acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Text Generation with Next-Concept Modeling (NCM)

20 views

Abstract

Diffusion language models (DLMs) are appealing for their parallel sampling, which can enable faster generation than autoregressive models; this is increasingly important as language model inference has become a dominant computational workload. In practice, however, DLM quality degrades quickly as the number of sampling steps shrinks, every step processes the full sequence, and learning to generate in every possible order makes training expensive. We introduce , a two-stage, coarse-to-fine framework in which a multiscale residual VQ-VAE encodes text into a few levels of discrete codes and a transformer generates them one level at a time, predicting all codes within a level in parallel. At the M-parameter scale, achieves a better generative perplexity on LM1B than the strongest DLM baseline while using fewer steps and running faster. On OpenWebText, it approaches the generative perplexity of the strongest diffusion baseline while running faster, and on XSum it matches the strongest diffusion baseline in ROUGE-L while sampling faster.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion, LLMs, next-concept, hierarchy

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_Zr7WOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Next-Concept Modeling (NCM), a hierarchical text generation framework designed to improve the generation efficiency of diffusion language models. The proposed method first employs a multiscale residual VQ-VAE to encode text into multiple levels of discrete latent codes. A Transformer then generates these codes in a coarse-to-fine manner, predicting all codes within each level in parallel. By replacing iterative denoising with hierarchical prediction, NCM aims to reduce the number of sequential generation steps while maintaining generation quality. The authors evaluate NCM on unconditional text generation, machine translation, and summarization tasks. The experimental results show substantial inference speedups over several diffusion and flow-based language models, with competitive or improved generation quality.

Critical strengths
  1. Important problem. Improving the inference efficiency of language models is an important research direction, particularly given the increasing computational cost of language generation.
  2. Intuitive and straightforward design. The proposed coarse-to-fine generation framework is conceptually simple and easy to understand. Combining multiscale tokenization with level-wise parallel prediction provides a natural approach to reducing sequential generation steps.
  3. Promising experimental results. The paper reports substantial inference speedups over several diffusion-based baselines across multiple tasks. The evaluation also includes experiments on model scaling and test-time scaling, which help demonstrate the potential efficiency advantages of the proposed framework.
Critical weaknesses

W1. The motivation is straightforward but not new, and the methodological novelty appears limited. The main motivation of this paper is to improve generation efficiency by reducing sequential prediction steps through hierarchical representations and parallel generation. However, this motivation and the underlying design principles have already been explored in several existing studies. In particular, the proposed framework closely resembles the next-scale prediction paradigm introduced by VAR, which also employs multiscale discrete representations and generates them in a coarse-to-fine manner with parallel predictions within each scale. Related studies, such as hierarchical diffusion language models and concept-level language modeling, have also explored hierarchical representations for text generation. Although adapting these ideas to language generation may involve additional challenges, the paper does not sufficiently clarify its novel technical contributions beyond combining existing multiscale tokenization and hierarchical prediction techniques. The proposed architectural components appear to be largely based on established methods, and their unique contributions are not clearly demonstrated.

W2. The method and objective are closely related to multi-token prediction (MTP), but relevant comparisons and discussions are missing. The proposed framework shares a similar high-level motivation with multi-token prediction (MTP) and parallel decoding approaches: reducing the sequential bottleneck of traditional autoregressive language generation by predicting multiple positions in parallel. Although NCM operates on hierarchical latent codes rather than directly predicting future tokens, both approaches aim to improve generation efficiency through parallel prediction. However, the paper neither discusses representative MTP-based approaches, such as Medusa and Hydra, nor provides experimental comparisons with them. More broadly, the evaluation focuses almost exclusively on diffusion and flow-based language models, without comparisons against conventional autoregressive Transformers or other efficient parallel decoding methods. This makes it difficult to assess whether the proposed framework offers meaningful advantages over existing efficient language generation approaches, rather than only over diffusion-based methods. In addition, prior MTP-related research has investigated the importance of modeling dependencies among simultaneously predicted tokens. NCM assumes conditional independence among codes within each level, but the potential limitations of this assumption are not sufficiently discussed or evaluated.

W3. Lack of systematic ablation experiments. The paper lacks ablation studies to validate the effectiveness and necessity of its key design components. For example, the proposed method incorporates multiscale residual tokenization, per-level convolutional refiners, a length-balanced training objective, and a code corruption strategy. However, the individual contributions of these components to generation quality and efficiency are not evaluated. In particular, code corruption is introduced to mitigate error propagation across hierarchical levels, but the paper does not demonstrate how much improvement it provides. Similarly, the benefits of the proposed multiscale hierarchy over simpler alternatives remain unclear. Without these ablation experiments, it is difficult to determine whether the performance improvements come from the specific proposed techniques or primarily from adopting the existing coarse-to-fine generation paradigm.

W4. The limitations of hierarchical parallel prediction are insufficiently investigated. NCM predicts all codes within each level in parallel under a conditional independence assumption. However, natural language exhibits strong syntactic and semantic dependencies among tokens, which may not be fully captured by coarse-level latent representations. Furthermore, the coarse-to-fine generation process may suffer from error propagation, as incorrect predictions at earlier levels can influence all subsequent levels. Unlike iterative refinement methods, NCM does not explicitly revisit previously generated codes. Although the authors introduce code corruption to address this issue, there is no systematic analysis of how well the framework preserves dependencies or recovers from incorrect coarse-level predictions. These limitations deserve more attention, especially given that parallel generation is central to the proposed method.

W5. The efficiency evaluation is not sufficiently comprehensive. The reported inference speedups are encouraging, but the comparisons are mainly conducted against diffusion-based baselines. The lack of comparisons with optimized autoregressive decoding makes it difficult to assess the broader practical efficiency advantages of NCM. Moreover, NCM adopts a two-stage architecture involving a separately trained tokenizer and generator. While the paper reports generator parameter counts and training-token budgets, the overall computational and memory costs of the complete framework are not sufficiently characterized. A more comprehensive evaluation of end-to-end training and inference costs would help substantiate the claimed efficiency improvements.

Questions for authors

Novelty and related work: What are the key technical differences between NCM and existing next-scale or hierarchical generation approaches, particularly VAR and hierarchical diffusion language models? Which components specifically address the challenges of adapting these approaches to text generation?

