acceptodds
Under review as a conference paper at ICLR 2027

PuzRian: Scalable Puzzle Reasoning for Vision-Language-Action Agents

Abstract

Vision-language models (VLMs) have shown remarkable progress as interactive agents. However, systematically evaluating their long-horizon spatial reasoning remains challenging. We introduce PuzRian, a scalable Gymnasium benchmark for interactive visual-spatial reasoning, featuring 3 task families with controllable complexity, action spaces, feedback, and observability. Evaluating 9 frontier and open-source VLMs reveals strong performance on symbolic tasks but substantial degradation as spatial and perceptual complexity increases, with near-zero success on challenging 6-piece puzzles. Controlled analysis further reveals that increasing interaction budgets alone provides limited benefit, while fine-grained feedback enables more effective error correction. We further find that performance is particularly sensitive to required spatial manipulations rather than the size of the action space, and that models become increasingly brittle at recovering from imperfect intermediate states as task complexity grows. PuzRian provides a controlled testbed for diagnosing how vision-language agents plan, use feedback, and recover from errors during long-horizon interaction.

Primary area:
datasets and benchmarks
Keywords:
agent evaluation, benchmarking, multi-modal agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7cysOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents PuzRian, a Gymnasium-based benchmark for testing visual-spatial reasoning in vision-language agents. It includes three types of puzzles with different levels of complexity and interaction. Nine models are evaluated, with results showing good performance on simple symbolic puzzles but significant difficulties with irregular shapes and visual fragments. The authors also examine how feedback, interaction limits, spatial transformations and error recovery affect performance.

Critical strengths
  1. Well-designed and flexible benchmark.

The main strength of PuzRian is its flexibility. The three puzzle types cover different levels of spatial complexity, while the configurable settings allow a closer look at how agents reason and interact. The process-level metrics are particularly useful for understanding failures beyond final success rates. The availability of the code is another positive aspect.

  1. Interesting experimental findings.

The evaluation of nine models highlights the gap between simple spatial tasks and more challenging geometric problems. The controlled experiments are also useful, particularly the finding that rotations and reflections cause more difficulty than simply increasing the number of available actions. The results on feedback and error recovery further show that agents often struggle to correct mistakes, even when given more opportunities to interact.

Critical weaknesses
  1. Handling of alternative valid solutions.

My main concern is how the benchmark handles puzzles with multiple valid solutions. As the authors acknowledge, Irregular Assembly may allow different correct arrangements, yet piece-level feedback appears to rely on a particular solution. This could incorrectly penalise valid placements and affect the feedback and recovery metrics. The authors should clarify how often this happens and how alternative solutions are handled.

  1. Limited discrimination at higher difficulty levels.

All nine models achieve zero success on Irregular-Hard, making it difficult to compare their capabilities. Additional intermediate difficulty levels would make the benchmark more informative. I also think the conclusions about the causes of failure could be more cautious, as perception and planning are difficult to separate. Finally, the ablation experiments use relatively few instances, and uncertainty estimates would help assess the reliability of the results.

Questions for authors

How does the benchmark handle multiple valid solutions in Irregular Assembly? Could these affect piece-level feedback and recovery metrics?

Could the authors include intermediate difficulty levels to better distinguish model performance, particularly for Irregular Assembly?

Could the authors report confidence intervals for the ablation results to show how consistent the observed differences are?

1 comment

We thank the reviewer for their positive feedback and constructive suggestions.

  1. Multiple valid solutions in Irregular Assembly

We clarify that episode success accepts any feasible reconstruction, without requiring the model to match the sampled ground-truth solution. For piece-level feedback, although our environment supports comparing placements against the sampled ground truth, we intentionally use a less restrictive setting in the reported Irregular experiments. The feedback only checks whether each piece is entirely contained within the target shape, avoiding incorrectly penalizing potentially valid alternative placements. For process-level metrics such as Piece Accuracy and Recovery Rate, we allow swapping pieces with identical shapes rather than requiring exact piece-ID matching. We also analyzed solution multiplicity on the same evaluation instances. For Irregular-Medium, 86% of instances have a unique solution and 13% have two solutions. For Irregular-Hard, 58% have a unique solution and 18% have two solutions. We will clarify the feedback design and this limitation in the revised paper. We have clarified the success criterion and piece-level feedback design for Irregular Assembly in 3.2, and added a discussion of the limitations of reference-based process metrics in 3.1.

