JNono-200K: Multimodal Nonogram Benchmark

• JNono-200K is a synthetic multimodal nonogram corpus combining empty grid images and encoded clues to generate both solved puzzles and structured textual answers.
• It employs joint formulation with bidirectional constraint propagation, remasking, and cross-modal coupling to ensure consistency with algorithmic ground truth.
• The dataset comprises 200K puzzles with varied grid sizes and rigorous quality control measures verifying unique solutions and enforcing complex row/column constraints.

JNono-200K is a synthetic multimodal nonogram corpus introduced as one of three joint multimodal generation datasets in “Concurrent Image Understanding and Generation: Self-Correcting Coupled Markov Jump Processes” (Le et al., 14 Jul 2026). It is designed for concurrent image understanding and generation in a setting where textual and visual outputs are logically coupled and jointly verifiable. The task is nonogram solving: the input comprises an empty $N \times N$ grid image annotated with row and column clues together with a textual puzzle specification, and the target comprises both a solved grid image and a structured textual answer describing the filled-cell ranges per row. The corpus is intended to stress generation, editing, and reasoning simultaneously, with supervision written in a style “native to parallel decoding” rather than autoregressive (Le et al., 14 Jul 2026).

1. Task definition and representational structure

JNono-200K targets visual reasoning via Nonogram, also known as Picross, in a regime where image and text are “logically interlocked and jointly verifiable against algorithmic ground truth” (Le et al., 14 Jul 2026). The task couples two input modalities—an empty grid image with marginal clue numbers, and a text prompt encoding puzzle size and all clues—with two output modalities: a solved grid image and a textual answer specifying filled-cell ranges per row.

The dataset formulation is explicitly joint rather than auxiliary. The image output must satisfy all row and column constraints, while the text output must decode to the same binary solution. The paper states that a thinking trace is included “to encourage reasoning style aligned with parallel decoding,” and that for nonogram specifically the trace emphasizes bidirectional row/column constraint propagation instead of row-by-row solving (Le et al., 14 Jul 2026). This makes the corpus a benchmark არა only for symbolic correctness, but also for whether a multimodal sampler can maintain cross-modal consistency over an iterative denoising trajectory.

The record schema is shared with the paper’s other corpora. Typical fields are prompt, source_image, target_image, structured_understanding, and thinking_trace. Row clues are represented as lists of positive integers for each row, and column clues analogously for each column. In the image modality, cells are rendered as black-filled versus empty; in the text modality, the solution is encoded as ranges or exact indices (Le et al., 14 Jul 2026).

2. Construction, scale, and quality control

JNono-200K contains 200k synthetic nonogram puzzles (Le et al., 14 Jul 2026). In the training corpus, grid sizes are uniformly sampled from $\{5,\ldots,25\}$. Puzzle generation uses “a mixture of pattern synthesizers biased toward multi-run row/column clues,” a design choice intended to increase logical complexity by creating more opportunities for bidirectional constraint propagation (Le et al., 14 Jul 2026).

The generation pipeline is fully specified at a high level. The source image is an empty $N \times N$ grid rendered with row clues on the left margin and column clues at the top margin. The target image is the same grid with cells filled black to satisfy all clues. The paper states that the solution is guaranteed unique and produced algorithmically. Text fields include a puzzle specification with grid size and all row/column clues, a thinking trace synthesized by an oracle Gemini 3 Flash conditioned on the source and target images together with the ground-truth solution, and a structured answer containing per-row filled-cell ranges (Le et al., 14 Jul 2026).

Quality control is grounded in algorithmic verification. The paper states that algorithmic ground-truth generation ensures puzzle validity and unique solutions, and that both the textual answer and the rendered solution path are independently verifiable against the algorithmic ground truth (Le et al., 14 Jul 2026). This architecture sharply limits annotation ambiguity relative to open-ended visual reasoning tasks. A plausible implication is that JNono-200K is designed to isolate reasoning and multimodal consistency failures rather than label-noise failures.

3. Formal constraint system and alignment with SC-CMJP

The nonogram target is formalized on a binary grid $X \in \{0,1\}^{H \times W}$. For row $i$ with clues $R_i = (r_{i,1},\ldots,r_{i,K_i})$, the row must contain exactly $K_i$ runs of consecutive $1$’s with lengths $r_{i,k}$ separated by at least one zero except at boundaries; columns obey the analogous condition for $C_j$ (Le et al., 14 Jul 2026). The paper writes: $\{5,\ldots,25\}$0 and

$\{5,\ldots,25\}$1

Overall feasibility is

$\{5,\ldots,25\}$2

The paper also gives the standard necessary capacity bounds

$\{5,\ldots,25\}$3

These definitions make the target space exact and fully checkable (Le et al., 14 Jul 2026).