Comparison with MTP: Given the similar motivation of reducing sequential generation through parallel prediction, why are multi-token prediction and parallel decoding approaches, such as Medusa and Hydra, not discussed or evaluated? How does NCM compare with these approaches and conventional autoregressive Transformers under comparable inference settings?

Ablation experiments: Can the authors provide ablation studies on the per-level convolutional refiners, length-balanced loss, and code corruption strategy? How sensitive is performance to the number of hierarchical levels and their corresponding code lengths?

Conditional independence and error propagation: How does NCM handle dependencies among codes predicted in parallel within the same level? Can the authors quantify the impact of incorrect coarse-level predictions on the quality of subsequent generations and demonstrate the effectiveness of code corruption?

End-to-end efficiency: Can the authors provide a more comprehensive comparison of computational costs, including the tokenizer and decoder, total training compute, inference latency, and memory consumption? How do these costs compare with optimized autoregressive generation?

8 comments

W.1 / Q.1: Novelty relative to VAR, Hierarchical Diffusion Language models and Concept Models

Novelty relative to Hierarchical Diffusion and Concept-Level Models

We thank the reviewer for relating our work to Hierarchical Diffusion and Concept Level Models, we discuss these in section 2 of our related work. Directionally, NCM and these methods share the same motivation; to generate text going from high level ideas to finer details, but these are achieved through different representation and generation techniques..

On the Hierarchical Diffusion Language Model end, HDLM (Zhou et al., 2026), builds a hierarchy over vocabulary tokens and iteratively refines a fixed length sequence through diffusion. HDLM processes the full sequence length at every iteration, compared to NCM that processes shorter sequences at the start of the generation.

Large concept models (LCM) (Barrault et al., 2024) autoregressively generate sentence representations while ConceptLM (Liu et al., 2026) uses discrete concepts to generate tokens autoregressively. NCM on the other hand learns a multiscale residual representation of the entire target sequence, with progressively longer code sequences. It generates all positions within each level in parallel and processes only that level’s positions. This provides a short generation path across levels, with cheaper computation at coarse levels.

Novelty Relative to VAR

We agree that NCM builds on the next-scale prediction paradigm introduced by VAR, as acknowledged in our related work section. Our contribution lies in the architectural and training adaptations that make this paradigm effective for text. As the reviewer rightly notes, adapting these ideas to language generation involves additional challenges. Below we discuss the challenges in adopting VAR built for vision to NCM that works well for text and then we provide a step by step ablation showing the effect of some of the major changes we make. Together, these move the generative perplexity from 169.44 to 13.19.

Challenges in Adapting Multiscale Methods to Text

We experience several challenges when adapting multiscale generation to text. Some of these include: 1) The tokenizer needs to be close to perfect reconstruction compared to in images where a lossy tokenizer could still be fine since many of the errors might not be perceptible. 2) Accumulated errors in the hierarchy in images can still lead to okay images if the errors do not lead to perceptible differences in the image, but this is not really the case for text. 3) It is easy to define the hierarchy in images since you can just spatially downsample but it is not clear what this should be for text. 4) Such VQVAE methods have not been extensively studied for text compared to images (Reviewer 1vbT notes this as well), requiring us to develop and validate suitable recipes for text.

From VAR to NCM: Design Ablation

To better outline our methodological contributions, we present the results of a cumulative ablation study. We start by taking the original VAR and then directly modifying it for text generation and then adding our changes on top of that one by one. For each change, we report the incremental change in generative perplexity and the mean per sample unigram entropy. The list is non-exhaustive and we hope that this makes it easier to isolate our contributions. To keep the cost of completing each ablation low, we run these on the TinyStories dataset using sequences of length 128 and a 12-level hierarchy with code lengths 1, 8, 16, 24, 32, 40, 52, 64, 76, 92, 108, and 128. The tokenizer and NCM are both trained for 40k steps each. These changes together move generative perplexity from 169.44 to 13.19.

#MechanismDescriptionMotivationNFEGen PPL ↓Δ PPLH
—Ground truthReal TinyStories text.Reference text statistics.—8.50—3.959
—ARAutoregressive baseline.Reference for autoregressive generation.1288.89—4.010
—DUODiffusion baseline.Reference for diffusion generation.12822.73—4.050
0Text VARDirect adaptation of VAR to text.Control baseline.12169.440.004.120
1Per-level vocabulariesGive each hierarchy level its own codebook.Allow specialization and reduce competition between coarse and fine levels for a shared codebook.12135.25-34.194.040
2Single-layer encoder and decoderUse a single transformer block in the VQ-VAE encoder and decoder.Limit the decoder’s ability to recover missing information through contextual inference, encouraging the codes to preserve information needed for reconstruction.12119.98-15.274.050
3Non-contextual encodingReplace the VQ-VAE encoder with an embedding lookup without cross-position mixing.Remove dependencies introduced by contextual encoding, aiming to make the resulting codes easier to predict in parallel.1284.31-35.674.040
4Learned resampling at the coarsest levelReplace fixed interpolation for downsampling and upsampling at the coarsest level with convolutional stacks.Learn which features to preserve and combine into coarse codes for subsequent refinement and accurate reconstruction.1276.55-7.764.040
5VQ-VAE dropout and code corruptionApply dropout with probability 0.3 and code corruption during VQ-VAE training.Encourage robust reconstruction from noisy or imperfect latent codes.1263.96-12.594.030
6Nested dropout over levelsDuring VQ-VAE training, randomly retain a prefix of the hierarchy and reconstruct text without the finer levels.Train earlier levels to support reconstruction on their own, encouraging useful coarse representations.1247.34-16.624.011
7Block-diagonal attentionRestrict target attention to the current level, supplying earlier levels through their cumulative reconstruction.Allow prediction from an updated cumulative reconstruction without relying on hidden states from earlier levels.1243.16-4.184.010
8Conditioning corruption, 0.75 → 0.25Corrupt earlier-level conditioning codes with a rate decreasing from 0.75 to 0.25 across levels, while keeping prediction targets unchanged.Reduce the training–inference mismatch by training the generator to handle imperfect earlier reconstructions.1216.34-26.823.968
9Length-balanced loss, weight ∝ length^−0.25Weight each code’s cross-entropy loss by its level length raised to −0.25.Reduce the aggregate loss imbalance between long and short levels, giving coarse-level prediction greater emphasis.1216.03-0.313.969
10Re-encodingRe-encode the current generated text to update the hierarchy, then perform further coarse-to-fine refinement.Allow coarse codes to be revised using information from finer levels.1713.19-2.843.964

W.3 / Q.3: Ablation studies

We ablate the hierarchy depth KK, level-length schedules ℓk\ell_k, loss weighting (α\alpha), and code corruption and include these results below as well as in Appendix F of the updated PDF. These four experiments use TinyStories with sequence length 128, training the tokenizers for 40k steps and the generators for 10k steps. Each table compares settings within its own experiment and thus absolute Gen. PPL values aross tables are not directly comparable.

