acceptodds
Under review as a conference paper at ICLR 2027

xRRMs: Explorative Recurrent Reasoning Models

Abstract

Recurrent Reasoning Models (RRMs), including Hierarchical Reasoning Models (HRM) and Tiny Recursive Models (TRM), achieve strong performance on combinatorial reasoning tasks such as Sudoku while remaining parameter efficient. HRM and TRM are in principle looped transformers in which a shared module is applied repeatedly to refine a neural state. Conceptually, this computation is grounded in a fixed-point equation, using deep supervision and stop-gradients between iterations. Recently, stochastic variants of RRMs, such as Equilibrium Reasoners (EqR) and Probabilistic Tiny Recursive Models (PTRM), have been proposed. Beyond enabling generative modeling, stochasticity provides a natural mechanism for improving exploration of the solution space. In particular, these variants induce stochasticity by randomizing the initial latent states and by injecting noise during recurrent iterations. However, their gains rely on exploring multiple trajectories at inference, which raises the question of how to encourage exploration more directly. We address this with xRRM, a Recurrent Reasoning Model trained with strategies from Explorative Modeling: multiple sampled neural states are refined in parallel. At each deep supervision step, the loss is evaluated for all sampled trajectories, but gradients are computed only from the trajectory whose output is closest to the correct solution. This effectively implements a winner-takes-all objective and restricts backpropagation to the best-matching trajectory. On Sudoku-Extreme, xRRM achieves an average fully solved rate of more than 90% with a single evaluation candidate, improving over EqR by more than 20 percentage points. Test-time scaling with 32 evaluation candidates increases this value to more than 99%. We extend our approach to generative problem settings and evaluate it on the Game-of-24 benchmark, where multiple valid solutions per problem exist, investigating whether effective exploration extends beyond deterministic problem solving.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
recurrent reasoning models, combinatorial reasoning problems, exploration, test-time-scaling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.80 over 5 reviews, above 85% of the others.
Reviewer_eR22OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose xRRM, a recurrent reasoning model trained using parallel stochastic trajectories, learned latent perturbations, and computing the gradients using only the candidate closest to a valid solution. Evaluations on Sudoku-Extreme, Pencil Puzzle Bench, and Game-of-24 show improved accuracy and test-time scaling.

Critical strengths
  1. At approximately matched training FLOPs, xRRM outperforms the second-best method (URM) by 10 points in single-candidate accuracy on Sudoku-extreme. This substantial improvement is the main reason for my positive assessment.

  2. Five-run evaluations strengthen the empirical evidence, supporting the practical value of this simple training strategy.

Critical weaknesses

1 The experiments do not isolate the contribution of learned noise from that of winner-takes-all training, making it difficult to assess the individual contribution of each component.

Questions for authors
  1. How does xRRM perform with a fixed noise scale, keeping the number of training candidates and the winner-takes-all objective unchanged?
2 comments

Thanks for the clear suggestion to assess the individual contribution of each component.

W1, Q1: Isolating effects

We ran the suggested control (fixed noise scale) proposed and added it to Table 2 (Sudoku-Extreme), Table 3 (Pencil Puzzle Bench) and Table 4 (Game-of-24):

  • Fixed noise (Fixed noise): Winner-takes-all training with a fixed perturbation scale instead of the learned one.

Following the requests of reviewer qsbA, GY5o, and, SmEt we also added further ablations to the same tables:

  • Mean over candidates (Mean K): The task loss is averaged over all candidates instead of taking the minimum.

  • Backprop through a random candidate (Random K)

On Sudoku-Extreme with a single evaluation candidate, xRRM reaches 90.2 ± 1.8%, compared with 86.5 ± 6.3% for fixed noise, 69.1 ± 4.5% for the mean loss, and, 70.8 ± 2.8% for the random-candidate control.

The ablations show that in overall selecting the candidate closest to a solution seems to contribute most to xRRM's performance. An exception to this is Pencil Puzzle Bench, were the learned noise scale seems to be equally important.

I thank the authors for their response. I have decided to maintain my positive rating.

