JMaze-200K: Joint Multimodal Maze Benchmark

• JMaze-200K is a large-scale joint text–image dataset composed of perfect mazes with a unique solution and uniformly varied grid sizes for robust multimodal benchmarking.
• The corpus integrates paired textual prompts, visual representations, and chain-of-thought reasoning traces to enable precise and self-correcting cross-modal generation.
• It serves as a diagnostic benchmark within the SC-CMJP framework to test continuous negotiation between textual description and image rendering under exact-match criteria.

JMaze-200K is a large-scale joint text–image corpus for training and evaluating multimodal generation systems on maze solving, introduced by Le et al. as one of three corpora accompanying the Self-Correcting Coupled Markov Jump Process (SC-CMJP) framework in "Concurrent Image Understanding and Generation: Self-Correcting Coupled Markov Jump Processes" (Le et al., 14 Jul 2026). It was designed to stress-test joint samplers on a visual-reasoning problem with a unique ground-truth solution, a clear verification procedure for both modalities, and difficulty variation controlled by grid size. Within that program, JMaze-200K serves as a controlled benchmark for continuous cross-modal negotiation between textual commitments and pictorial execution during a single diffusion trajectory.

1. Role in joint multimodal reasoning

Visual-reasoning tasks such as maze solving tightly intertwine logical, textual reasoning and spatial, pictorial execution. The motivating claim behind JMaze-200K is that prior multimodal diffusion samplers tended to decode text and image in parallel but independently, so they could commit to a textual solution that the image branch would never render, or conversely produce an image inconsistent with the text. JMaze-200K was therefore created to force a model to negotiate text and image commitments continuously, detect cross-modal inconsistencies, and self-correct them within a single diffusion trajectory (Le et al., 14 Jul 2026).

The corpus occupies a specific place within the broader SC-CMJP benchmark suite. The paper introduces three large-scale joint multimodal generation corpora—JEdit-1M, JMaze-200K, and JNono-200K—and reports that the associated sampler achieves best joint performance for image understanding and editing as well as visual reasoning, including maze solving. This suggests that JMaze-200K is intended less as a generic maze repository than as a diagnostic environment for coupled text–image decoding under exact-match constraints.

2. Generation procedure and maze structure

JMaze-200K contains 200,000 perfect mazes produced with the publicly available maze-dataset library (Ivanitskiy et al. 2023) (Le et al., 14 Jul 2026). In this setting, “perfect” means that there is exactly one simple path between any two cells, with no loops, and the mazes are typically generated via depth-first search backtracking. The use of perfect mazes is consequential because it removes ambiguity in the target path: from the designated start to the designated end, only one valid simple solution exists.

Grid sizes are sampled uniformly as

$n \sim \mathrm{Uniform}\{6,7,\dots,20\},$

so each size from $6\times 6$ through $20\times 20$ is equally likely. Within each maze, one cell is marked green as the start and a different cell is marked red as the end; walls are rendered in black and open corridors in white. By varying $n$, the construction controls both average path length and the number of dead-end branches, since a perfect $n\times n$ maze has approximately $n^2$ cells and $(n^2-1)$ passages. Beyond grid size, structure is fully randomized by DFS order, yielding a broad distribution of corridor lengths, branching factors, and path lengths.

This combination of uniqueness, randomized topology, and explicit start/end marking makes the corpus particularly suitable for benchmarking exact correspondence between symbolic path output and spatial overlay. A plausible implication is that many common confounds in open-ended multimodal generation—especially multiple equally acceptable outputs—are deliberately minimized.

3. Representations, targets, and reasoning traces

Each training example includes both textual and visual conditioning, together with paired text and image targets (Le et al., 14 Jul 2026). The text prompt contains a text-based adjacency description of the grid plus explicit coordinates of the start and end, for example: “Grid 12×12. Neighbors of (r,c): … . Start at (3,1), end at (12,10).” The image input is an $n\times n$ grayscale, or binary, grid rendered as PNG, with black walls, white open corridors, a green start cell, and a red end cell.

The target text answer is the ordered solution path encoded as a sequence of global coordinates

$\bigl\{(r_1,c_1),(r_2,c_2),\dots,(r_T,c_T)\bigr\},$

packaged inside $\langle\mathrm{answer}\rangle$ tags, for example

$6\times 6$0

The target image is the same maze grid with the correct solution path overlaid in blue. The text and image outputs therefore represent the same latent solution in two synchronized modalities.