Hierarchy depth KK

Each hierarchy depth has a separate tokenizer trained for 40,000 steps, 4,096 codes per level, and a linear level schedule from 1 to 128. Each generator is trained for 10,000 steps.

Hierarchy depth KKGen. PPL ↓Entropy HH
4306.94.22
8148.34.15
16101.24.11
3281.24.08

We find that increasing the hierarhcy depth improves the Gen. PPL with diminishing returns. Note that all four tokenizers we trained for this experiment achieved at least 99.95% final reconstruction accuracy.

Level schedule

All schedules use 12 levels, starting from a single code up to 128 codes.

Level scheduleExact level lengths (ℓ1,…,ℓ12)(\ell_1,\ldots,\ell_{12})Gen. PPL ↓Entropy HH
Linear[1, 13, 24, 36, 47, 59, 70, 82, 93, 105, 116, 128]88.44.09
Quadratic[1, 2, 5, 10, 18, 27, 39, 52, 68, 86, 106, 128]59.64.06
Square-root[1, 39, 55, 67, 78, 87, 95, 102, 109, 116, 122, 128]196.74.21
Logarithmic[1, 34, 54, 69, 81, 91, 99, 106, 112, 118, 123, 128]168.24.20
Exponential[1, 3, 6, 9, 14, 20, 29, 39, 53, 72, 96, 128]57.74.05
Cosine[1, 4, 11, 23, 38, 55, 74, 91, 106, 118, 125, 128]77.84.08

We find that quadratic and exponential scheduels perform the best among the options tested.

Loss weighting (α\alpha)

The weights per code at level kk satisfy wk∝ℓk−αw_k\propto\ell_k^{-\alpha}. This means α=0\alpha=0 gives equal weight to each code, whereas α=1\alpha=1 gives equal total weight to each level.

Loss exponent α\alphaGen. PPL ↓
019.0
0.2518.9
0.518.9
0.7519.0
118.9
1.519.4

Code corruption

In this experiment we vary the mean rate of a linear corruption schedule across levels. Corruption replaces conditioning codes with entries uniformly sampled from the same level's codebook.

Mean corruption rateGen. PPL ↓Entropy HH
049.24.06
0.2523.44.00
0.3520.33.97
0.5019.03.97
0.6518.83.95
0.7522.83.96

Moderate corruption improves Gen. PPL but we find excessive corruption then degrades quality.

Additional Ablations

We also investigate ablations of the attention patterns (Appendix F.2.1), context encoders (Appendix F.2.2.), classifier-free guidance (Appendix F.4), and MBR candidate counts (Appendix F.4).

Together, these ablations provide empirical support for NCM's architectural and training choices.

Further Analyses

We also provide analyses listed in the below table. In particular, Appendix G demonstrates progressively improving reconstruction accuracy across the hierarchy and the qualitative interventions in Appendix H suggest coarse levels largely capture parts of global structure while finer levels allow for variation in local details.

AnalysisLocation in Updated PDF
Per-level reconstruction and codebook utilizationAppendix G, Figure 17
Fixing coarse codes and resampling fine levelsAppendix H, Figure 18
Mixing coarse representationsAppendix H, Figure 19
Perturbing early codesAppendix H, Figure 20
Reviewer_vtQHOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper aims to make non-autoregressive text generation faster and more powerful than diffusion LMs. It proposes Next Concept Modeling (NCM), which first trains a multiscale residual VQ-VAE to encode text into multiple levels of discrete codes, then trains a transformer to predict all codes within each level in parallel. Experiments at the ~100M scale on unconditional and conditional generation show that NCM matches or outperforms diffusion baselines with fewer sampling steps.

Critical strengths
  1. This work tackles an important problem by providing a new paradigm for diffusion LMs that models text more accurately with fewer steps.

  2. Bringing next-scale prediction from image generation to language is a fresh and clean idea.

  3. The efficiency gains are large and consistent. The evaluation also covers conditional generation, model-size scaling, and test-time scaling with MBR decoding.

Critical weaknesses
  1. The first concern is whether NCM carries over to real-world LMs. Diffusion LMs are known to converge much more slowly than AR models; for example, prior work has reported 16× slower convergence. A comparison of the scaling behavior of NCM and AR models would therefore be highly desirable. However, the paper includes no autoregressive baseline.

  2. Another major concern is the lack of ablations and analysis supporting the design choices. Per-level refiners, code corruption, level weighting, and the level schedule are not ablated. There is also no analysis of what the coarse levels encode, so it is unclear whether they capture anything resembling “concepts.”

  3. The quality evaluation is limited. Unconditional generation quality is evaluated only using Gen. PPL and entropy, which the paper itself notes can be gamed, without likelihood evaluation or a MAUVE-style metric. NCM’s step count is fixed by the number of levels, so Figure 4 shows it as a single point rather than a quality–speed curve. In addition, few-step distilled DLMs are not compared directly.

  4. NCM currently does not support KV caching.

Questions for authors

My questions have been covered by the weakness section.

6 comments

We thank the reviewer for the constructive and actionable feedback. We are glad they found NCM to be a "fresh and clean idea" and recognized the "large and consistent" efficiency gains. We address each of the reviewer's concerns and questions below.

W.1 Autoregressive baselines and scaling

Scaling NCM

We thank the reviewer for raising this important question. We agree that scaling NCM to the size of real-world language models is an important direction, and we are actively working toward this goal. However, training models at this scale requires substantial computational resources, making a comprehensive large-scale investigation beyond the scope of the present work.

As a first step, we have introduced and released NCM at a smaller research scale, while demonstrating preliminary scaling behavior on LM1B at 97M, 317M, and 725M parameters, with consistent improvements in generation quality as model size increases.

