acceptodds
Under review as a conference paper at ICLR 2027

Loopholing All the Way: Language Modeling with pure Self-Conditioning

Abstract

Recent work proposes continuous flow language models (FLMs) to address the limitations of discrete diffusion models (DDMs): at every step, DDMs sample tokens independently, which hurts few-step generation, and discard the uncertainty of the denoiser's predictions (*information collapse*). Empirically, on reasoning tasks such as Sudoku and GSM8K, FLMs at best match standard DDMs and fall well behind DDMs with self-conditioning (SC). In particular, Loopholing DDMs (LDDMs) mitigate information collapse by propagating an SC state across steps, along the discrete tokens. While SC is traditionally seen as an auxiliary heuristic to improve performance, its empirical success makes us wonder whether **we could define a continuous generative model based purely on self-conditioning?** We answer affirmatively with LAW (“Loopholing All the Way”), which propagates the *expected embedding* under the denoiser's predictions across steps, decoding the tokens once, at the end. On hard Sudoku, LAW comes close to LDDMs (90.4% vs. 92.7%), and on Maze, it is slightly more accurate than all baselines with few steps. On TinyGSM, fine-tuning a pre-trained Duo checkpoint with LAW increases the GSM8K accuracy from 18.6% to 55.2%. On OpenWebText, trained from scratch, LAW matches or improves over S-FLM with the standard DiT architecture.

Primary area:
generative models
Keywords:
discrete diffusion, diffusion language models, continuous latent models, loopholing, self-conditioning, uniform-state diffusion, few-step generation, non-autoregressive generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qq65OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Loopholing All the Way (LAW), a continuous generative model based entirely on self-conditioning. Unlike existing methods, LAW propagates token embeddings across iterations and decodes discrete tokens only at the final step. The effectiveness of the proposed method is validated on Sudoku, Maze, GSM8K, and OpenWebText, and ablation studies show the effectiveness of key design choices. LAW shows strong few-step performance on several reasoning tasks, particularly when fine-tuning a pretrained Duo model on TinyGSM.

Critical strengths

The principal contribution of this paper, in my view, is that it shows continuous self-conditioning alone can support competitive generation without intermediate discrete sampling. This provides a novel perspective on the role of discrete sampling in diffusion language models (DLMs), which is also shown to be compatible with pre-trained checkpoints. The effectiveness is also shown for few-step generation experiments on multiple datasets.

Critical weaknesses

The major concern is that the explanation of the key performance gains remains limited. The chosen design is validated by experimental results and ablations, but it remains unclear whether the performance gains come from removing sampling noise or changing how inputs are combined. An explanation of why the advantage is strong after fine-tuning but not when training from scratch is also expected. The lack of such intuitive discussion limits the understanding of why pure self-conditioning is beneficial.

From my perspective, it is intuitive that LDDM might learn to rely mainly on its continuous state when discrete tokens are unhelpful. What prevents it from recovering LAW-like behavior in some cases?

Questions for authors

Please refer to the 'Critical Weaknesses' section and address the concern.

2 comments

We thank the reviewer for their positive review and for emphasizing the novelty and effectiveness of our method across different experimental settings.

General Remark: Motivated by the suggestion of the reviewer XATP that "fewer training-sequence presentations" could contribute to the gap between LDDM [1] and LAW, we found that a small adjustment to the training procedure significantly improves LAW's performance, without changing the model architecture or the method itself. In particular, we explore a training procedure closer to LDDM. With probability pp, we unroll M=2M=2 passes and apply the loss to the last pass only, rather than averaging over all passes. Otherwise, we perform a single pass, as in uniform diffusion. Since we stop gradients through the earlier passes, this allows us to train LAW for the same number of steps as LDDM-U at matched FLOPs. Additionally, with this modification in place, we observed that using cross-entropy (CE) gave more diverse samples on OpenWebText.

These minor adjustments lead to substantial improvements. When trained from scratch on TinyGSM, LAW now solves 42.2% of GSM8K problems (T=1T=1, 1024 NFE), compared to 34.5% for LDDM-U. On OpenWebText (64 NFE), LAW now achieves lower Gen. PPL than the diffusion and flow baselines at matched entropy (Fig. 24, Suppl. D.4).

