LightsOut: Puzzle Paradigm in Quantum and AI

• LightsOut is a discrete puzzle paradigm with binary state spaces and local toggling, underpinning applications in quantum error correction, AI planning, and image preprocessing.
• Its methodology models the puzzle as a linear system over F₂, enabling efficient solutions to XOR constraints and minimal-weight decoding problems.
• The framework bridges theory and practice through neural-symbolic planning and diffusion-based outpainting, achieving notable performance improvements in quantum and vision domains.

LightsOut refers to a family of problems, methodologies, and algorithms rooted in the classic LightsOut puzzle. It serves as both a canonical discrete mathematics and planning benchmark and a structural prototype for a variety of computational and physical phenomena, including quantum code decoding and symbolic model learning from subsymbolic data. The formal structure of LightsOut—binary state spaces and local toggling operations—has led to its adoption in neural-symbolic planning, quantum error correction, and image pre-processing pipelines, each leveraging the puzzle’s algebraic and algorithmic properties.

1. Classic LightsOut Puzzle: Formalism and Structural Properties

The classic LightsOut puzzle is defined on an $n\times m$ grid of binary lights, each light in state ON ($1$) or OFF ($0$). The fundamental operation is the “press” move: selecting a cell (a “switch”), which toggles the state (ON↔OFF) of the selected cell and each of its orthogonal (up, down, left, right) neighbors. The objective is, from a given initial configuration, to determine a sequence of presses that turns every light OFF.

The group structure underpinning LightsOut is crucial: every move is involutive (reapplying the same switch twice yields no net effect), and the order of moves does not affect the final board configuration—the space of actions forms an Abelian group over $\mathbb{F}_2$. The solution set for a given initial state is characterized by the linear system $A x = b$ over $\mathbb{F}_2$, where $A$ is the adjacency matrix encoding toggling rules, $x$ is the vector of sought-after presses, and $b$ is the vector of initial light states (Berent et al., 2023).

2. LightsOut in Quantum Error Correction: Color Code Decoding as an XOR Puzzle

Quantum low-density parity-check (LDPC) codes, specifically CSS color codes embedded on three-valent, three-colorable 2D lattices (e.g., hexagonal), exhibit a mapping from code decoding to LightsOut-style XOR systems. Parity-check syndrome data (ON/OFF states of “lights”) sits on faces of the code’s lattice, while possible Pauli $X$ or $1$0 corrections (the “switches”) reside on vertices. The action of toggling a switch flips all lights in its adjacency set (vertices incident to faces), mirroring the puzzle's rules.

Solving for a recovery operation compatible with a measured syndrome reduces to a LightsOut parity system: for each face $1$1, the parity (XOR) of adjacent switches equals the face’s syndrome value,

$1$2

The minimal-weight decoding problem—critical for reducing logical error rates in quantum computation—then corresponds to finding a minimal-cardinality solution $1$3 subject to XOR constraints, precisely the MaxSAT/LightsOut minimization (Berent et al., 2023). This algebraic approach yields competitive decoder thresholds ($1$4), nearly matching the optimal tensor-network decoder ($1$5), yet at lower computational cost for low physical error rates. The LightsOut analogy further extends to more general code structures by adjusting neighborhood definitions and integrating time-like (syndrome noise) switches in three-dimensional stacking.

3. Neural-Symbolic Classical Planning: LightsOut as a PDDL Induction and Planning Benchmark

In AI planning, LightsOut has emerged as a canonical domain for evaluating neural-symbolic representation learning and action model induction. Latplan, an unsupervised framework, demonstrates that an agent can learn a fully-propositional STRIPS action model of the 5×5 LightsOut puzzle directly from raw images with unknown symbolic structure (Asai et al., 2021, Asai et al., 2017). The key steps include:

• State Abstraction: A discrete Variational Autoencoder (VAE) encodes each grid image into a propositional bit-vector $1$6, where $1$7 matches the board size (e.g., $1$8 for $1$9).
• Action Model Extraction: Given a dataset of valid image pairs $0$0, the system learns anonymous symbolic actions via a Gumbel-Softmax VAE and regresses STRIPS-style preconditions and effects or compiles a ground transition model into PDDL.
• Symbolic Planning: For novel start/goal images, encoded into binary vector states, an off-the-shelf planner (e.g., Fast Downward, A*, LAMA) can synthesize an action sequence that realizes an all-OFF solution, and the symbolic plan can be visualized step-wise at pixel-level by decoding latent vectors.