We view this work as establishing the foundations of a promising new language modeling paradigm rather than presenting a fully scaled foundation model. Much like other early-stage language modeling research that begins with smaller research prototypes, we hope that making NCM available will enable the broader community to explore, improve, and scale the approach. We are continuing this effort ourselves and believe that the results presented here provide a useful starting point for further investigation.

Convergence

NCM demonstrates substantially better training efficiency than the diffusion language model baselines we study. In Section 4.5.2, we show that NCM outperforms these baselines on LM1B while using 4.9× fewer training tokens, including both tokenizer and generator training. One possible explanation is that NCM learns a fixed coarse-to-fine generation order, much like AR models learn a fixed left-to-right ordering. In contrast, diffusion language models are typically trained across many different masking patterns and denoising contexts, potentially making the learning problem more challenging.

Comparisons to Autoregressive Baselines

We've updated our manuscript to include more comparsions provide comparisons to autoregressive baselines:

Conditional Generation

NCM outperforms and autoregressive baseline of similar parameter size in both translation and summarition tasks while running ~6x faster. We show results on XSUM and WMT below.

MethodWMT14 De–En SizeBLEU ↑WMT14 NFE ↓XSum SizeROUGE-1 ↑ROUGE-2 ↑ROUGE-L ↑XSum NFE ↓
AR99M25.26499M30.510.224.464
NCM120M22.61188M35.012.828.09
NCM + Soft decisions120M26.21188M36.213.929.69
NCM + CFG120M23.92288M35.313.028.218
NCM + Soft + CFG120M26.52188M36.413.929.617

Soft decisions pass probability-weighted code embeddings to later levels instead of committing to a single discrete code at each position.

Unconditional Generation

For uncoditional generation, we compare NCM to AR models on lm1b and OpenWebtext.

Model / settingCyclesNumber of Function Evaluations (NFE)Gen PPL ↓Entropy
Ground-truth LM1B text——53.514.31
AR Transformer—12866.704.32
DUO—12897.304.30
NCM small — baseline01195.804.32
Re-encode to level 811582.354.34
Re-encode to level 821975.914.35
Re-encode to level 842767.964.35
Re-encode to level 863565.034.36

From the table above, on LM1B, NCM 11 steps outperforms diffusion language models like Duo but is worst than a strong AR baseline. However we could increase the number of sampling steps in NCM by re-encoding the genrated text to a specific coarse level and then regenerating from there. Applying this to LM1B, NCM is able to match a strong AR baseline while using 3.65x fewer sampling steps. We discuss the results for OpenWebtext under the response to W3.

Reasoning

In addition to the above, we also compare to AR and diffusion language models on the reasoning task of sudoku puzzles where the final board must respect several constraints (Section 4.6). NCM outperforms all autoregressive and diffusion language model baselines using significantly fewer sampling steps (Table below).

TypeMethodSizeNFE ↓Easy accuracy (%) ↑Medium accuracy (%) ↑Hard accuracy (%) ↑
ARGreedy25M12814.65.11.0
DiscreteMDLM29M12892.077.130.2
DiscreteDUO29M12896.384.758.4
ContinuousFLM29M12894.282.744.5
ContinuousS-FLM29M12894.885.245.0
ContinuousFMLM+29M1697.792.881.3
OursNCM25M1299.697.890.0

W.2: Ablations and concept analysis

We ablate the hierarchy depth KK, level-length schedules ℓk\ell_k, loss weighting (α\alpha), and code corruption and include these results below as well as in Appendix F of the updated PDF. These four experiments use TinyStories with sequence length 128, training the tokenizers for 40k steps and the generators for 10k steps. Each table compares settings within its own experiment and thus absolute Gen. PPL values aross tables are not directly comparable.

Hierarchy depth KK

Each hierarchy depth has a separate tokenizer trained for 40,000 steps, 4,096 codes per level, and a linear level schedule from 1 to 128. Each generator is trained for 10,000 steps.

Hierarchy depth KKGen. PPL ↓Entropy HH
4306.94.22
8148.34.15
16101.24.11
3281.24.08

We find that increasing the hierarhcy depth improves the Gen. PPL with diminishing returns. Note that all four tokenizers we trained for this experiment achieved at least 99.95% final reconstruction accuracy.

Level schedule

All schedules use 12 levels, starting from a single code up to 128 codes.

Level scheduleExact level lengths (ℓ1,…,ℓ12)(\ell_1,\ldots,\ell_{12})Gen. PPL ↓Entropy HH
Linear[1, 13, 24, 36, 47, 59, 70, 82, 93, 105, 116, 128]88.44.09
Quadratic[1, 2, 5, 10, 18, 27, 39, 52, 68, 86, 106, 128]59.64.06
Square-root[1, 39, 55, 67, 78, 87, 95, 102, 109, 116, 122, 128]196.74.21
Logarithmic[1, 34, 54, 69, 81, 91, 99, 106, 112, 118, 123, 128]168.24.20
Exponential[1, 3, 6, 9, 14, 20, 29, 39, 53, 72, 96, 128]57.74.05
Cosine[1, 4, 11, 23, 38, 55, 74, 91, 106, 118, 125, 128]77.84.08

We find that quadratic and exponential scheduels perform the best among the options tested.

Loss weighting (α\alpha)

The weights per code at level kk satisfy wk∝ℓk−αw_k\propto\ell_k^{-\alpha}. This means α=0\alpha=0 gives equal weight to each code, whereas α=1\alpha=1 gives equal total weight to each level.

Loss exponent α\alphaGen. PPL ↓
019.0
0.2518.9
0.518.9
0.7519.0
118.9
1.519.4

Code corruption

In this experiment we vary the mean rate of a linear corruption schedule across levels. Corruption replaces conditioning codes with entries uniformly sampled from the same level's codebook.

Mean corruption rateGen. PPL ↓Entropy HH
049.24.06
0.2523.44.00
0.3520.33.97
0.5019.03.97
0.6518.83.95
0.7522.83.96

Moderate corruption improves Gen. PPL but we find excessive corruption then degrades quality.

Additional Ablations

We also investigate ablations of the attention patterns (Appendix F.2.1), context encoders (Appendix F.2.2.), classifier-free guidance (Appendix F.4), and MBR candidate counts (Appendix F.4).