  1. Intermediate difficulty levels for Irregular Assembly

We thank the reviewer for this suggestion. We agree that 0 success rates on Irregular-Hard make it difficult to distinguish model capabilities. Although our process-level metrics provide some additional information, more fine-grained difficulty levels would be useful. Due to computational costs, we have not systematically evaluated all intermediate settings. We plan to include additional experiments with 30 instances per setting, covering 5×5 and 6×6 puzzles with and without required rotation, to better investigate performance across difficulty levels.

  1. Confidence intervals for the ablation results

We thank the reviewer for this suggestion. We will include Standard Error of the Mean, which is p^(1−p^)n−1\sqrt{\frac{\hat{p}(1-\hat{p})}{n-1}} in the revised paper. Here we present in ablation Effect of Additional Actions as an example. As shown in the table, the performance of GPT-5.4 and Gemini-3.1 drops substantially when spatial transformations are required, consistent with our main observation.

SettingGPT-5.4Gemini-3.1Gemini-2.5
Baseline0.900 ± 0.0550.833 ± 0.0680.167 ± 0.068
Flip-10.500 ± 0.0910.333 ± 0.0860.000 ± 0.000
Flip-20.167 ± 0.0680.067 ± 0.0460.000 ± 0.000
Rotate0.033 ± 0.0330.067 ± 0.0460.000 ± 0.000
Reviewer_XCuxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

PuzRian is a configurable benchmark for interactive spatial reasoning. It includes three puzzle families and evaluates nine vision-language models. Performance drops sharply on harder tasks. The benchmark is useful, but several analyses do not isolate the causes of failure.

Critical strengths
  1. The paper studies a relevant problem: how agents plan, use feedback, and recover from errors.
  2. The environment supports varied task difficulty, actions, and observations.
  3. The evaluation covers multiple models and measures both final success and intermediate progress.
Critical weaknesses
  1. The contribution beyond VisGym, VSP, and Jigsaw-R1 is not clearly established.
  2. Irregular puzzles can have multiple valid solutions, as acknowledged in Appendix C.1. Feedback based on one reference solution may mislabel valid placements. This raises concerns about the feedback and recovery analyses.
  3. The budget study uses different puzzle sizes for global and piece-level feedback. It therefore does not isolate how feedback affects the value of additional turns.
  4. The interactive baseline uses repeated model calls. The non-interactive baseline uses one. Without matching total inference resources, the gains cannot be attributed to feedback alone.
  5. Many ablations use only 30 instances and report no uncertainty estimates. Small differences are difficult to assess.
Questions for authors
  1. How do terminal success, piece-level feedback, and recovery metrics handle alternative valid solutions?
  2. Does the feedback–budget finding hold on the same puzzle instances with matched budgets?
  3. Do the gains from interaction remain when total reasoning resources are matched?
  4. Section 5.3.2 describes correctly pre-placed pieces, while Appendix C.7 includes incorrect placements. Which setting produced Figure 12?
  5. What evidence distinguishes spatial planning failures from errors in coordinate handling, rotation, or state tracking?
4 comments
  • Novelty compared with existing benchmarks

We would like to clarify that PuzRian is not just another collection of visual puzzles, but a configurable benchmark specifically designed to evaluate complex interactive spatial reasoning and investigate the limitations of current agents.

Compared with VSP and Jigsaw-R1: VSP mainly provides game-like settings for studying visual perception and spatial planning, and Jigsaw-R1 focuses on visual reinforcement learning and generalization. Both benchmarks primarily focus on generating solutions rather than interacting with an environment. In contrast, PuzRian provides an interactive environment that requires agents to iteratively manipulate puzzle pieces, observe updated states, and adjust their actions based on feedback. Also, we introduce rich geometric manipulation and compositional spatial reasoning, including irregular shape reconstruction, multi-piece arrangements, rotation, and flipping. These tasks require agents to combine multiple spatial reasoning skills rather than simply predict piece ordering or generate a sequence of movements.