We provide full details on this variant in Suppl. D of the updated manuscript.

We address the reviewer's concerns below.

Do gains come from how the inputs are combined?

Yes. With the Expected Embedding state, the GSM8K accuracy peaks at 38.5% when the state is added to the discrete tokens, as in LDDMs, vs. 53.4% when it replaces them, as in LAW (256 NFE, T=1T=1; Fig. 7).

Do gains come from removing the sampling noise?

Yes. To isolate the sampling noise, we introduce two new baselines that replace the Expected Embedding in LAW by either:

  1. the embedding of the argmax token (argmax),
  2. the embedding of a token sampled from the prediction (sampled).

As in the submission, we fine-tune the Duo checkpoint on TinyGSM for 50k steps with batch size 64. We find that

  • with the Expected Embedding, LAW reaches the highest peak accuracy,
  • with the Expected Embedding, LAW learns faster than the two new baselines.

Suppl. E.3 shows the GSM8K accuracy of all variants with 256 NFE at T=1T=1, without regularization (Fig. 26a) and with temperature regularization (ρ=4\rho=4; Fig. 26b).

Why is LAW stronger after fine-tuning compared to training from scratch?

It is not entirely clear why. The gap is not explained by the difference in training duration: we evaluated the LAW checkpoints trained from scratch every 10k steps, and the accuracy plateaus after ~70k steps (out of 192k; Fig. 23). The Duo initialization might help LAW converge to a better solution, possibly thanks to the representations learned during pre-training. Understanding why is an interesting question for future work.

Why do LDDMs not learn to rely on their continuous state when the tokens are unhelpful, and recover LAW-like behavior?

We found that LDDMs rely mainly on the discrete tokens. When we set the token embeddings of LDDM-U and LDDM-M to zero during sampling, they solve almost none of the GSM8k problems. With a zeroed SC state, the accuracy drops by half or more but remains well above 0% (Fig. 25). We see three reasons, now in Suppl. A.3:

  • Training signal: LDDMs train on corrupted data zt∼qt(⋅∣x)\mathbf{z}_t \sim q_t(\cdot \mid \mathbf{x}), whose tokens are informative, so the loss rewards relying on them.
  • Architecture: LDDMs feed the backbone the sum ztEθ+ct\mathbf{z}_t E_\theta + \mathbf{c}_t. While LDDMs could in principle learn to use the SC state alone, neural networks struggle to ignore input features (e.g. here by setting all token embeddings to zero), even when some features are pure noise [2, 3, 4, 5]. It is therefore not surprising that LDDMs rely on the discrete tokens and do not converge to the same solution as LAW.
  • State map: LAW with the LDDM state map, (ΦθHid.,replace)(\Phi^{\text{Hid.}}_\theta, \text{replace}), solve close to 0% of the problems, vs. 53.4% with the expected embedding (256 NFE; Fig. 7). One observation is that with the Expected Embedding, the input of the denoiser remains in the convex hull of the Embeddings. There are no such guarantees when using the final layer hidden state.

[1] Loopholing Discrete Diffusion: Deterministic Bypass of the Sampling Wall, Jo et al., ICLR 2026.

[2] Why Do Tree-Based Models Still Outperform Deep Learning on Typical Tabular Data?, Grinsztajn et al., NeurIPS 2022.

[3] The Data Efficiency of Deep Learning Is Degraded by Unnecessary Input Dimensions, D'Amario et al., Frontiers in Computational Neuroscience 2022.

[4] Large Language Models Can Be Easily Distracted by Irrelevant Context, Shi et al., ICML 2023.

[5] Feature Selection, L1 vs. L2 Regularization, and Rotational Invariance, Ng, ICML 2004.

Dear reviewer,

We wanted to follow up and ask if our response has addressed your concerns. Otherwise, we would be happy to address any remaining questions and reach a consensus before the discussion period ends.

Thank you for your time!

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

LAW generates sequences by repeatedly refining expected token embeddings and decoding discrete tokens only at the end. It achieves competitive results on Sudoku and Maze and substantially improves GSM8K accuracy when fine-tuned from Duo. Also results for OpenWebText are presented

Critical strengths
  1. A simple method that tests a clear research question: can self-conditioning support generation on its own.