Together, these ablations provide empirical support for NCM's architectural and training choices.

Further Analyses

We also provide analyses listed in the below table. In particular, Appendix G demonstrates progressively improving reconstruction accuracy across the hierarchy and the qualitative interventions in Appendix H suggest coarse levels largely capture parts of global structure while finer levels allow for variation in local details.

AnalysisLocation in Updated PDF
Per-level reconstruction and codebook utilizationAppendix G, Figure 17
Fixing coarse codes and resampling fine levelsAppendix H, Figure 18
Mixing coarse representationsAppendix H, Figure 19
Perturbing early codesAppendix H, Figure 20

W.3: Distributional Metrics and Few Step Language Models

Dibstributional Metrics (Mauve)

We compare NCM to both AR and diffusion baselines using the Mauve score on OWT 1024 sequence length and show results in the table below. NCM trails only the AR baseline. Due to NCM's training objective, the likelihood is not very tractable so we do not do any likelihood comparisons.

MethodNFE ↓Gen. PPL ↓EntropyMAUVE ↑
AR102435.35.580.925
MDLM1024110.55.640.827
SEDD1024110.05.620.780
ELF-M6422.05.200.862
DUO102477.75.550.609
FLM102462.85.340.396
FMLM3245.15.180.343
NCM6459.65.400.871

Few Step Language Models

We compare NCM with few step diffusion language models at 8 sampling steps and show the results in the table below. NCM and show them in the table below. NCM outpeforms both Duo and MDLM distilled and trails FMLM although FMLM does have a lower entropy.

MethodStepsGen. PPL ↓Entropy ↑
MDLM + SDTT8105.965.38
Duo + DCD8111.625.41
FMLM886.595.35
NCM897.815.39

NCM Step Count

NCM's step count is fixed by the number of levels but in you can add additional steps to improve quality by reencoding later finer generation into lower levels of the hierarchy and continuing generation from there.

For example, in the first pass you could do:

1 → 2 → 3 → 4 → 5 → 6 → 7 → 8 → 9 → 10 → 11 → 12

And then you could do one cycle that re-encodes the generated text back to e.g the 8 level and then generate to the end again e.g

12 → 8 → 9 → 10 → 11 → 12

Gen PPL Results with Additional Judges

To assess robustness of Gen. PPL to the choice of reference model, we have added Appendix E.1 (Table 8) to the updated PDF, which evaluates LM1B generations using four autoregressive reference models (Mistral-7B, Llama-3.1-8B, Gemma-2-9B, and Qwen3-14B).

ModelSizeTraining TokensMistral-7B ↓Llama-3.1-8B ↓Gemma-2-9B ↓Qwen3-14B ↓Avg. ↓
RDLM108M66B179.54315.401293.67379.00541.90
CANDI139M66B67.7599.96208.59125.87125.54
MDLM139M66B63.6391.84186.53115.25114.31
DUO139M66B55.0877.24141.1696.5992.52
FLM139M66B68.0889.20150.11110.98104.59
NCM97M13.36B59.6975.23106.4191.6483.24

W.4 KV caching

NCM supports caching the prefix in conditional generation settings. However, intermediate computations across generation levels are not currently cached. In some regimes, this can be advantageous, as NCM avoids the growing KV-cache memory overhead associated with autoregressive decoding. Nevertheless, KV caching could be incorporated by extending NCM to a block-wise generation setting, where previously generated blocks are cached and reused when generating subsequent blocks. We view this as a promising direction for further improving NCM's efficiency, particularly for long-context generation.

Reviewer_j9SMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NCM encodes text with a multiscale residual VQ-VAE into K levels of discrete codes, from short coarse levels down to one code per token. A transformer then generates the levels from coarse to fine, predicting all codes within a level in parallel, so the number of sampling steps equals K (8–16) and does not grow with sequence length. At about 100M parameters, NCM reaches 93.05 Gen. PPL on LM1B in 12 steps, against 97.30 for DUO at 128 steps, at similar entropy and 21× lower wall-clock time. On XSum it reaches 28.0 ROUGE-L in 8 steps, and MBR decoding over several cheap samples improves it further. On OpenWebText it approaches DUO.

Critical strengths
  1. The few-step result on LM1B is strong and fairly well controlled. At 12 steps NCM reaches 93.05 Gen. PPL against 97.30 for DUO at 128 steps, with matched sample entropy, and the 21× wall-clock gain is measured end to end on the same hardware. On XSum, NCM without guidance beats every unguided baseline on all three ROUGE scores with 8 forward passes. Poor quality at few steps is the main weakness of current diffusion LMs, so the paper targets the right problem.

  2. The approach is simple and transfers an idea that already works for images. Next-scale prediction over a residual-VQ hierarchy decouples the number of sequential steps from sequence length. Single samples are therefore cheap, so MBR over several samples overtakes the guided baselines at similar wall-clock time, and Gen. PPL improves with model size at constant entropy. This makes NCM a credible alternative to masked and continuous diffusion.

Critical weaknesses
  1. The evidence for the quality claims is weaker than the headline suggests. On OpenWebText, the Fig. 4 caption and §4.2 say NCM matches or beats the best baseline with 8× fewer steps. In the plot, NCM at 16 steps (about 86 Gen. PPL) falls between DUO at 32 and 64 steps, while DUO at 128 steps is about 81, and NCM has the lowest entropy of all methods on OWT, which by itself lowers Gen. PPL. Unconditional quality is measured only by Gen. PPL and unigram entropy, with no variance. The few-step baselines named in the introduction (distilled DUO, FLM) and an autoregressive reference are missing. Because the contribution is a better quality–speed trade-off, this limits how far the results hold beyond LM1B.

  2. The method is under-specified, and none of its parts is ablated. The length-balanced VQ-VAE objective, the per-level heads, code corruption and the loss weight α are mentioned but never defined. Level lengths, codebook sizes, tokenizer size, how guidance is applied and the pretraining data for XSum and WMT are not given, and the appendix contains only samples. No experiment shows which component matters or how sensitive generation is to the tokenizer, and the "concept" framing is never examined, for example by decoding partial reconstructions. The work is therefore hard to reproduce, and it is unclear what produces the gains.