Compared with VisGym: In VisGym, puzzle-solving tasks are just only a small subset. It mainly focuses on broader evaluation of visual interaction capabilities. In contrast, PuzRian is specifically differ on evaluating spatial recognition and reasoning, with tailored variations that test different levels of abstraction and spatial reasoning complexity. The related tasks in VisGym mainly involve relatively simple operations, such as piece placement and edge matching, with limited requirements for complex geometric transformations or combining multiple spatial skills. In contrast, we introduce tasks that require agents to reason about geometric relationships, plan multi-piece arrangements, and perform transformations such as rotation and flipping. Moreover, PuzRian provides fine-grained diagnostic analysis to understand why agents succeed or fail, instead of just the overall success rate. Beyond final success rates, we systematically investigate the effects of puzzle complexity, required versus redundant spatial transformations, feedback granularity, interaction budgets, and observation access, as well as agents' ability to recover from mistakes. These controlled variations allow us to identify specific challenges in interactive spatial reasoning rather than only measuring overall task performance.

  • Sample size and uncertainty

We thank the reviewer for this concern. Due to the computational and time cost of evaluating multiple models across a large number of configurations, we limited most ablation experiments to 30 instances per setting, with a single run per instance. Our goal is primarily to identify general performance trends and understand how different factors affect model behavior. We will provide a Standard Error of the Mean, which is p^(1−p^)n−1\sqrt{\frac{\hat{p}(1-\hat{p})}{n-1}} in the revised paper. Here we present in ablation Effect of Additional Actions as an example. As shown in the table, the performance of GPT-5.4 and Gemini-3.1 drops substantially when spatial transformations are required, consistent with our main observation.

SettingGPT-5.4Gemini-3.1Gemini-2.5
Baseline0.900 ± 0.0550.833 ± 0.0680.167 ± 0.068
Flip-10.500 ± 0.0910.333 ± 0.0860.000 ± 0.000
Flip-20.167 ± 0.0680.067 ± 0.0460.000 ± 0.000
Rotate0.033 ± 0.0330.067 ± 0.0460.000 ± 0.000

Q1. Multiple valid solutions and correctness evaluation

We want to clarify that we have already considered multiple valid solutions in the design, and the issue does not affect the success evaluation or piece-level feedback used in our experiments. For Irregular Assembly, episode success does not require matching the sampled ground-truth solution, and any feasible reconstruction is accepted. For piece-level feedback, although our environment supports comparing placements with the sampled ground truth, we intentionally use a less restrictive setting in the reported experiments. The feedback only checks whether each piece is entirely inside the target shape, without requiring it to match the reference position or orientation. For process-level metrics such as Piece Accuracy and Recovery Rate, we allow swapping pieces with identical shapes rather than requiring exact piece-ID matching. We also analyze number of solutions . For Irregular-Medium, 86% of instances have a unique solution and 13% have two solutions. For Irregular-Hard, 58% have a unique solution and 18% have two solutions. So most Medium instances have only one solution. We acknowledge that reference-based process metrics cannot perfectly handle all valid solutions, since it is difficult to determine which solution the model intends to reach. We clarified the success criterion and piece-level feedback design for Irregular Assembly in 3.2 , and added a discussion of the limitations of reference-based process metrics in 3.1.

Q2. Matched feedback and interaction budgets

Our original experiments used different puzzle sizes because global feedback already resulted in near-zero success on the 5×5 setting, while larger settings generally had 0 success. In contrast, piece-level feedback allowed us to investigate how performance changes with additional interaction steps on larger puzzles. To address this, we conducted additional experiments on the same 5×5 Visual-Medium puzzles with rotation, comparing global and piece-level feedback under matched interaction budgets (15, 20, and 25 turns). These matched results support that additional interaction steps are more beneficial with informative feedback. We have included this in Appendix C.1 and will add more models results.

Feedback15 turns20 turns25 turns
Global0.0670.0330.067
Piece-level0.6330.9330.900

Q3. Interaction gains and inference resources In our non-interactive baseline, the model generates the complete action sequence in a single response, while the interactive setting allows sequential execution with updated observations and feedback. To avoid restricting non-interactive reasoning, we set the maximum output token budget to 65,535, and none of the responses reached this limit. Strictly matching total inference resources is challenging, since interactive solving naturally has additional reasoning and replanning based on newly observed states and feedbacks. Non-interactive tends to achieve lower performance, suggesting that the value of sequential interaction. It is also consistent with our experiments showing that more informative piece-level feedback often improves performance.

