acceptodds
Under review as a conference paper at ICLR 2027

ELF-REG: Scaling Continuous Diffusion Language Models to Reasoning Tasks

Abstract

Fully continuous diffusion language models (dLMs) denoise continuous representations without intermediate discretization, then decode all response tokens in parallel at the final step. Their performance on challenging reasoning tasks remains less established than that of autoregressive (AR) LLMs and masked dLMs. We scale Embedded Language Flows (ELF) to mathematical reasoning and code generation on GSM8K, MATH-500, HumanEval, and MBPP. We introduce ELF-REG, which improves learning with representation alignment and entanglement (REPA+REG), where a frozen AR teacher supervises intermediate denoiser features and supplies a global representation that is jointly denoised with the response. ELF-REG-L achieves 55.96% pass@1 on GSM8K at 64 network function evaluations (NFE), and 13.39% on MATH-500 and 22.56% on HumanEval at 128 NFE. It outperforms the evaluated comparable-scale dLMs in pass@1 on GSM8K and code, and improves MATH-500 pass@1 from 10.55% for the ELF-L baseline to 13.39% with ELF-REG-L. Without few-step training, the same task-specific checkpoints support strong low-NFE performance through early-stop, which decodes an intermediate clean prediction without completing the denoising trajectory. At 16 NFE, ELF-REG-L reaches 41.21% HumanEval pass@10, outperforming recent continuous dLMs of comparable scale.

Primary area:
generative models
Keywords:
diffusion language models, representation alignment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Xn8ROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors tackle the setting of pretraining continuous diffusion language models with a downstream evaluation of mathematical reasoning. They combine three existing methods, ELF, REPA, and REG, into a pretraining recipe that performs well in absolute terms, e.g., their largest ~800M parameter model achieves 13.39% MATH pass@1. They also provide favorable comparisons to dLMs of similar scale, although these comparisons are not training FLOP-matched. Lastly, they find that their ELF-REG recipe supports early stopping by decoding tokens from intermediate continuous states without any explicit few-shot distillation.

Critical strengths
  • The most encouraging part of this paper is that it could serve as a "hero run", showing that combining a bag of tricks such as ELF, REPA, and REG together yields a model that is closer to the ballpark of decent AR LLMs. e.g., 13.39% MATH pass@1, compared to 24.8% for Llama-3.2-1B-Instruct.
  • It is interesting that the authors find that early exiting works even without dedicated few-shot distillation. Ultimately, few-shot inference is one of the main motivations for exploring (continuous) diffusion LMs at all, so this is a nice result.
Critical weaknesses
  • The critical weakness here is that the comparisons are not compute-controlled. This leads to a number of questions: (1) compared to the strongest baseline recipe (such as PlaidQ+CFG), how does ELF-REG compare (at various inference budgets) when one carefully controls training FLOPs? The authors provide the "Task-training tokens" in Fig. 1(d), but this is not a standard approach to comparing pretrained LMs and so I have a hard time buying this as a replacement for an experiment that controls data and compute. This is ultimately the relevant experiment in deciding whether one pretraining recipe is better than another. (2) Similarly, it is unclear if the variant with REG+REPA is even better than ELF alone, when one adjusts for the extra training FLOPs of REG+REPA. Given that this paper combines existing methods ELF+REG+REPA and therefore provides little methodological novelty, I place more weight on this experiment setup being sound.
  • I found the writing in places to be extremely terse and AI-sounding. e.g., the writing in Sections 3.2 and 3.3.
Questions for authors
  • Can you fix a data distribution and a set of FLOP counts, and build pretraining scaling laws for ELF-REG, ELF, and the strongest baseline (e.g., PlaidQ)? This approach has been taken in the continuous diffusion literature since Plaid [1], and has been standard in language modeling since Kaplan et al. [2]. e.g., you could follow the isoFLOP analysis of Chinchilla [3]. This would be the proper way to compare scaling recipes and would make the tentative results here sound.
  • Can you take a pass through the paper and clean up the writing, which sounds terse and AI-generated in many places such as S3.2, S3.3?