Questions for authors
  1. Could you give a table of OpenWebText Gen. PPL and entropy for NCM and DUO at each step count, together with a distributional metric such as MAUVE? If NCM at 16 steps does not match DUO at 128 steps at matched entropy, the 8× claim should be revised.
  2. Please specify the level lengths, codebook sizes, tokenizer architecture and size, the length-balanced objective, the per-level heads, α, the corruption scheme, the sampling settings, how guidance is applied, and the pretraining corpus for XSum and WMT. Did the Table 2 baselines receive similar pretraining?
  3. Ablations of code corruption, α, the number of levels and the codebook size, and decodes of partial reconstructions, would show what drives the gains and whether the coarse levels carry anything like concepts.
4 comments

W.1 / Q.1: Generation quality and evaluation

NCM Quality at Different Steps & Mauve Score

NCM achieves competitive Gen. PPL at substantially lower sampling cost, but at 16 steps does not match DUO at 128 steps. Thus we have corrected the claim of 8x fewer steps in the updated PDF (Figure 5). We have also added NCM gen ppl points at 32 and 64 steps to the plot.

Below we compare NCM and DUO with Mauve metric as requested by the reviewer.

MethodNFE ↓Gen. PPL ↓EntropyMAUVE ↑
AR102435.35.580.925
MDLM1024110.55.640.827
SEDD1024110.05.620.780
ELF-M6422.05.200.862
DUO102477.75.550.609
FLM102462.85.340.396
FMLM3245.15.180.343
NCM6459.65.400.871

Below we present the table requested with NCM vs Duo gen ppl at different sampling steps.

NFENCM Gen. PPL ↓NCM EntropyDUO Gen. PPL ↓DUO Entropy
1686.35.41122.85.58
3267.05.4196.25.57
6459.65.4085.65.57
128——80.05.55
256——78.65.55
512——78.15.55
1024——77.75.55

Few Step Language Models

We compare NCM with few step diffusion language models at 8 sampling steps and show the results in the table below. NCM and show them in the table below. NCM outpeforms both Duo and MDLM distilled and trails FMLM although FMLM does have a lower entropy.

MethodStepsGen. PPL ↓Entropy ↑
MDLM + SDTT8105.965.38
Duo + DCD8111.625.41
FMLM886.595.35
NCM897.815.39

Autoregressive Reference

We provide an AR reference with the mauve scores above. We've also updated our manuscript to include more comparsions provide comparisons to autoregressive baselines:

Conditional Generation

NCM outperforms and autoregressive baseline of similar parameter size in both translation and summarition tasks while running ~6x faster. We show results on XSUM and WMT below.

MethodWMT14 De–En SizeBLEU ↑WMT14 NFE ↓XSum SizeROUGE-1 ↑ROUGE-2 ↑ROUGE-L ↑XSum NFE ↓
AR99M25.26499M30.510.224.464
NCM120M22.61188M35.012.828.09
NCM + Soft decisions120M26.21188M36.213.929.69
NCM + CFG120M23.92288M35.313.028.218
NCM + Soft + CFG120M26.52188M36.413.929.617

Soft decisions pass probability-weighted code embeddings to later levels instead of committing to a single discrete code at each position.

Unconditional Generation

For uncoditional generation, we compare NCM to AR models on lm1b and OpenWebtext.

Model / settingCyclesNumber of Function Evaluations (NFE)Gen PPL ↓Entropy
Ground-truth LM1B text——53.514.31
AR Transformer—12866.704.32
DUO—12897.304.30
NCM small — baseline01195.804.32
Re-encode to level 811582.354.34
Re-encode to level 821975.914.35
Re-encode to level 842767.964.35
Re-encode to level 863565.034.36

From the table above, on LM1B, NCM 11 steps outperforms diffusion language models like Duo but is worst than a strong AR baseline. However we could increase the number of sampling steps in NCM by re-encoding the genrated text to a specific coarse level and then regenerating from there. Applying this to LM1B, NCM is able to match a strong AR baseline while using 3.65x fewer sampling steps.

Reasoning In addition to the above, we also compare to AR and diffusion language models on the reasoning task of sudoku puzzles where the final board must respect several constraints (Section 4.6). NCM outperforms all autoregressive and diffusion language model baselines using significantly fewer sampling steps (Table below).

TypeMethodSizeNFE ↓Easy accuracy (%) ↑Medium accuracy (%) ↑Hard accuracy (%) ↑
ARGreedy25M12814.65.11.0
DiscreteMDLM29M12892.077.130.2
DiscreteDUO29M12896.384.758.4
ContinuousFLM29M12894.282.744.5
ContinuousS-FLM29M12894.885.245.0
ContinuousFMLM+29M1697.792.881.3
OursNCM25M1299.697.890.0

W.2 / Q.2: Method Details and reproducibility

We have added Appendix C in our updated PDF, which provides detailed tokenizer and generator configurations for every dataset in Tables 6 and 7 respectively. These include depth, latent and hidden widths, attention heads, codebook sizes, level counts, prefix lengths, optimizer settings, and training budgets.

We further provide additional details about loss weighting (α\alpha) in Appendix F.2.3, code corruption in Appendix F.3, and classifier-free guidance in Appendix F.4.

Apologies for the confusion, when we say pre-train we just mean that for each task we first train the NCM unconditionally with the passages/summaries/translations for that task e.g for XSUM we first train unconditionally before doing conditional training.

For the Guidance, we follow the same recipe as in MaskGIT[1] and VAR[2].

W.2 / Q.3: Ablations and concept analysis

Visualizing the Hierachy

Figure 3 and Tbale 4 show a translation example at each partially decoded stage of translation which gives insights into the the information carried by the coarse levels. In Appendix H, we also probe the hierarchy providing more intuition on how it works.

We ablate the hierarchy depth KK, level-length schedules ℓk\ell_k, loss weighting (α\alpha), and code corruption and include these results below as well as in Appendix F of the updated PDF. These four experiments use TinyStories with sequence length 128, training the tokenizers for 40k steps and the generators for 10k steps. Each table compares settings within its own experiment and thus absolute Gen. PPL values aross tables are not directly comparable.

Step-by-Step Ablation KK To better understand whwere the gains come from. We provide a step by step ablation in response to reviewer Zr7W W1/Q1 "From VAR to NCM: Design Ablation". This helps explain how every change we make layers ontop of the other to produce the receipe we have.