Q4. Partial-state experiment settings

Thanks for pointing out this. We clarify that all pre-placed pieces in the reported partial-state experiments are initially at their correct ground-truth positions and orientations. Specifically, we use pre-generated puzzles and manually place (k) pieces at their ground-truth configurations, leaving the remaining (N-k) pieces unplaced. The pre-placed pieces can still be moved by the model during interaction. Although our environment supports incorrect initial placements, this option was not used in these experiments. In Figure 12, the number of unplaced pieces refers to those initially not on the playground. Since all pre-placed pieces are correct, there are no initial placement errors that require correction. We have clarified the partial-state initialization procedure in 5.3.2 and revised the corresponding experimental description in Appendix C.7.

Q5. Distinguishing spatial planning failures from other errors

While it is difficult to fully disentangle different failure sources, PuzRian provides several analysis to investigate them. First, our text-representation experiments provide explicit grid-cell information and achieve higher performance, means that visual interpretation contributes to the difficulty. Second, our action-space ablations show that spatial transformations such as rotation and flipping do introduce additional challenges. We also track invalid actions, including rejected placements, movements, rotations, and out-of-bounds actions. For example, Gemini-3 produces 17/0 invalid actions on Irregular/Visual-Medium, increasing to 80/8 on Hard settings. This indicates that it’s more difficult when complexity increases. Our process-level metrics further reveal difficulties in preserving correct placements and recovering from mistakes.

Thank you for the clarifications and additional experiments. These address several of my concerns, particularly the feedback rules and partial-state setup. The matched-budget experiment is also a useful addition.

However, my original concerns are not fully resolved. The distinction from existing benchmarks remains insufficiently established. Reference-based process metrics still have limitations on puzzles with multiple valid solutions. The interaction comparison also leaves inference resources unmatched, so the gains cannot be attributed to feedback alone.

Related to my original concern about the contribution, I still wonder what distinctive insight this benchmark would retain if the next model generation largely solved the current configurations. This is not a requirement to predict future progress, but a reason to clarify its diagnostic contribution beyond documenting current failures.

I appreciate the authors’ efforts, but these remaining concerns lead me to retain my original rating.

Thanks Reviewer XCux for sharing the thoughts! We believe there are necessary details that we need to make clear that would clarify these concerns.

  1. Regarding multiple valid solutions and process metrics

We acknowledge that reference-based process metrics have limitations when multiple valid solutions exist. However, these metrics are designed to provide meaningful diagnostic signals for models as our testing environment proves to be very hard for most of the existing models with 0 end task success rates. The difficulty also scales steeply with our configuration with even just one or two additional grid size and pieces. Furthermore, these metrics also remain meaningful for instances with unique solutions, which account for the majority of most of our subtasks, such as Irregular-Medium cases (86% with a unique solution, see our statistics above), and are particularly useful for Visual tasks.

  1. Regarding inference resources

We want to clarify that the interactive versus non-interactive comparison is intended to evaluate the benefit of interactive solving, rather than isolate the effect of feedback alone. The interactive setting naturally involves sequential observations, feedback, and replanning, on the other hand the non-interactive baseline generates a complete solution in a single response. We also allow up to 65,535 output tokens for non-interactive reasoning, but none of the responses reached this limit. While this does not strictly match total inference resources, the comparison remains meaningful for evaluating the overall interactive solving paradigm.

  1. Regarding the benchmark's long-term contribution

We agree that future, more capable models may solve the current configurations, but this does not mean interactive spatial reasoning is fully solved. Our environment is designed to be scalable, allowing puzzle complexity to increase to tens or potentially hundreds of pieces easily. Even as models improve, evaluating performance under increasing complexity and limited inference budgets remains meaningful. For Visual Assembly, analysis of action efficiency, spatial transformations, and error recovery remain valuable as complexity increases. For Irregular Assembly, larger puzzles may introduce more valid solutions, limiting the applicability of reference-based process metrics. However, the harder planning problem itself with native terminal success and global feedback remain valid for evaluating spatial reasoning, even when multiple solutions exist.