Reviewer_SmEtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces xRRM, a training framework for recurrent reasoning models that aims to improve exploration of the latent solution space. Instead of optimizing a single stochastic trajectory, xRRM maintains multiple parallel latent trajectories during training and applies a winner-takes-all objective, backpropagating the task loss only through the candidate closest to a valid solution. The method also learns the perturbation scale and trains a halting/correctness head over all candidates for test-time selection. Compared with prior works, xRRM show significant performance improvement and relatively moderate average performance improvement on all baselines.

Critical strengths
  • The paper is overall well-written.

  • The proposed method is simple but seemingly effective.

  • Under the matched-FLOPs setting, the performance gain on Sudoku is significant.

Critical weaknesses
  • The perforamnce gain is a bit overclaimed in the abstract. According to Table 3, the average performance gain is less significant than that on Sudoku.

  • Ablation study is missing. It remains unclear whether the empirical advantage comes from generating more trajectories or from WTA loss. For example, how does WTA compare empirically with a simple average loss?

Questions for authors
  • Although the WTA objective is differentiable whenever the minimizing candidate is unique, its gradient changes discontinuously when the winning candidate switches. Will this cause instability during training?

  • How does WTA compare empirically with a simple average loss?

2 comments

W1: Perforamnce gain is a bit overclaimed

a) Performance gain in the abstract

Thank you for pointing this out. We agree that the abstract should not give the impression that gains of this size hold on every benchmark. We have revised it in two ways: it now states that the improvement is largest on Sudoku-Extreme, and it makes explicit that these numbers refer to the attention variant.

We would like to keep the comparison with EqR in the abstract, as we consider it our main comparison. EqR is the closest prior method, and the single-candidate setting compares one trajectory with one trajectory, so the difference reflects the training procedure rather than additional test-time compute.

b) Table 3: Pencil Puzzle Bench

We agree that the gains on Pencil Puzzle Bench are smaller. We would like to point out, however, that URM, which achieves the best single-candidate result there, has to be considered separately. URM has twice as many attention layers, executes more steps, and has an additional convolution in the MLP. The convolution introduces a locality prior that can help more or less depending on the puzzle type on Pencil Puzzle Bench. The comparison with URM is therefore the least informative one about the effect of our training procedure.

W2, Q2: Ablation study missing/Comparison to Average loss

We ran the suggested controls proposed and added it to Table 2 (Sudoku-Extreme), Table 3 (Pencil Puzzle Bench) and Table 4 (Game-of-24):

  • Mean over candidates (Mean K): The task loss is averaged over all candidates instead of taking the minimum.

Following the requests of reviewers qsbA, GY5o, and, eR22, we also added a further ablation to the same tables:

  • Fixed noise (Fixed noise): Winner-takes-all training with a fixed perturbation scale instead of the learned one.

  • Backprop through a random candidate (Random K)

On Sudoku-Extreme with a single evaluation candidate, xRRM reaches 90.2 ± 1.8%, while the mean loss achieves 69.1 ± 4.5%.

The ablations show that in overall selecting the candidate closest to a solution seems to contribute most to xRRM's performance. An exception to this is Pencil Puzzle Bench, were the learned noise scale seems to be equally important. We did not observe a major difference between backpropagating the mean of all candidates and random candidate selection.

Q1: Gradient discontinuity

The winner-takes-all loss is the minimum over finitely many smooth functions. It is continuous and piecewise smooth, and it is non-differentiable only where two candidates attain exactly the same loss, which generically is a set of measure zero. This is the same situation as for ReLU at 0 or max-pooling. Moreover, the noise is resampled at every step, so training optimizes the expected minimum.

For problems with multiple valid solutions, the same argument holds for the exact minimum over candidates and targets (SelectWinner in Algorithm 2). EfficientSelectWinner is a heuristic approximation that is cheaper to compute.

Empirically, we did not observe instabilities during training. The variation across the five runs is not larger than for the baselines: on Sudoku-Extreme, the standard deviation is 1.8 for xRRM, compared with 9.2 for PTRM and 3.8 for EqR.

Thank you for the detailed rebuttal and additional experiments.

Most of my concerns are solved. I maintain my positive rating of 3.