Hierarchy depth KK

Each hierarchy depth has a separate tokenizer trained for 40,000 steps, 4,096 codes per level, and a linear level schedule from 1 to 128. Each generator is trained for 10,000 steps.

Hierarchy depth KKGen. PPL ↓Entropy HH
4306.94.22
8148.34.15
16101.24.11
3281.24.08

We find that increasing the hierarhcy depth improves the Gen. PPL with diminishing returns. Note that all four tokenizers we trained for this experiment achieved at least 99.95% final reconstruction accuracy.

Level schedule

All schedules use 12 levels, starting from a single code up to 128 codes.

Level scheduleExact level lengths (ℓ1,…,ℓ12)(\ell_1,\ldots,\ell_{12})Gen. PPL ↓Entropy HH
Linear[1, 13, 24, 36, 47, 59, 70, 82, 93, 105, 116, 128]88.44.09
Quadratic[1, 2, 5, 10, 18, 27, 39, 52, 68, 86, 106, 128]59.64.06
Square-root[1, 39, 55, 67, 78, 87, 95, 102, 109, 116, 122, 128]196.74.21
Logarithmic[1, 34, 54, 69, 81, 91, 99, 106, 112, 118, 123, 128]168.24.20
Exponential[1, 3, 6, 9, 14, 20, 29, 39, 53, 72, 96, 128]57.74.05
Cosine[1, 4, 11, 23, 38, 55, 74, 91, 106, 118, 125, 128]77.84.08

We find that quadratic and exponential scheduels perform the best among the options tested.

Loss weighting (α\alpha)

The weights per code at level kk satisfy wk∝ℓk−αw_k\propto\ell_k^{-\alpha}. This means α=0\alpha=0 gives equal weight to each code, whereas α=1\alpha=1 gives equal total weight to each level.

Loss exponent α\alphaGen. PPL ↓
019.0
0.2518.9
0.518.9
0.7519.0
118.9
1.519.4

Code corruption

In this experiment we vary the mean rate of a linear corruption schedule across levels. Corruption replaces conditioning codes with entries uniformly sampled from the same level's codebook.

Mean corruption rateGen. PPL ↓Entropy HH
049.24.06
0.2523.44.00
0.3520.33.97
0.5019.03.97
0.6518.83.95
0.7522.83.96

Moderate corruption improves Gen. PPL but we find excessive corruption then degrades quality.

Additional Ablations

We also investigate ablations of the attention patterns (Appendix F.2.1), context encoders (Appendix F.2.2.), classifier-free guidance (Appendix F.4), and MBR candidate counts (Appendix F.4).

Together, these ablations provide empirical support for NCM's architectural and training choices.

  1. Chang, H., Zhang, H., Jiang, L., Liu, C. and Freeman, W.T., 2022, June. Maskgit: Masked generative image transformer. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 11305-11315). IEEE.

  2. Tian, K., Jiang, Y., Yuan, Z., Peng, B. and Wang, L., 2024. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37, pp.84839-84865.

Further Analyses

We also provide analyses listed in the below table. In particular, Appendix G demonstrates progressively improving reconstruction accuracy across the hierarchy and the qualitative interventions in Appendix H suggest coarse levels largely capture parts of global structure while finer levels allow for variation in local details.

AnalysisLocation in Updated PDF
Partially Decoded hierarchyFigure 3, Table 4
Per-level reconstruction and codebook utilizationAppendix G, Figure 17
Fixing coarse codes and resampling fine levelsAppendix H, Figure 18
Mixing coarse representationsAppendix H, Figure 19
Perturbing early codesAppendix H, Figure 20

We hope the clarifications and additional details address the reviewer's concerns. We have also revised our draft to take into account the reviewer's feedback. Should any important concern remain, we would be grateful if the reviewer could bring it to our attention during the discussion period. In light of the additional evidence, we would appreciate it if the reviewer reconsiders their score.

Dear reviewer j9SM,

We have worked tirelessly to incorporate your valuable feedback, provide relevant clarifications, and address any concerns. With about one day remaining in the discussion period, we would really like the opportunity to resolve any important concerns the reviewer may still have. If there are no such concerns given our rebuttal responses and updated PDF draft, we would appreciate it if the reviewer reconsiders their score.

Thank you so much for your time and consideration!

Sincerely, the authors

Reviewer_1vbTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a novel approach to language modeling via next concept modeling (NCM). The idea lies in constructing KK discrete latent representations with different compression levels and then training a model to generate data autoregressively from the highest to the lowest compression level. This paradigm enables significantly faster convergence while achieving comparable performance. The authors validate this approach across different settings, including unconditional and conditional generation, comparing it against multiple baselines.

Critical strengths
  1. The experiments are excellent in terms of positioning the different models and demonstrating how the method scales. The authors have put considerable effort into analyzing different setups, making it clear how NCM speeds up generation and what maximum quality it achieves. The improvements in the results also appear substantial.

  2. The method itself is simple yet elegant. It has strong connections to the VAR literature, which lends further credibility to the reported results.

Critical weaknesses

[W.1] Clarity of experiments. There are many experiments that must be conducted for a novel approach in discrete diffusion research, and the authors have done a great job of presenting a substantial number of them. However, the way the experiments and their results are presented makes it somewhat difficult to independently assess how well the evidence supports the authors' conclusions and whether important aspects of the method have been adequately accounted for.

(1) For example, the authors claim that the method is efficient in both training and sampling. While the analysis of sampling appears sufficient in terms of wall-clock time, the analysis of training only reports the number of tokens consumed without providing any timing information. This is particularly relevant because the method consists of two stages, and token consumption does not necessarily reflect training speed. Additionally, although the authors mention that the AE is retrained for each task, their efficiency analysis seems to focus primarily on the generator, without adequately accounting for the AE's training cost and model size. This is particularly relevant given the relatively limited number of established approaches for training foundational AEs for text. The cost of training the AE should be taken into account when assessing the overall efficiency of the method. Moreover, it remains unclear how this approach would scale to larger models, such as LLaDA.

(2) Next, it seems that the same symbol KK is used to denote both the additional MBR groups and the depth of the AE. This makes it particularly difficult to analyze Table 2 and Figure 5, understand the correspondence between experiments, and thus verify the authors' claims.

