore with a verifier-backed oracle, then solve a held-out test input. The Streamlit UI lives in interface/active_arc_app.py.

Quick start

pip install -r requirements.txt
streamlit run interface/active_arc_app.py

Pass app flags after -- (Streamlit does not recognize them otherwise). Flags are combinable:

# One free training pair is shown at start by default (no query cost); disable with:
streamlit run interface/active_arc_app.py -- --no-hot-start

# Randomly corrupt query outputs (default p=0.12; clamped to 0.05–0.20)
streamlit run interface/active_arc_app.py -- --noisy-science
streamlit run interface/active_arc_app.py -- --noisy-science --noise-probability 0.15

# Wrong test answer sends you back to exploration (+10 query count)
streamlit run interface/active_arc_app.py -- --re-trials

# Fixed RNG for task/verifier selection
streamlit run interface/active_arc_app.py -- --seed 42

# Pin a specific ARC task (needs a validated verifier + ARC-GEN dynamic pair)
streamlit run interface/active_arc_app.py -- --task-id 8eb1be9a --seed 42

# Combine features
streamlit run interface/active_arc_app.py -- --hot-start --noisy-science --re-trials --seed 42

Legacy single-feature alias: --mode hot_start (same as --hot-start; OR-combined with the flags above).

For the pixel grid editor, build the frontend once:

cd interface/arc_grid_component/frontend && npm install && npm run build

Experimentation

ARC task 8eb1be9a is a good example to try when experimenting with the interface and modes.

External ARC editor (external/arc-app)

The optional Node editor under external/arc-app needs a MongoDB Atlas URI. Credentials are not stored in source; each developer keeps a local external/arc-app/.env (gitignored):

cd external/arc-app
cp .env.example .env
# set MONGODB_URI=... (ask a teammate / project lead for the shared URI)
npm install
npm run build && npm start

See external/arc-app/README.md for full setup.

ConceptARC dataset

ActiveARC can also run trials over ConceptARC DSL programs, kept fully separate from the ARC-AGI pool. Choose the pool with --dataset (or the Dataset selector in the sidebar); ARC-AGI stays the default:

# Random ConceptARC program
streamlit run interface/active_arc_app.py -- --dataset conceptarc

# A specific ConceptARC program (ids are <concept>/<task>, e.g. count/count11, copy/copy12)
streamlit run interface/active_arc_app.py -- --dataset conceptarc --task-id count/count11

The same --dataset conceptarc flag works for the headless agent runner (pipelines/run_active_arc_agent.py) and create_trial_session(..., dataset="conceptarc").

Exported programs live under external/conceptarc/programs/<concept>/<task>.json (five concepts: count, center, insideoutside, abovebelow, copy). Official tasks 1–10 reuse the ConceptARC corpus examples; generated families >10 ship three freshly generated examples plus a held-out test pool. The live query verifier and the dynamic test generator are rebuilt from each program via the conceptarc_gen package, imported from CONCEPTARC_GEN_ROOT (default: the sibling ../../ConceptARC-Generator/ConceptARC-GEN checkout). ConceptARC requires Python 3.12 and the ConceptARC-GEN repo's pinned external/ARC-GEN submodule (for the common module).

To (re)generate the exported programs, run from the ConceptARC-GEN repo root:

PYTHONPATH="$(pwd)/external/ARC-GEN" .venv/bin/python export_to_activearc.py \
    --out /path/to/ActiveARC/external/conceptarc/programs