Reviewer_GY5oOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes xRRM, which trains recurrent reasoning models using the winner-takes-all objective of Explorative Modeling. During training, K stochastic trajectories are rolled out in parallel; each has a random initial state and per-step perturbations whose scale is learned. The halting head is trained on all candidates and is used to pick a candidate at inference. Training FLOPs are matched to the baselines and results are averaged over five seeds/splits. On Sudoku-Extreme, xRRM reaches a 90.2% fully-solved rate with a single evaluation candidate (vs. 67.8% for EqR and 61.4% for TRM) and 99.1% with 16–32 candidates. It also improves on Pencil Puzzle Bench with 32 candidates (85.2% vs. 81.0% for EqR) and on Game-of-24 (12.4% vs. ≈4–5% for the baselines with one candidate).

Critical strengths
  1. WTA over parallel trajectories is easy to add to existing RRMs. On Sudoku-Extreme the improvement is large even with a single evaluation candidate. This holds under matched training FLOPs, five seeds on different train/test splits, and K_train selected on a held-out part of the training data. The K_train = 1 ablation suggests the gain comes from multi-candidate training rather than from the learned noise alone.
  2. The paper gives full results for every K_train and for alternative multi-solution objectives. Its limitations section is candid about sensitivity to K_train, the small Pencil test set, and the weak halting head on Game-of-24.
Critical weaknesses
  1. K trajectories with the loss averaged over all candidates instead of the min, and WTA with a fixed noise scale as in EqR. Without them it is unclear whether the gain comes from winner selection, from training on more trajectories per sample, or from the learned perturbation.
  2. Under FLOP matching, the batch size drops from 768 (TRM) to 132 at K_train = 8 and to 38 at K_train = 32 (Table C1). Performance peaks at intermediate K and collapses for large K (Sudoku 90.2 → 54.8%, Pencil 73.4 → 34.8%). It is therefore unclear whether this behavior comes from the objective or from the shrinking batch. On Game-of-24, K_train ≤ 4 is no better than the baselines (3.1–3.8%).
  3. The scope is limited and the baselines may be weak. Only Sudoku-Extreme shows a clear win. On Pencil Puzzle Bench (61 test puzzles), xRRM is below URM with one candidate (73.4 vs. 76.1%), and with 32 candidates it leads EqR by 4.2 pp with stds of 2.0–3.4.
  4. Only the attention variant of TRM is used. Its Sudoku-Extreme result is below the original paper's (App. C.1), and the MLP variant, which App. C.1 notes is stronger on Sudoku, is excluded.
Questions for authors
  1. Should add (a) K trajectories with the mean loss over candidates, (b) WTA with a fixed noise scale, and (c) xRRM vs. TRM/EqR at the same batch size? This would separate winner selection, learned noise, and the batch-size effect of FLOP matching.
  2. In Table 4, the EqR rows are identical under q-value selection and under the functional evaluator, and they are flat in K_eval. Does this mean EqR's candidates are effectively identical?
  3. How does xRRM compare with, or apply to, the TRM-MLP variant? Do the results transfer to Maze-Hard or ARC-AGI?
1 comment

We want to thank the reviewer for the detailed review including the suggested experiments.

W1, Q1(a), Q1(b): Isolating effects

We ran the suggested controls proposed and added it to Table 2 (Sudoku-Extreme), Table 3 (Pencil Puzzle Bench) and Table 4 (Game-of-24):

  • Mean over candidates (Mean K): The task loss is averaged over all candidates instead of taking the minimum.
  • Fixed noise (Fixed noise): Winner-takes-all training with a fixed perturbation scale instead of the learned one.

Following the requests of reviewer qsbA, we also added further ablations to the same tables:

  • Backprop through a random candidate (Random K)

On Sudoku-Extreme with a single evaluation candidate, xRRM reaches 90.2 ± 1.8%, compared with 70.8 ± 2.8% for the random-candidate control, 69.1 ± 4.5% for the mean loss and 86.5 ± 6.3% for fixed noise.