An auxiliary supervision signal is also included. An LLM, Gemini 3 Flash, is prompted to produce a “parallel” reasoning trace that first identifies large dead-end regions, major corridors, and global bottlenecks, then commits the coordinate sequence, mirroring the joint text–image decoding style of SC-CMJP. This natural-language chain-of-thought is placed inside $6\times 6$1. The presence of this trace indicates that the corpus is not restricted to answer supervision alone; it also encodes a specific global-to-local reasoning decomposition.

4. Dataset composition and benchmark partitions

The corpus and its held-out benchmark are organized as follows (Le et al., 14 Jul 2026).

Component	Size	Grid sizes
Training set	200,000 examples	$6\times 6$2
JMaze-Test500	500 additional mazes	$6\times 6$3
ID subset	Part of Test500	$6\times 6$4
OOD subset	Part of Test500	$6\times 6$5

No explicit validation set is mentioned; models are fine-tuned on the full 200,000 examples and evaluated on Test500. The grid-size distribution in training is uniform over $6\times 6$6, so each size has exactly $6\times 6$7 examples in training. By construction, each maze has exactly one simple path from start to end, so there is no class imbalance in the sense of multi-solution versus single-solution instances.

The split design is notable because the test set spans both smaller and larger mazes than those seen during training. The out-of-distribution subset therefore probes extrapolation at the level of grid size rather than only random seed variation within a fixed support.

5. Evaluation protocol and exact-match criteria

JMaze-200K uses exact-match evaluation centered on joint correctness across modalities (Le et al., 14 Jul 2026). The primary metric is joint accuracy:

$6\times 6$8

A sample is counted as correct only if both modalities exactly match the ground truth. The text sequence is parsed by a tolerant extractor, while the image is compared either pixel-wise or via a verifier model.

Separate text accuracy, $6\times 6$9, and image accuracy, $20\times 20$0, are also reported as diagnostics. Metrics are computed on the in-distribution and out-of-distribution subsets of Test500, as well as overall. No path-length error or edit distance is reported; only exact-match accuracy is used, because each path is unique.

A frequent misunderstanding in maze-oriented evaluation is to treat near-miss paths or partially correct renderings as informative continuous outcomes. JMaze-200K does not do this. Its benchmark definition is deliberately discrete: correctness is established only when the symbolic path and the visual overlay both coincide with the unique ground truth.

6. Integration into SC-CMJP and reported results

For experiments on JMaze-200K, the authors start from a unified text–image Masked Diffusion Model, Lumina-DiMOO, and fine-tune it for 200,000 steps with batch size 512 and learning rate $20\times 20$1 (Le et al., 14 Jul 2026). At inference, the reported CO sampler scores text positions by their self-confidence $20\times 20$2, while image positions are scored by a mixture

$20\times 20$3

with

$20\times 20$4

where $20\times 20$5 is the mean token entropy in modality $20\times 20$6.

The sampler also applies a birth–death, unmask–remask jump process with quotas

$20\times 20$7

and

$20\times 20$8

where $20\times 20$9 is the length of modality $n$0, $n$1, and $n$2 is the remasking ratio. In the broader SC-CMJP framework, one modality’s transition rates are functionals of the other modality’s confidence score as weighted by cross-modal attention, and a remasking jump retracts commitments when cross-modal evidence turns against them.

On JMaze-Test500, Figure 1 (right) shows $n$3 as a function of the number of function evaluations. Only the CO sampler yields a monotonically increasing curve, rising from approximately $n$4 at $n$5 to $n$6 at $n$7, whereas uncoupled baselines plateau or degrade. In Table 2, CO reaches $n$8 ID accuracy versus $n$9 for the best baseline, $n\times n$0 OOD accuracy versus $n\times n$1 for the runner-up MDM, and $n\times n$2 overall joint accuracy compared with $n\times n$3 for the next best method. Within the paper’s interpretation, these results confirm that cross-modal coupling and self-correction improve both in-distribution and generalization performance.

The benchmark’s empirical role is therefore tightly coupled to the paper’s methodological claim: if text and image branches can revise one another during denoising, exact joint correctness on a unique-solution reasoning task improves. This suggests that JMaze-200K is especially useful for evaluating multimodal samplers whose objective is not merely parallel decoding, but synchronized commitment and retraction across modalities.