(3) Finally, I found that most of the hyperparameters are left implicit (I couldn't find what are the values used for the experiments). Specifically, the structure of the AE is not sufficiently clear. The authors do not specify its KK (the depth), ℓk\ell_k, α\alpha (despite introducing it in lines 260-261), or the code corruption method (lines 268-269). This issue is further compounded by the fact that these parameters are introduced but never analyzed, which is particularly important for a novel method. This additionally raises a question on reproducibility of the method.

Questions for authors

Please see weaknesses

5 comments

We thank the reviewer for their constructive and actionable feedback. We are glad they found NCM’s method to be “simple yet elegant” and also appreciated the extensive experiments, while recognizing the substantial improvements in results. We address each of the reviewer's questions and concerns below.

W.1(1): Training efficiency and scaling

We report training budgets that include both the tokenizer and generator, with a separate tokenizer trained for each dataset. On LM1B, the tokenizer (111.4M parameters) is trained on 6.80B tokens and the generator (97.1M parameters) is trained on 6.55B tokens, for a total of 13.36B training tokens. This is 4.9× fewer than the 66B tokens used to train the diffusion baselines (Appendix C.2). On OpenWebText, the corresponding total is 127.2B training tokens, roughly 4× fewer than the 500B+ tokens used by the MDLM and DUO baselines.

We have benchmarked both NCM stages and the diffusion baselines on the same 8 H100 GPUs.

Model / StageWall Clock Time (h)H100 GPU-hours
NCM tokenizer1.5912.75
GPT-2-small (13B tokens)2.7321.86
NCM generator6.6553.21
NCM total8.2565.96
GPT-2-small (66B tokens)13.76110.10
MDLM16.80134.42
DUO18.85150.82
FLM17.76142.08
CANDI26.61212.92
RDLM60.63485.04

As shown in the table above, NCM requires fewer GPU-hours than MDLM (the cheapest baseline) when we train each model on their full training schedule on LM1B.

We show preliminary scaling on LM1B by at 97M, 317M, and 725M params. Figure 6(a) shows consistent quality improvements with model size. We agree that how these gains extend to the 8B parameter scale of LLaDA remains an open question. LLaDA’s full pretraining used 2.3T tokens and 130,000 H800 GPU-hours. This is beyond the scope of the present study. We are actively investigating larger-scale NCM models as a natural extension of this work. In the meantime, our results establish promising scaling across nearly an order of magnitude and provide a foundation for further research.

W.1(2): Notation

We apologize for overloading KK and any resulting confusion. In the revised PDF, KK denotes only the number of levels in the hierarchy, and MM denotes the number of MBR candidates throughout. This change has been propagated in Section 4.4.1, Figure 6 (b, c), Appendix F.4, and Figure 16.

W.1(3): Architecture, hyperparameters, and ablations

We have added Appendix C in our updated PDF, which provides detailed tokenizer and generator configurations for every dataset in Tables 6 and 7 respectively. These include depth, latent and hidden widths, attention heads, codebook sizes, level counts, prefix lengths, optimizer settings, and training budgets.

We ablate the hierarchy depth KK, level-length schedules ℓk\ell_k, loss weighting (α\alpha), and code corruption and include these results below as well as in Appendix F of the updated PDF. These four experiments use TinyStories with sequence length 128, training the tokenizers for 40k steps and the generators for 10k steps. Each table compares settings within its own experiment and thus absolute Gen. PPL values aross tables are not directly comparable.

Hierarchy depth KK

Each hierarchy depth has a separate tokenizer trained for 40,000 steps, 4,096 codes per level, and a linear level schedule from 1 to 128. Each generator is trained for 10,000 steps.

Hierarchy depth KKGen. PPL ↓Entropy HH
4306.94.22
8148.34.15
16101.24.11
3281.24.08

We find that increasing the hierarhcy depth improves the Gen. PPL with diminishing returns. Note that all four tokenizers we trained for this experiment achieved at least 99.95% final reconstruction accuracy.

Level schedule

All schedules use 12 levels, starting from a single code up to 128 codes.

Level scheduleExact level lengths (ℓ1,…,ℓ12)(\ell_1,\ldots,\ell_{12})Gen. PPL ↓Entropy HH
Linear[1, 13, 24, 36, 47, 59, 70, 82, 93, 105, 116, 128]88.44.09
Quadratic[1, 2, 5, 10, 18, 27, 39, 52, 68, 86, 106, 128]59.64.06
Square-root[1, 39, 55, 67, 78, 87, 95, 102, 109, 116, 122, 128]196.74.21
Logarithmic[1, 34, 54, 69, 81, 91, 99, 106, 112, 118, 123, 128]168.24.20
Exponential[1, 3, 6, 9, 14, 20, 29, 39, 53, 72, 96, 128]57.74.05
Cosine[1, 4, 11, 23, 38, 55, 74, 91, 106, 118, 125, 128]77.84.08

We find that quadratic and exponential scheduels perform the best among the options tested.

Loss weighting (α\alpha)

The weights per code at level kk satisfy wk∝ℓk−αw_k\propto\ell_k^{-\alpha}. This means α=0\alpha=0 gives equal weight to each code, whereas α=1\alpha=1 gives equal total weight to each level.

Loss exponent α\alphaGen. PPL ↓
019.0
0.2518.9
0.518.9
0.7519.0
118.9
1.519.4

Code corruption

In this experiment we vary the mean rate of a linear corruption schedule across levels. Corruption replaces conditioning codes with entries uniformly sampled from the same level's codebook.

Mean corruption rateGen. PPL ↓Entropy HH
049.24.06
0.2523.44.00
0.3520.33.97
0.5019.03.97
0.6518.83.95
0.7522.83.96

Moderate corruption improves Gen. PPL but we find excessive corruption then degrades quality.

Additional Ablations

We also investigate ablations of the attention patterns (Appendix F.2.1), context encoders (Appendix F.2.2.), classifier-free guidance (Appendix F.4), and MBR candidate counts (Appendix F.4).

Together, these ablations provide empirical support for NCM's architectural and training choices.

We hope the clarifications and additional details address the reviewer's concerns. We have also revised our draft to take into account the reviewer's feedback. Should any important concern remain, we would be grateful if the reviewer could bring it to our attention during the discussion period. In light of the additional evidence, we would appreciate it if the reviewer reconsiders their score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.