The ablations show that in overall selecting the candidate closest to a solution seems to contribute most to xRRM's performance. An exception to this is Pencil Puzzle Bench, were the learned noise scale seems to be equally important. We did not observe a major difference between backpropagating the mean of all candidates and random candidate selection.

W2, Q1(c)

a) Effect of the batch size

We ran batch-size ablations on Sudoku-Extreme (see new Table C5), with the number of training steps kept fixed.

Baselines at the batch sizes of xRRM: Smaller batches hurt the baselines substantially: TRM drops from 61.4% (batch size 768) to 34.4% (132) and 17.8% (38), and EqR from 67.8% (768) to 27.7% (132) and 16.1% (38). At the batch size of 132, xRRM-8 reaches 90.2%. FLOP matching therefore puts xRRM at a disadvantage with respect to batch size rather than favouring it.

Change of xRRM batch size: We trained xRRM-32 with the batch size of xRRM-8 (K=8, batch size 132 instead of 38, at about 3.5 times the training FLOPs). With a single evaluation candidate it improves from 54.8% to 67.9%, but stays well below xRRM-8 (K=8) at the same batch size (90.2%). An explanation for this decrease in single-candidate accuracy with increasing KK might be with many training candidates only one of them has to be correct.

b) Note on Game-of-24

We agree that xRRM does not improve over the baselines for K<=4 on this task, and we now state this more explicitly in the text.

W3: Scope and strength of the results

We agree that the clear win is on Sudoku-Extreme and that the gains on Pencil Puzzle Bench are smaller.

Regarding URM: The baselines are well known RRMs and are often applied on these kind of datasets. Our intention was to improve on a known architecture, to be better able to disentangle model architecture and our contribution. Therefore, we used the same backbone as TRM, EqR and PTRM. Nevertheless, we also included URM, which has twice as many attention layers, executes more steps, and has an additional convolution in the MLP. The convolution introduces a locality prior that can help more or less depending on the puzzle type on Pencil Puzzle Bench. The comparison with URM is therefore the least informative one about the effect of our training procedure.

W4, Q3: MLP-based token mixing instead of attention

We have added the MLP-mixer variant of TRM, PTRM, EqR and xRRM on Sudoku-Extreme (see new Table C6). With a single evaluation candidate, EqR reaches 84.7 ± 1.3%, PTRM reaches 86.3 ± 1.3%, and, xRRM reaches 94.0 ± 1.1% with a single candidate. The improvement of xRRM therefore transfers to the MLP-mixer variant, although the margin is smaller than for the attention variant because the baselines start from a higher level.

Q2: Identical EqR rows in Table 4

Thanks for spotting this. The EqR row in the block with the functional evaluator was indeed a copy error, which we have now corrected. The corrected values do not change the qualitative conclusions regarding the comparison with the other methods.

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

This work proposes xRRM, a step-level best-of-N training method for recurrent reasoning models (RRMs). The method uses an approximation to score intermediate RRM steps, selects the best candidate from multiple samples, and updates the parameters using cross-entropy loss against the ground truth. The idea is to drive each step of the RRM toward the ground truth under the best context and perturbation found during rollout, making the update more on-policy. The paper reports experiments on Sudoku-Extreme, Pencil Puzzle Bench, and Game-of-24, showing the effectiveness of the method.

Critical strengths

The motivation is interesting. The method selects the best context and perturbation by rolling out a pool of candidates and choosing the one with the lowest loss.

The experiments show that the method is effective. On Sudoku-Extreme, xRRM solves 90.2% of puzzles with a single test-time candidate, compared with 67.8% for EqR.

Critical weaknesses

The paper does not clearly explain the method’s novelty relative to previous work such as PTRM, EqR, and Forward XM. This is especially true of Forward XM, as the authors acknowledge that the best-of-N idea comes from it. This makes it hard to judge the novelty of this work.

Although the authors acknowledge that the best-of-N idea comes from Forward XM, the experiments do not include a performance comparison with Forward XM.

There is no ablation study on the key contribution of this work, the best-of-K algorithm. The ablation on K alone cannot establish the importance of the algorithm.

Questions for authors

See the weaknesses above.

