Lingjing-Solo is a single-agent world-model framework for interactive reasoning benchmarks. Instead of treating a benchmark as a sequence of isolated guesses, it gives one agent a compact loop for observing frames, maintaining hypotheses about the environment, choosing informative actions, planning short horizons, and reflecting when its current strategy stops making progress.
From multi-agent ideas to a single decision loop
The project reframes the original Lingjing engine for settings such as ARC-AGI-3 : hidden rules, turn-based interaction, zero-shot tasks, and a sensitive step budget. Its canonical lingjing_solo package combines five layers:
- Perception: frame differences, segmentation, and compact visual encodings.
- World model: transition records, state hashes, rule confidence, and loop detection.
- Exploration: information-gain scoring, rule induction, and goal inference.
- Planning: lightweight BFS-style search with an optional LLM strategy advisor.
- Reflection: targeted reconsideration when the agent detects loops, conflicts, or a shrinking action budget.
The design keeps the LLM in a strategic-advisor role rather than calling it on every step. The README describes a default per-level reflection budget, while evaluation can fall back to a lightweight route without external network access.
Why it belongs in Model Engineering & Applications
Lingjing-Solo is a useful example of model engineering around a model, not just model selection. The practical system includes state representation, action legality, transition evidence, resource budgets, fallback behavior, and benchmark adapters. Its public package has a NumPy-first core, an optional Torch CNN encoder, and a separate ARC-AGI-3/Kaggle integration path.
The repository also makes its current boundaries explicit: environment-specific win detection, deeper planner search, and optional LLM wiring remain integration points rather than hidden promises. That separation makes the project easier to test, adapt, and inspect as a research prototype.
Try the project
The repository documents two entry points: install the core package for local development, or use the ARC-AGI-3 Starter integration for a fixed local/Kaggle environment. Core self-checks include pytest -q, python test_solo.py, and the notebook template smoke test.
Project source: github.com/LingJing1001/Lingjing-Solo-