[1] Ishaan Gulrajani, Tatsunori B Hashimoto. Likelihood-Based Diffusion Language Models. Advances in Neural Information Processing Systems 36 (NeurIPS 2023). [2] Jared Kaplan et al. Scaling Laws for Neural Language Models, 2020. https://arxiv.org/abs/2001.08361 [3] Jordan Hoffman et al. Training Compute-Optimal Large Language Models, 2022. https://arxiv.org/abs/2203.15556

4 comments
AuthorsOpenReview ↗

Response to reviewer Xn8R (1)

We added consistency-sampler results, comparisons of accuracy at lower training and inference cost, and edited section 3 for readability in the updated version of the paper. Below, we report training and inference FLOPs, including the frozen encoder's inference cost, and GSM8K isoFLOP comparisons. Additions to the paper are coloured blue.

[Strength 1] The most encouraging part of this paper is that it could serve as a "hero run"

To further demonstrate the superior performance of our recipe, we added consistency-sampler results without additional training. Using the consistency sampler and same NFE, ELF-REG-L outperforms ELF-L baseline across all six benchmarks (Table 4). ELF-REG's highest pass@1 accuracy using the consistency sampler are 61.50% on GSM8K (NFE 128) and 20.15% on MATH-500 (NFE 128), compared with 58.40% and 16.85% for the ELF baseline. On HumanEval and HumanEval+, ELF-REG reaches 33.08% and 31.40% (both NFE 128), versus 28.35% and 26.83%. On MBPP-378 and MBPP-500, ELF-REG reaches 38.76% and 25.60% (both NFE 128), versus 36.51% and 24.55%. Please see Table 4 for more details.

[W1] [Q1] Training FLOPs - We add GSM8K isoFLOP comparisons.

... fix a data distribution and a set of FLOP counts, and build pretraining scaling laws for ELF-REG, ELF ...?

We agree that task-training tokens alone do not control training compute; we note that the ELF paper, which we build on, exclusively reports training tokens as a proxy for training compute. We have therefore added isoFLOP experiments comparing ELF baseline and ELF-REG on the same training distribution across four model sizes, together with training and inference FLOP comparisons.

Following your suggestions and the isoFLOP approach in Plaid [1] and Chinchilla [2], we evaluate four model sizes at four training FLOP budgets on our GSM8K training data (Figure 10). We evaluate performance using the consistency sampler with early stop at NFE 64. We report reference training FLOPs that count one full-dataset encoder/teacher feature-preparation pass per training run and assume reuse of those cached features.

ELF-REG improves accuracy at the same model size and training FLOPs in almost all cases. The table below reports the highest accuracy across the model sizes evaluated at each budget:

Training FLOPs (×1019\times 10^{19})ELF baseline pass@1 (%)ELF-REG pass@1 (%)
1.5024.6825.63
2.5234.7639.12
4.6639.8445.26
6.6543.3352.18

Each entry is the highest mean accuracy chosen from the model sizes evaluated at that budget (averaged over 4 generation seeds). Thus, the improvement persists after accounting for the extra training FLOPs of REPA+REG. We report reasoning accuracy directly because improvements in pretraining loss or NLL for scaling laws don't reliably predict downstream performance [3].

Figure 10 also includes base training loss and GPT-2-large response NLL. The response NLL scores ELF's generated responses under GPT-2-large, conditioned on the prompt and it is not ELF's own likelihood. GPT-2-large is also used to evaluate generated text in prior dLM papers, including SEDD [4], Block Diffusion [5] and Duo [6]. Our study compares ELF-baseline and ELF-REG at fixed training budgets and we don't extrapolate a general pretraining scaling law to larger models from these results.

The new efficiency comparisons also extend to MATH and code (Table 31). For example, ELF-REG-L achieves 17.40% on MATH-500 compared with 16.85% for ELF-L baseline, using 19.1% fewer training FLOPs and 12.3% fewer inference FLOPs.

... compared to the strongest baseline recipe (such as PlaidQ+CFG), how does ELF-REG compare ... when one carefully controls training FLOPs?