  2. Evaluation across several tasks, with informative ablations.

  3. Transparent reporting of limitations, including weaker results when training from scratch and reduced text diversity.

Critical weaknesses
  1. My main concern is the fairness of the fine-tuning comparison. LAW is fine-tuned from a pretrained Duo checkpoint, whereas the LDDM baselines are trained from scratch. This does not provide a fair, controlled comparison of the adaptation methods. In the compute-matched from-scratch experiments, LAW underperforms both LDDM variants.I think in order to make fair comparison authors should test whether LAW retains its advantage under matched fine-tuning conditions, particularly against LDDM-U adapted from the same Duo checkpoint, using the same data, comparable compute, and the same model selection protocol.

  2. Text quality is assessed mainly through generative perplexity and unigram entropy on OWT, which can miss repetition and semantic errors. Please include additional metrics, such as MAUVE and Gradient Moment [1], to provide a broader assessment of generation quality and diversity. The handling of end-of-text tokens also needs clarification.

[1] Beyond Single Tokens: Distilling Discrete Diffusion Models via Discrete MMD

Questions for authors
  1. Could you compare LAW with LDDM-U fine-tuned from the same Duo checkpoint, using the same data and training budget?

  2. Could you add MAUVE, GM, and repetition metrics to better assess text quality and diversity?

2 comments

We thank the reviewer for the positive review of our work, and we are glad that the reviewer appreciated our evaluation across several tasks, and found our ablations informative.

General Remark: Motivated by the suggestion of the reviewer XATP that "fewer training-sequence presentations" could contribute to the gap between LDDM [1] and LAW, we found that a small adjustment to the training procedure significantly improves LAW's performance, without changing the model architecture or the method itself. In particular, we explore a training procedure closer to LDDM. With probability pp, we unroll M=2M=2 passes and apply the loss to the last pass only, rather than averaging over all passes. Otherwise, we perform a single pass, as in uniform diffusion. Since we stop gradients through the earlier passes, this allows us to train LAW for the same number of steps as LDDM-U at matched FLOPs. Additionally, with this modification in place, we observed that using cross-entropy (CE) gave more diverse samples on OpenWebText.

These minor adjustments lead to substantial improvements. When trained from scratch on TinyGSM, LAW now solves 42.2% of GSM8K problems (T=1T=1, 1024 NFE), compared to 34.5% for LDDM-U. On OpenWebText (64 NFE), LAW now achieves lower Gen. PPL than the diffusion and flow baselines at matched entropy (Fig. 24, Suppl. D.4).

We provide full details on this variant in Suppl. D of the updated manuscript.

We answer the reviewer's comments below.

Is LAW better than LDDM-U when both are fine-tuned from the same Duo checkpoint for the same number of updates?

Yes. We fine-tuned a LDDM-U (p=0.9p=0.9) from a Duo checkpoint for 50k steps, matching the data and batch size (64, on a single H200) used to fine-tune LAW. On GSM8K (T=1T=1, 256 NFE), LDDM-U solves 19.2% after 20k steps and 31.0% after 50k, vs. 55.2% and 52.0% for LAW (ρ=4\rho=4), and 52.2% and 38.7% for LAW without regularization (Suppl. E.4, Fig. 26).

Can you evaluate unconditional generation with other metrics?

We thank the reviewer for the suggestion, which enriches our empirical protocol. We have extended our generative performance evaluations with two extra metrics:

  • We computed the Gradient Moment (GM) [2] on 5000 samples per model (T=1T=1, 1024 NFE). With the revised training (see general remark), only the autoregressive model has a lower GM than LAW.
  • To also measure the diversity across samples, we have computed the nn-gram overlap, the share of the distinct 4-grams of a sample that occur in 5 other randomly selected samples. The nn-gram overlap of LAW is close to that of AR.
ModelGM (lower is better)Overlap-4 (lower is better)
OWT (data)0.001 ± 0.0040.68 ± 0.01
AR0.150 ± 0.0050.69 ± 0.01
MDLM0.883 ± 0.0070.73 ± 0.01
Duo0.849 ± 0.0080.79 ± 0.01
FLM1.200 ± 0.0142.47 ± 0.01
S-FLM (S-arch.)1.632 ± 0.0172.03 ± 0.02
LDDM-M0.629 ± 0.0060.79 ± 0.01
LDDM-U0.658 ± 0.0070.84 ± 0.01
LAW0.565 ± 0.0080.70 ± 0.02