Reviewer_nErEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PuzRian, a Gymnasium-based benchmark for interactive spatial reasoning, with three task families: assembling colored rectangles, irregular shapes, and natural-image fragments. The authors generate solvable instances through an exact-cover formulation and evaluate nine vision-language models while varying puzzle size, available actions, feedback granularity, and observation access. Table 2 shows that GPT-5.4 and Gemini-3.1 perform nearly perfectly on the symbolic tasks, but both have zero success on Irregular-Hard, which requires rotation. On Visual-Hard, piece-level feedback raises their success rates from zero to 16% and 26%, respectively. The authors further attribute failures to difficulties with spatial transformations, feedback use, and error recovery. The experiments directly support a clear capability gap on the tested puzzle configurations, although the independent effects of these factors need more careful interpretation.

Critical strengths
  1. The benchmark offers a controlled setting for studying interactive spatial reasoning.** The three task families share an action interface while allowing changes to geometric complexity, visual content, feedback, and observations. Solvable puzzles are generated through exact cover (Section 3, Table 1, and Appendix B.1). This design makes it practical to investigate specific capabilities, which is the benchmark's main value as a research tool.

  2. The spatial-transformation experiments offer useful diagnostics beyond a leaderboard.** Figures 5–6 and Appendix C.2 separately test required and redundant actions. In the Visual setting with a fixed 5×5 grid and five pieces, requiring rotation substantially reduces performance, while merely adding actions that do not need to be used has a relatively smaller effect. This provides focused evidence that the difficulty is not simply due to having more action choices, and shows how the environment can help formulate and test specific hypotheses.

Critical weaknesses
  1. It is unclear whether feedback and process metrics match the task objective when multiple solutions are valid, which affects the credibility of the diagnostic findings. [Needs clarification]** The Irregular prompt only asks the agent to cover the target shape correctly; it does not require a unique arrangement of pieces (Appendix E, page 28). Appendix C.1 also explicitly acknowledges that these tasks can have multiple valid solutions, making piece-level feedback less reliable. However, Edit Distance in Appendix D uses reference positions and rotations, and the paper does not explain how Piece Accuracy, Recovery Rate, and piece-level feedback handle other valid arrangements. If they only recognize the reference solution used during generation, they could mark valid progress as wrong or even steer the model away from a feasible solution. This would directly weaken the interpretation of feedback effects and error recovery. It does not, by itself, establish that the success criterion in Table 2 has the same problem.

  2. Some central explanations are not supported by sufficiently matched comparisons. [Confirmed limitation of the evidence]** For the claim that piece-level feedback makes extra interaction budget more useful, the global-feedback budget experiment in Figure 4 and Table 5 uses a 5×5 grid with five pieces, while the piece-level experiments use 6×6 grids with six pieces and 7×7 grids with seven pieces. This makes it difficult to attribute the difference to feedback alone. For reasoning from partial states, the main text says the pre-placed pieces are correct, but Appendix C.7 includes incorrect pieces. If Figure 12 does require additional error correction, having the same number of unplaced pieces does not mean having the same amount of work left. These issues do not invalidate the observed performance differences, but they weaken a key part of the paper's value as a benchmark for controlled diagnosis.

Questions for authors
  1. How do the success criterion, piece-level feedback, and process metrics handle multiple valid solutions in Irregular?** Please clarify whether arrangements that differ from the generated reference solution but satisfy the task requirements are accepted, and whether valid intermediate arrangements can receive negative feedback. Showing that these mechanisms account for multiple solutions, or that any resulting bias does not change the main findings, would substantially address the first concern. If feedback does mislead models, the affected conclusions would need to be reconsidered.

  2. Are there budget-scaling results for both feedback types using the same puzzle instances, sizes, and budgets?** This would help establish whether the benefit of extra interaction really depends on feedback granularity. Without such a matched comparison, it seems more appropriate to report the trends within each configuration separately rather than attribute their difference to feedback type.