We have not included PlaidQ in a scaling law experiment because the repository linked in PlaidQ is not accessible and returns HTTP 404. Without its implementation, we can't faithfully implement PlaidQ's training on same data and FLOP budgets, and we can't directly measure the training FLOP count. However, based on the architecture using dense attention, training settings, and other hyperparameters reported by PlaidQ, we estimate PlaidQ's training FLOPs to be approximately ≈22.82×1020\approx 22.82 \times 10^{20} (we assume 380k optimiser steps, 512 batch size, and 11.731 TFLOPs per input), whereas ELF-REG uses 2.861×10202.861 \times 10^{20} FLOPs, almost an order of magnitude less compute than PlaidQ to achieve superior performance when considering training compute.

Additionally, we also estimate the inference FLOPs of PlaidQ from its architecture and sampling settings, including CFG. On HumanEval, ELF-REG-L achieves 24.24% at NFE 16, using 24.954 TFLOPs per answer, including the frozen encoder. PlaidQ+CFG achieves 22.04% at NFE 257 on HumanEval, with around 84.977 TFLOPs per answer. Therefore, ELF-REG out-performs PlaidQ+CFG at fewer inference FLOPs.

AuthorsOpenReview ↗

Response to reviewer Xn8R (2)

We provide a comprehensive comparison of ELF-REG with baselines regarding training and inference FLOPs, consistently showing superior ELF-REG performance at lower compute. On GSM8K, ELF-REG-B achieves higher accuracy with fewer training and inference FLOPs. On code, ELF-REG-L achieves higher accuracy with fewer inference FLOPs. These comparisons are also included in Table 32 in the updated manuscript.

GSM8K:

ModelNFEPass@1 (%)Training FLOPs (×1019\times 10^{19})Inference TFLOPs/answer
ELF-REG-B, epoch 91637.915.0814.691
FMLM+ (Init)6433.60≈\approx 13.99311.616
MDLM1024≈\approx 33.905.196138.555
Duo1024≈\approx 36.025.196138.554
S-FLM (top-1)512≈\approx 18.185.19569.272

ELF training FLOPs include one-time encoder/teacher feature preparation, assuming we generate text features from Qwen encoder/teacher and reuse them.

Code (HE = HumanEval; HE+ = HumanEval+):

ModelNFEHE (%)HE+ (%)MBPP-378 (%)TFLOPs/HETFLOPs/MBPP-378
ELF-REG-L, epoch 121624.2423.1727.1824.95424.954
Open-dCoder12820.8017.6023.9033.94631.719
oDLM12817.8716.4022.9942.12739.323
PlaidQ12919.0218.0522.18≈\approx 42.654≈\approx 39.816
PlaidQ+CFG25722.0419.7024.58≈\approx 84.977≈\approx 79.324

ELF-REG uses consistency sampling, and its inference FLOPs include the frozen encoder and final decoding. Accuracies are averaged over 4 generation seeds, using EvalPlus for code evaluation. Baseline accuracies are values extracted from the papers and ≈\approx denotes estimated training/inference costs or accuracy read from figures. Bold marks the highest accuracy or lowest compute in each column.

For example, ELF-REG-L exceeds PlaidQ+CFG on all three code metrics at NFE 16 compared to PlaidQ+CFG's NFE 257, while using fewer inference FLOPs.

[W2] [Q2] terse and AI-sounding writing.

We recognise that section 3 is more terse than intended and we have edited section 3 to make the writing sound less terse. We experimented with writing section 3 using a LLM, but it was too long so we wrote it from scratch. The terseness was due to aggressive space saving, and we should have written more natural prose instead.

[1] Likelihood-Based Diffusion Language Models, https://arxiv.org/pdf/2305.18619

[2] Training Compute-Optimal Large Language Models, https://arxiv.org/pdf/2203.15556

[3] Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check, https://arxiv.org/abs/2507.00885

[4] Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution, https://arxiv.org/pdf/2310.16834

[5] Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models, https://arxiv.org/pdf/2503.09573

[6] The Diffusion Duality, https://arxiv.org/pdf/2506.10892

I thank the authors for running a compute-controlled pretraining experiment. Unfortunately the setup is still quite non-standard for LLM pretraining, where it appears the authors perform epoched pretraining on GSM8K directly rather than pretraining on a standard corpus such as DCLM or FineWeb. I therefore have a hard time trusting that their approach actually generalizes to relevant pretraining settings, in which models are pretrained from scratch on mixes of web and high-quality data, and later finetuned/post-trained on tasks such as GSM8K.