We prioritized GM and the nn-gram overlap over MAUVE, which prior work found unreliable in several ways, e.g.:

  • Insensitive to corruption: with GPT-2 features, replacing 10 tokens at the start or middle of each text with random tokens lowers MAUVE by only 1.3% and 6.5% [3].
  • Oversensitive to formatting: appending a newline to each text gives a MAUVE below 0.01 against the original texts [4].
  • High variance: with the same samples, changing only the random seed gives a MAUVE between 0.89 and 0.93 [5], a range comparable to the gaps between models.

How are end-of-text tokens handled?

We compute the Gen. PPL as in MDLM [6]: GPT-2 Large scores all tokens of the sample except the end-of-text tokens after the first one. The unigram entropy uses all tokens of the sample, end-of-text tokens included. We have updated Suppl. B.7 accordingly.

[1] Loopholing Discrete Diffusion: Deterministic Bypass of the Sampling Wall, Jo et al., ICLR 2026.

[2] Beyond Single Tokens: Distilling Discrete Diffusion Models via Discrete MMD, Hoogeboom et al., arXiv 2026.

[3] On the Blind Spots of Model-Based Evaluation Metrics for Text Generation, He et al., ACL 2023.

[4] MixCE: Training Autoregressive Language Models by Mixing Forward and Reverse Cross-Entropies, Zhang et al., ACL 2023.

[5] Promises, Outlooks and Challenges of Diffusion Language Modeling, Deschenaux and Gulcehre, arXiv 2024.

[6] Simple and Effective Masked Diffusion Language Models, Sahoo et al., NeurIPS 2024.

Reviewer_qWyZOpenReview ↗

Rebuttal Acknowledgement

Thank you for addressing my concerns. I hope that the additional experiments broaden and strengthen the paper’s empirical evaluation. I’m more confident in my initial assessment after reading your response, new results can be seen as positive enough that I continue to lean accept. Good luck.

Reviewer_XATPOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

LAW replaces token sampling between denoising steps with the expected token embedding under the model’s predictions. The model is trained to consume this continuous state and decodes tokens only at the end. Fine-tuning Duo gives strong GSM8K results with few model calls. Training from scratch is competitive on Sudoku and Maze, but trails LDDM on GSM8K. The OpenWebText results are weaker: LAW improves over Duo only at low entropy with 1,024 calls and otherwise has a worse quality–diversity tradeoff.

Critical strengths
  1. The few-step GSM8K result is strong. Fine-tuned LAW reaches 51.2% with 32 model calls, compared with LDDM-U’s 49.5% with 1,024 calls at its better reported temperature. The additional training compute to reach this checkpoint is approximately 2% of Duo’s pretraining compute. The competitive Sudoku results and small advantage over the strongest Maze baselines at 2–64 calls also show that this simple procedure is useful.

  2. Figure 7 provides a useful controlled comparison. Starting from the same Duo checkpoint, with the same objective and training budget and without regularization, expected-embedding replacement reaches a peak accuracy of 53.4%, compared with 38.5% when sampled-token inputs are added. This shows that the replacement operation helps independently of temperature regularization.

Critical weaknesses
  1. The OpenWebText results make me less convinced of LAW as a general approach to language generation. In Figure 5, at entropy around 5.3 and 1,024 calls, LAW’s Gen. PPL is roughly 80 versus 40 for Duo. At 64 calls, Duo is better throughout the overlapping entropy range. LAW only leads at low entropy with 1,024 calls, and the examples in Figures 19–20 contain substantial repetition. These results limit the evidence for the method beyond the reasoning tasks.