  3. Which correct, incorrect, and unplaced pieces are present in the initial states used in Figure 12?** Can pre-placed pieces be moved, and does the comparison with solving from an empty board match the remaining required actions and rotation demands? If this is only a wording inconsistency and the actual controls are adequate, the concern can be resolved. If the experiment also introduces extra correction work, it does not sufficiently support the current explanation of the difficulty of reasoning from partial states.

1 comment

We thank the reviewer for raising these concerns, and we clarify as follows.

  1. Multiple valid solutions and correctness evaluation

We want to clarify that our experiments already account for multiple valid solutions, and the main experimental results are not affected. Irregular Assembly does not require matching exactly the sampled ground-truth decomposition. Episode success is determined by whether the final state forms a valid reconstruction, so any feasible solution is accepted. For piece-level feedback, although our environment supports comparing each piece against the sampled ground truth, we intentionally do not use this option in the reported experiments, as multiple valid solutions may exist. Instead, the feedback only indicates whether all cells of each placed piece are entirely contained within the target shape, without requiring it to match its reference position or orientation. This avoids incorrectly penalizing potentially valid alternative placements, although it provides less precise guidance than the reference-based feedback used in Visual Assembly. This also helps explain why piece-level feedback provides smaller improvements on Irregular Assembly than on Visual Assembly.

For process-level metrics such as Piece Accuracy and Recovery Rate, we account for interchangeable pieces with identical shapes rather than requiring exact piece-ID matching. We additionally analyzed solution multiplicity on the same evaluation instances. For Irregular-Medium, 86% of instances have a unique solution and 13% have two solutions, after accounting for interchangeable identical shapes. For Irregular-Hard, 58% have a unique solution and 18% have two solutions. Therefore, most instances in the Medium setting have a unique solution. For cases with multiple solutions, process metrics may not perfectly reflect the model's progress, since the model's intended solution cannot be determined from intermediate placements alone. We clarified the success criterion and piece-level feedback design for Irregular Assembly in 3.2, and added a discussion of the limitations of reference-based process metrics in 3.1.

  1. Matched comparison of feedback and interaction budgets

Thanks for the suggestion. Our original experiments used different puzzle sizes because performance under global feedback was already close to 0 in the 5×5 setting, and larger puzzles generally resulted in 0 success, making it difficult to observe meaningful trends. In contrast, piece-level feedback allowed us to better investigate how performance changes with additional interaction steps. To address this concern, we conducted additional experiments using the same 5×5, 5-piece setting with rotation, comparing global and piece-level feedback under the same puzzle instances and interaction budgets (15, 20, and 25 turns) as following. For GPT-5.4, performance with global feedback is close to 0 across different interaction budgets as reported in main paper while with piece-level feedback, success increases from 0.633 at 15 turns to 0.933 at 20 turns, and remaining similar at 25 turns. These results suggest that additional interaction steps are more beneficial when informative feedback is available, rather than simply increasing the interaction budget. We additionally included the current comparison results in Appendix C.1 of the revised version, and we will further add other models results to provide further evidence for the importance of informative feedback under matched interaction budgets.

Feedback15 turns20 turns25 turns
Global0.0670.0330.067
Piece-level0.6330.9330.900
  1. Partial-state experiment design and interpretation

We thank the reviewer for pointing out this ambiguity. We clarify that, in the reported partial-state experiments, all pre-placed pieces are initially at their correct ground-truth positions and orientations based on sample solution, consistent with the description in the main paper. Specifically, we start from pre-generated puzzles and manually place (k) pieces according to the ground-truth configurations, leaving the remaining (N-k) pieces unplaced. The pre-placed pieces are not fixed and can still be moved by the model during interaction. Although our environment supports generating partial states with both correct and incorrect placements, this option was not used in the reported experiments. We clarified the actual setting in Section 5.3.2 and corrected the description in Appendix C.7.

In Figure 12, the number of unplaced pieces refers to the pieces that are initially not on the playground. For example, in a 5 piece puzzle with 2 correctly pre-placed pieces, the number of unplaced pieces is 3. Since all pre-placed pieces are initially correct, the experiments do not introduce additional correction requirements from incorrect initial placements. And under piece-level feedback, the model will receive information indicating that these pre-placed pieces are correct.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.