Given that this paper is a recipe combining several established methods in the literature, and hence has limited methodological novelty, I believe the bar should be higher for a sound empirical evaluation. Because of the issues in soundness above, I cannot justify giving an accept recommendation.

We thank the reviewer for the response. We don't claim generalisation to general-corpus LLM pretraining, which is beyond our available compute. Instead, we study task-specific math reasoning and code generation with fully continuous dLMs. We follow existing works in training on task-specific data instead of general-corpus pretraining, including S-FLM, FMLM+, and DBTM [1-3], which we already cited. The table contains recent papers that do GSM8K task-specific training:

StudyGeneral pretraining?Benchmark Task
S-FLM [1]NoGSM8K
FMLM+ [2]NoGSM8K
DBTM [3]NoGSM8K
DoT (from scratch) [4]NoGSM8K
CoBit-S [5]NoGSM8K
Simplex Diffusion Models [6]NoGSM8K
AlphaDLM [7]NoGSM8K
ClockDLM [8]NoGSM8K
DoubleCat / SemiCat [9]NoGSM8K
H-CoBit [10]NoGSM8K

The reviewer cited Plaid [11], which performs isoFLOP analysis of language-model likelihood without directly evaluating math/code reasoning. Moreover, [12] shows that better perplexity doesn't reliably translate into better downstream performance. We therefore measure reasoning accuracy directly when controlling for training FLOPs.

Our primary contribution is establishing an effective recipe for training fully continuous dLMs for reasoning that achieves higher GSM8K accuracy with fewer training FLOPs than ELF baseline, S-FLM (top-1) [1], and FMLM+ (Init) [2]. REPA supplies supervision at individual token positions, while REG requires the shared denoiser to predict a representation of the entire prompt and response. With consistency sampling at NFE 64, ELF-REG-L reaches 45.26% at 4.66e19 training FLOPs, exceeding ELF-L baseline's 43.33% at 6.65e19.

The reasoning-specific motivation for early-stop is that obtaining the correct answer does not require every reasoning token to stabilize. ELF-REG-B reaches 37.64% accuracy at NFE 64 versus 38.29% at NFE 512, although complete-response agreement is only 8.61% (Table 11). Thus, one eighth of the NFE budget already provides near-final reasoning accuracy without dedicated few-step training. This concerns answer correctness, which unconditional generation perplexity does not measure.

[1] Language Modeling with Hyperspherical Flows, https://arxiv.org/abs/2605.11125

[2] Posterior Refinement: Fast Language Generation via Any-Order Flow Maps, https://arxiv.org/abs/2606.24773

[3] Discrete Beckmann Transport Models for One-Step Language Modeling and Reasoning, https://arxiv.org/abs/2609.15903

[4] Diffusion of Thoughts: Chain-of-Thought Reasoning in Diffusion Language Models, https://arxiv.org/abs/2402.07754

[5] CoBit: Language Modeling with Bitstream Diffusion, https://arxiv.org/abs/2605.07013

[6] Simplex Diffusion Models, https://arxiv.org/abs/2609.35553

[7] Alpha Diffusion Language Models: Factorization Alone Is Not the Problem, https://arxiv.org/abs/2609.38066

[8] Clock Diffusion: Efficient Semi-Autoregressive Continuous Diffusion Language Models, https://arxiv.org/abs/2610.00894

[9] On the Design Space of Continuous Diffusion Language Models, https://arxiv.org/abs/2610.13834

[10] Denoising Hierarchical Representations: Joint Continuous Diffusion for Language Modeling, https://arxiv.org/abs/2610.08738

[11] Likelihood-Based Diffusion Language Models, https://arxiv.org/abs/2305.18619

[12] Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check, https://arxiv.org/abs/2507.00885

Reviewer_tfNhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ELF-REG, which extends ELF to math reasoning and code tasks. The authors combine REPA and REG to improve training, basically using a frozen AR teacher to provide intermediate representations for alignment. The paper also applies early-stop decoding to generate answers with fewer NFE, without additional few-step training. Experiments on some datasets like GSM8K and MATH-500 show improvements over the original ELF baseline.