1 comment

Novelty relative to PTRM, EqR and Forward XM

xRRM applies Explorative Modeling (Gladstone et al., 2026) to recurrent reasoning models. The best-of-K principle itself is older than XM and goes back to multiple choice learning (Guzman-Rivera et al., 2012; Lee et al., 2016), which we cite in the introduction and in the remark before Proposition 1.

The differences to the prior RRM variants are:

  • PTRM is trained deterministically and adds noise only at inference.
  • EqR trains a single stochastic trajectory per input with a fixed noise scale; several trajectories are used at inference.
  • xRRM trains a pool of K stochastic trajectories per input with a learned noise scale. At every deep supervision step only the lowest-loss candidate receives the task gradient, while the halting head is trained on all candidates.

We have added a short paragraph to Section 2.2 (Relation to Prior Work) that makes this more clearly. We thank the reviewer for making it clear to us, that this was not well described.

Our main contribution therefore is to show that this combination already improves individual trajectories, e.g., on Sudoku-Extreme, xRRM reaches 90.2% with a single evaluation candidate, compared with 67.8% for EqR.

Comparison with Forward XM

Forward XM is a training procedure rather than a model or an architecture for these benchmarks; it needs an underlying architecture for predicting solutions to be applied to. xRRM is Forward XM applied to RRMs, so the xRRM results are the Forward XM results in this setting. The comparison that isolates its effect is against the same architecture without winner selection, which we have now added (see below).

Ablation of the best-of-K selection

We agree that the sweep over K alone does not isolate the selection rule.

In response to this comment and to related suggestions by the other reviewers, we have added three ablations to Table 2 (Sudoku-Extreme), Table 3 (Pencil Puzzle Bench), and, Table 4 (Game-of-24):

  • Backprop through a random candidate (Random K)
  • Mean over candidates (Mean K)
  • Fixed noise (Fixed noise)

Random K differs from xRRM only in the selection. On Sudoku-Extreme with a single evaluation candidate it reaches 70.8 ± 2.8%, compared with 90.2 ± 1.8% for xRRM, which shows the importance of selecting the lowest-loss candidate.

Reviewer_qsbAOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces xRRM, a training approach for stochastic recurrent reasoning models. Multiple latent trajectories evolve in parallel, and at each deep-supervision step only the candidate with the lowest task loss contributes to the task gradient. The model also learns state-dependent perturbations and trains a correctness head for candidate selection. Experiments on Sudoku-Extreme, Pencil Puzzle Bench, and Game-of-24 examine single-candidate performance and inference-time scaling. The main finding is that training with multiple candidates substantially improves performance even when only one candidate is used at inference.

Critical strengths
  1. The single-candidate Sudoku result is substantial: xRRM achieves (90.2±1.8(90.2\pm1.8%), compared with (67.8±3.8(67.8\pm3.8%) for EqR. Since no inference-time candidate selection is involved, this provides clear evidence that the training procedure improves the individual solver.

  2. The experiments make a meaningful effort to control training compute. Batch sizes compensate for the additional forward passes, and Table C1 reports closely matched measured FLOPs. Results are averaged over five runs, and the full training-candidate sweep is provided rather than only the best configuration.

  3. The paper reports limitations candidly. Performance deteriorates at large training-candidate counts, and the Game-of-24 results clearly distinguish the learned selector from symbolic verification. The latter comparison usefully identifies candidate selection as a major bottleneck.

Critical weaknesses
  1. The contribution of winner-takes-all supervision to the proposed exploration mechanism is not sufficiently isolated.

The comparison with EqR changes both the supervision rule and the perturbation parameterization, while compute matching also changes the batch size. The candidate-count sweep therefore supports the complete training recipe, but does not isolate the explanation that removing direct supervision from losing trajectories enables better exploration. A useful control would retain the same candidate pool, learned perturbations, batch size, and halting policy, but backpropagate through a randomly selected candidate instead of the lowest-loss candidate. This would preserve the single backward pass while directly testing the selection rule. Figure 2 also counts changes to previously incorrect digits without distinguishing corrections from changes to another incorrect value. Reporting these transitions separately would make the evidence for productive exploration more informative.