  2. It remains unclear how much of the gain comes from carrying a soft state. Figure 7 shows two things. Removing the sampled-token input helps (53.4% vs 38.5%), and the expected embedding works far better than the hidden state. But no variant carries a hard token alone, so a soft state is never compared with a hard one. This matters because the paper motivates LAW by its ability to retain the denoiser’s uncertainty. Models trained and evaluated with only an argmax-token or a sampled-token embedding would help separate the value of softness from simply avoiding sampling noise. The strongest GSM8K results also depend on Duo initialization. Trained from scratch at matched compute, LAW reaches 42.3% versus 49.5% for LDDM-U (T=0.1T=0.1, 1,024 calls). Fewer training-sequence presentations could contribute to this gap. On GSM8K, the strongest evidence is therefore for adapting a pretrained model.

Overall, I lean weak reject. The method is simple and the few-step reasoning results are valuable. But the paper presents LAW as a language model, and its OpenWebText results are clearly weaker than Duo’s. The contribution of the soft state is also not yet isolated. The hard-token controls could change my assessment, and a real-data reference in Figure 5 would help judge how much the generation gap matters.

Questions for authors
  1. Where does real OpenWebText lie in Figure 5 under your evaluation pipeline?

  2. Under the Figure 7 setup, how does accuracy change if the model is trained and evaluated with sampled-token or argmax-token embeddings as the sole carried state, with the rest of the setup held fixed?

3 comments

We thank the reviewer for their feedback and appreciate the comment highlighting the few-step performance of LAW.

General Remark: Motivated by the suggestion of the reviewer that "fewer training-sequence presentations" could contribute to the gap between LDDM [1] and LAW, we found that a small adjustment to the training procedure significantly improves LAW's performance, without changing the model architecture or the method itself. In particular, we explore a training procedure closer to LDDM. With probability pp, we unroll M=2M=2 passes and apply the loss to the last pass only, rather than averaging over all passes. Otherwise, we perform a single pass, as in uniform diffusion. Since we stop gradients through the earlier passes, this allows us to train LAW for the same number of steps as LDDM-U at matched FLOPs. Additionally, with this modification in place, we observed that using cross-entropy (CE) gave more diverse samples on OpenWebText.

These minor adjustments lead to substantial improvements. When trained from scratch on TinyGSM, LAW now solves 42.2% of GSM8K problems (T=1T=1, 1024 NFE), compared to 34.5% for LDDM-U. On OpenWebText (64 NFE), LAW now achieves lower Gen. PPL than the diffusion and flow baselines at matched entropy (Fig. 24, Suppl. D.4).

We provide full details on this variant in Suppl. D of the updated manuscript.

We now address the reviewer's concerns.

LAW underperforms Duo on OpenWebText. Can LAW operate as a general language model?

While the original LAW performed worse, with the improved training and 64 NFE, LAW has a lower Gen. PPL than the diffusion and flow baselines at matched entropy (see Fig. 24, Suppl. D.4 of the updated manuscript), and approaches AR. This demonstrates that LAW is not only applicable to reasoning tasks but also effective as a general language model, especially in the low NFE regime.

How much of the gain comes from the soft state? How does the accuracy change with a sampled or argmax token instead of the Expected Embedding?

We thank the reviewer for the suggestion. To answer this question, we introduce two new baselines for LAW, which replace the Expected Embedding, in training and sampling, by:

  • the embedding of the argmax token (argmax);
  • the embedding of a token sampled from the prediction (sampled).

As in the submission, we fine-tune the Duo checkpoint on TinyGSM for 50k steps with batch size 64. Suppl. E.3 shows the GSM8K accuracy of all variants with 256 NFE at T=1T=1, without regularization (Fig. 26a) and with temperature regularization (ρ=4\rho=4; Fig. 26b). We find that:

  • With the Expected Embedding, LAW reaches the highest peak accuracy, with and without regularization.
  • With the Expected Embedding, LAW learns faster: after 10k steps, 51.5% vs. 44.0% (sampled) and 41.9% (argmax) without regularization, and 50.9% vs. 39.4% (sampled) and 41.9% (argmax) with ρ=4\rho=4.

Where does OpenWebText lie in Fig. 5 under your evaluation pipeline?