Empirical results on both canonical and “twisted” variants of LightsOut show 100% instance coverage and optimality for most heuristic planners, and robust generalization to minor image perturbations or board swirls (Asai et al., 2021, Asai et al., 2017). This instantiates a neural-symbolic bridge whereby fully propositional classical representations are induced without domain-specific supervision.

4. LightsOut as Preprocessing: Diffusion-Based Outpainting for Imaging Applications

Recent research exploits the structural motifs of LightsOut in vision pipelines, notably in the LightsOut diffusion-based outpainting method for lens flare removal (Tsai et al., 17 Oct 2025). Here, the analogy is operational: the pipeline “fills in” off-frame (invisible) light sources to allow more effective downstream single-image flare removal (SIFR) by standard neural networks. The methodology comprises:

• Light-Source Prediction: A multitask regression module predicts the parameters and confidences of $0$1 latent circular light sources outside the image frame, rendering a soft light mask $0$2.
• Diffusion-Based Outpainting: A LoRA-finetuned Stable Diffusion v2 U-Net inpaints the padded region, explicitly conditioned on the original image (padded), a boundary mask, BLIP-2 text prompt, and the synthetic $0$3. Composite output exhibits photorealistic and physically consistent halo and streak artifacts.
• Plug-and-Play Flare Removal: The completed image, inclusive of simulated off-frame light effects, is processed by any SIFR model (e.g., Zhou et al., Flare7K++, MFDNet) without network retraining or structural modification.

Quantitatively, LightsOut outpainting yields consistent SIFR improvements: across multiple backbones, PSNR increases by $0$40.3–1.4 dB, LPIPS decreases by $0$50.002–0.015, and SSIM improves by $0$60.005–0.02 in scenarios where light sources are incomplete or missing. Ablation indicates that both LoRA fine-tuning and light-source conditioning contribute substantially and synergistically to these gains (Tsai et al., 17 Oct 2025).

5. Computational and Theoretical Implications

The essential algebraic structure—linear systems over $0$7, locally constrained toggling, and irrelevance of action ordering—underpins the application of LightsOut in theoretical computer science, planning, and quantum information. In quantum decoding, the ability to cast syndrome matching as a LightsOut parity problem enables optimization via MaxSAT and analytical mapping to established Ising percolation thresholds. In AI planning, the propositional encoding and action model regularities allow for accurate, tractable learning and full transparency of induced symbolic models.

A plausible implication is that other toggling or parity-propagation settings—including error correction on bipartite graphs, neural circuit inference, or visual editing tasks—can be reduced to LightsOut-like systems, where XOR constraints and action minimization play central roles. This suggests a broader generality beyond its historic origins as a puzzle.

6. Limitations, Variations, and Future Directions

Several limitations and opportunities characterize the LightsOut paradigm:

• In quantum decoding, highly structured lattices with complex neighborhoods or multi-layer time dependencies may challenge the direct applicability of the classic LightsOut mapping, necessitating generalizations to nonlocal or conjunctive parity constraints (Berent et al., 2023).
• In Latplan, scalability to larger, less regular action/state spaces remains constrained by the combinatorics of transition enumeration and the expressiveness of learned action models (Asai et al., 2021, Asai et al., 2017).
• In vision applications, LightsOut-based outpainting introduces additional inference cost and may struggle with highly complex multicentric light geometries or severe color/intensity mismatches.

Proposed future work includes end-to-end training of outpainting and flare removal, conditioning on both light source and geometry information (e.g., depth, multiple illuminants), and extending the methodology to temporally coherent video outpainting scenarios (Tsai et al., 17 Oct 2025).

7. Summary Table: LightsOut in Key Research Domains

Domain	Role of LightsOut	Key Reference
Quantum error correction	Syndrome-to-correction mapping	(Berent et al., 2023)
Neural-symbolic planning	State/action model induction	(Asai et al., 2021, Asai et al., 2017)
Computer vision	Outpainting for preprocessing	(Tsai et al., 17 Oct 2025)

LightsOut thus exemplifies a recurring structural motif within discrete mathematics, computational planning, quantum information, and modern vision pipelines, offering a tractable, interpretable prototype for problems governed by local, invertible, binary operations.