  1. The generative extension demonstrates solution finding, but not coverage of multiple valid solutions.

The introduction motivates sampling from the full set of valid solutions. However, the objective (min⁡k,mL(Y^k,Ym))(\min_{k,m}L(\hat{Y}_k,Y_m)) can be minimized even if every candidate converges to the same valid solution for a given input. The Game-of-24 fully solved rate does not distinguish this behavior from generating several different valid expressions. In particular, the 69.8% result with symbolic selection demonstrates that the candidate pool often contains a valid answer, but does not establish broader solution coverage. Reporting the number of distinct valid solutions per problem, with a clear treatment of equivalent expressions, would directly test this aspect of exploration. Otherwise, the generative claim should be narrowed to improved success at finding a valid solution.

Questions for authors

See Critical Weaknesses above.

1 comment

W1:

a) Isolating the contribution of winner-takes-all supervision

Thank you for this precise suggestion. We ran the control proposed and added it to Table 2 (Sudoku-Extreme), Table 3 (Pencil Puzzle Bench) and Table 4 (Game-of-24):

  • Backprop through a random candidate (Random K)

Following the requests of reviewers GY5o, SmEt, and, eR22, we also added two further ablations to the same tables:

  • Mean over candidates (Mean K): The task loss is averaged over all candidates instead of taking the minimum.
  • Fixed noise (Fixed noise): Winner-takes-all training with a fixed perturbation scale instead of the learned one.

The ablations show that in overall selecting the candidate closest to a solution seems to contribute most to xRRM's performance (with the exception of Pencil Puzzle Bench, were the learned noise scale seems to be equally important). On Sudoku-Extreme with a single evaluation candidate, xRRM reaches 90.2 ± 1.8%, compared with 70.8 ± 2.8% for the random-candidate control, 69.1 ± 4.5% for the mean loss and 86.5 ± 6.3% for fixed noise. Since the random-candidate control differs from xRRM only in the selection rule, this isolates lowest-loss selection as the main source of the gain.

b) Figure 2: direction of the changes

Figure 2 was intended to show how much xRRM explores compared to the other models. To also show the direction of these changes, the new Figure C1 separates them into "Correct to incorrect" and "Incorrect to correct". The direction in which xRRM changes its predictions closely follows its performance on the task.

W2: Coverage of multiple valid solutions

We agree, that the objective min⁡k,mL(Y^k,Ym)\min_{k,m} L(\hat{Y}_k, Y_m) only requires the closest solution and does not reward different candidates for reaching different solutions.

To gain further insight, we measured coverage directly, as suggested. The exact formula, we used, is given in Appendix C.3. Across all tested values of K, xRRM reaches a coverage of up to about 30% (at K=16), see Table C2. Note that in practice K would have to be selected on validation data. In comparison, EqR and PTRM seem to be below 5%.

Overall, the comment raised an important point. In response, we have toned down the claims about the generative capabilities of our method throughout the paper and state this explicitly in the limitations. The coverage results nevertheless suggest that xRRM may be better able than EqR and PTRM to find multiple valid solutions.

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback and the very helpful suggestions for further experiments. We respond to each review individually and summarize the main changes here.

The main concern raised by most reviewers was that the effects of winner selection, training on multiple trajectories, and, the learned noise scale were not sufficiently separated. We therefore added three ablations of xRRM to Tables 2, 3 and 4.

Further additions:

  • Batch-size ablations for TRM, EqR and xRRM on Sudoku-Extreme (Table C5)
  • Results for the MLP-mixer variant of TRM, PTRM, EqR and xRRM on Sudoku-Extreme (Table C6)
  • Solution coverage on Game-of-24 (Appendix C3)
  • A breakdown of the prediction changes by direction (Figure C1)
  • Revised claims on generative problem settings, a paragraph on the relation to prior work, and an extended limitations section
  • A slightly revised abstract, based on the points raised by the reviewers
  • Minor changes throughout the manuscript to integrate the new results and correct typos

For the reviewers' convenience, we have highlighted the main changes in the revised manuscript in blue. Minor edits such as corrected typos are not always explicitly marked.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.