Within the SC-CMJP framework, the same paper associates JNono-200K with cross-modal attention, an entropy-based gate $\{5,\ldots,25\}$4, and remasking-based self-correction (Le et al., 14 Jul 2026). The mechanisms emphasized as relevant are shared percentile rank space, entropy-based gating, and self-correction via remasking. The paper’s nonogram-specific interpretation is that remasking is useful when locally confident image fills later become globally inconsistent with row/column constraints. This suggests that JNono-200K functions as a stress test for whether multimodal samplers can retract commitments rather than merely refine them.

4. Evaluation protocol and benchmark splits

Evaluation is performed on a held-out benchmark, JNono-Test500, spanning grid sizes $\{5,\ldots,25\}$5 (Le et al., 14 Jul 2026). The benchmark is divided into in-distribution and out-of-distribution subsets. In-distribution comprises sizes $\{5,\ldots,25\}$6, matching the training regime, while OOD comprises $\{5,\ldots,25\}$7 (Le et al., 14 Jul 2026).

The headline metric is joint accuracy. A sample is counted as correct only when both the text answer and the generated image are correct. The text answer is parsed by “a tolerant Gemini extractor,” and the image is judged “side-by-side against algorithmic ground truth” by Gemini 3 Flash (Le et al., 14 Jul 2026). The paper states that text accuracy and image accuracy are also computed separately, but headline reporting uses joint correctness only. No partial credit or tie-breaking is used.

Success criteria are strict in both modalities. For the image, correctness means cell-by-cell agreement with the algorithmic solution and satisfaction of all clues. For the text, correctness means that the per-row filled ranges decode to the same binary grid as the algorithmic solution (Le et al., 14 Jul 2026). A caveat noted in the paper is that the exact implementation details of tolerance thresholds in the Gemini-based evaluation are not enumerated.

5. Baselines and reported performance

The paper compares four samplers on JNono-Test500: MDM, ReMDM, MMaDA-Parallel, and CO (Self-COrrecting COupled Jump) (Le et al., 14 Jul 2026). MDM is described as independent across modalities with no remasking; ReMDM adds remasking but remains independent across modalities; MMaDA-Parallel interleaves modality updates across steps but still lacks within-step cross-modal feedback; CO combines cross-modal coupling with remasking (Le et al., 14 Jul 2026).

Method	Split	Joint accuracy
MDM	In-Distribution	0.110
ReMDM	In-Distribution	0.002
MMaDA-Parallel	In-Distribution	0.143
CO (Ours)	In-Distribution	0.167
MDM	OOD	0.050
ReMDM	OOD	0.000
MMaDA-Parallel	OOD	0.113
CO (Ours)	OOD	0.175
MDM	Total	0.100
ReMDM	Total	0.002
MMaDA-Parallel	Total	0.138
CO (Ours)	Total	0.168

The paper reports that CO is best across ID, OOD, and Total, and that the advantage is largest on OOD, where it exceeds the runner-up by $\{5,\ldots,25\}$8 (Le et al., 14 Jul 2026). It also reports that MDM generalizes poorly OOD on nonogram, dropping by $\{5,\ldots,25\}$9 from ID to OOD, while MMaDA drops by $N \times N$0; CO instead rises slightly on OOD. ReMDM yields near-zero joint correctness, which the paper interprets as evidence that remasking without cross-modal coupling is insufficient for bidirectional constraint reasoning (Le et al., 14 Jul 2026).

These results are specific to joint accuracy rather than unimodal correctness. As reported, the benchmark is therefore less a measure of isolated image synthesis or text generation than of whether both modalities converge to the same valid nonogram solution.

6. Position within the benchmark suite, use cases, and naming

JNono-200K is introduced alongside JEdit-1M and JMaze-200K (Le et al., 14 Jul 2026). The paper distinguishes JNono-200K from JMaze-200K by the structure of its constraints: mazes encode a single path through walls from start to end, whereas nonograms encode many local constraints that must be satisfied simultaneously. The paper further states that the nonogram setting especially benefits from remasking under cross-modal contradictions, because late-stage fills can invalidate previously confident cells (Le et al., 14 Jul 2026).

The intended use is training and evaluation of joint multimodal generation and understanding, especially for visual reasoning tasks with algorithmic ground truth. The project page is listed as https://coupled-jump.github.io, and the paper states that model checkpoints, code, and datasets are planned for release to the community; licensing is not specified in the paper (Le et al., 14 Jul 2026).

A plausible source of confusion is the similarity of the string “JNono-200K” to unrelated JUNO terminology. That resemblance is incidental. Separate JUNO detector papers concern a 20 kiloton liquid scintillator experiment, and one of them explicitly states that JUNO is not a 200 kiloton detector (He, 2014). In current arXiv usage relevant to multimodal reasoning, JNono-200K denotes the 200k-example nonogram corpus introduced in 2026 (Le et al., 14 Jul 2026).