We have included the (entropy,Gen. PPL)=(5.46,15.2)(\text{entropy}, \text{Gen. PPL}) = (5.46, 15.2) data point corresponding to the validation set of OpenWebText (last 100k documents of the dataset) in Fig. 5 of the updated manuscript.

Is the fine-tuned LAW better only because it received more updates (Duo pre-training + LAW fine-tuning)?

No, we do not believe that the fine-tuned LAW performs better than training from scratch because of additional updates. Indeed, we evaluated intermediate LAW checkpoints trained from scratch every 10k steps and found that the accuracy plateaus after ~70k steps (out of 192k, Fig. 23, Suppl. D.3). Understanding why Duo initialization performs better is an interesting question for future work. We thank the reviewer for this question, and now highlight this finding in the revised conclusion.

[1] Loopholing Discrete Diffusion: Deterministic Bypass of the Sampling Wall, Jo et al., ICLR 2026.

Dear reviewer,

We wanted to follow up and ask if our response has addressed your concerns. Otherwise, we would be happy to address any remaining questions and reach a consensus before the discussion period ends.

Thank you for your time!

Thank you for the additional experiments. These substantially address my concerns about the soft state and the OpenWebText results. Some limitations remain, but they no longer outweigh the strengths of the contribution. I am therefore moving to weak acceptance and have no further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Area Chair and Reviewers,

We are grateful for the detailed feedback and questions. Reviewers found the few-step reasoning results strong (XATP, qq65), the method simple with a clear research question (qWyZ), the ablations informative (XATP, qWyZ), the report of limitations transparent (qWyZ), and the paper clear (XATP, qWyZ, qq65). Reviewer qq65 also found that LAW offers a novel perspective on the role of discrete sampling in diffusion language models. Below, we summarize the main questions, our answers, and the changes to the manuscript, which are marked in blue in the PDF file.

Last-pass training (XATP, qWyZ, qq65). Following the suggestion of XATP, we introduced a small adjustement to the training procedure. On TinyGSM, LAW now beats LDDM-U when trained from scratch (42.2% vs 34.5%). On OpenWebText, with 64 NFEs, the LAW outperforms the diffusion and flow baselines in terms of the Gen. PPL / Entropy Pareto frontier. In the manuscript, we:

  • added Suppl. D, which describes last-pass training and its results on all tasks,
  • summarized these results in Sec. 3.2, Sec. 4 (including a new row in Table 1) and the conclusion.

Soft state and sampling noise (XATP, qq65). We fine-tune LAW from Duo with the embedding of the argmax token or of a sampled token as the carried state. The expected embedding still reaches the highest peak accuracy and learns faster.

Fine-tuning from Duo (XATP, qWyZ, qq65). We fine-tuned LDDM-U from the same Duo checkpoint as LAW, with the same data and number of steps. After 50k steps, LDDM-U solves only 31% of GSM8K problems, vs. 52% for LAW. We also show that the higher accuracy of fine-tuned LAW over LAW trained from scratch does not come from more training steps: trained from scratch, LAW plateaus after ~70k of its 192k steps (Fig. 23). We revised the conclusion accordingly.

Why LDDMs do not recover LAW-like behavior (qq65). We demonstrate that LDDMs rely mainly on the discrete token inputs. When we set the token embeddings to zero, LDDMs solve almost none of the tasks. During training, both the tokens and the SC state are informative, so the model learns to use both. In principle, LDDMs could recover LAW by ignoring the token inputs, but neural networks generally struggle to ignore part of their input, even if pure noise (Suppl. A.3).

Further metrics on OpenWebText (XATP, qWyZ). We added the OpenWebText validation set to Fig. 5 (Gen. PPL 15.2, unigram entropy 5.46). We also report the Gradient Moment and a 4-gram overlap across samples (Suppl. E.1). With 1024 NFE, only the AR model has a lower Gradient Moment than LAW with last-pass training, and its 4-gram overlap (0.70%) is close to that of the data (0.68%) and of the AR model (0.69%). We clarified the handling of end-of-text tokens in Suppl. B.7.

We thank the reviewers for these suggestions, which helped improving the paper. We would be very grateful if the reviewers could share any remaining questions or concerns. We would be happy to discuss them further, so that we can reach a consensus before the discussion period ends.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.