Critical strengths

Simple and easy-to-follow method. The overall method is straightforward and the paper clearly explains how REPA and REG are added to ELF. To me, it is interesting to see these existing representation learning techniques applied to fully continuous dLMs for reasoning and code generation.

Interesting low NFE results. The early-stop results are promising, especially on GSM8K and code generation. I like that the model can produce useful answers before completing the full denoising process, without needing separate few-step training.

Useful ablations and analysis. The authors provide ablations on REPA, REG, alignment depth, and early-stop settings. The analysis of endpoint error and token stability is also interesting.

Critical weaknesses

Limited novelty and unconvincing technical motivation. My biggest concern is that almost every component comes from existing work, including ELF, REPA, REG, and early stopping. The paper mainly combines these techniques so the technical contribution feels quite incremental to us readers. More importantly, the motivation seems somewhat circular like REPA and REG have worked in previous diffusion settings so the authors would like to apply them to ELF. The paper does not clearly identify what specific problem in continuous reasoning requires these components, or why this particular combination is the right solution.

The role and necessity of REG are unclear. The ablation results actually raise my questions. In Tab 4, removing the REG token at inference barely changes performance, from 28.92% to 28.66%. Appendix Tab 25 also shows that REG alone provides almost no improvement over the baseline in that setting. This makes me question what REG actually contributes. It seems possible that the benefit mainly comes from the additional training supervision from the teacher model rather than having a global representation during inference.

Heavy dependence on specific pretrained models. The method relies on a frozen Qwen3-0.6B encoder and a Qwen3-1.7B teacher. The extra supervision from the pretrained teacher could explain part of the improvement, yet it is unclear how much of the benefit comes from the specific REPA+REG design rather than teacher-based supervision more generally. Although the authors compare some text encoders, they do not systematically evaluate different teacher models. I wonder whether the improvements would remain similar with a weaker teacher, a stronger teacher, or a teacher from another model family especially from dLLMs like LLaDA and Dream instead of AR models. Without these comparisons, it is hard to judge how robust the method is to the choice of pretrained representations.

Some improvements are relatively small, especially after longer training. For example, ELF-REG-B only improves GSM8K pass@1 from 37.71 to 38.11. Another interesting observation is that REPA+REG gives a much larger improvement at an early training epoch, while the final ELF-B baseline catches up considerably after more training. This makes me wonder whether the method mainly helps with training convergence rather than substantially improving the final capability. Maybe indepth analysis with matched training budgets and longer training would help clarify this.

The sensitivity analysis of the loss weights and other hyperparameter is extremely limited. The method introduces two additional loss terms, but their weights are almost not studied. The REPA weight is fixed at 0.5, and the REG weight is only compared with 0.03. Interestingly, the smaller REG weight gives almost the same test performance as the default setting in Appendix Table 25, which makes me question the effectiveness of REG component.

The model-size and efficiency comparisons are not fully convincing. The paper emphasizes strong performance at low NFE and comparisons with similarly sized dLMs. However, the reported ELF parameter counts do not include the frozen Qwen3 encoder required at inference, and NFE also excludes its forward pass. Different baselines also use different prompting, response lengths, and code evaluation settings. So I think end-to-end latency and total inference cost analysis would make the efficiency claims more convincing. Please see my Q5 below.

Questions for authors

Q1 What is the main technical motivation for introducing REPA and REG? The original ELF already learns to reconstruct clean representations from a pretrained encoder. What specific limitation does the additional AR teacher address in your dLM settings? The current motivation seems to rely mainly on the fact that REPA and REG worked in previous diffusion models. Could the authors explain why these components are particularly needed for reasoning in your settings, rather than simply providing additional training supervision?

Q2 Does REG actually contribute? Table 4 shows that removing REG at inference barely affects performance, and Appendix Table 25 shows little improvement from REG alone. Does REG mainly act as an additional training objective, or does the jointly denoised global representation provide something important during inference? Could the authors better explain why this component is necessary?

Q3. How sensitive is ELF-REG to the loss weights? The REPA weight is fixed at 0.5, and only two REG weights are compared (Don’t know why select and compare 0.03 and 0.1 only). Could the authors provide a more complete sensitivity analysis of both loss weights? In particular, the smaller REG weight appears to perform similarly to the default setting, this makes me really confused about the real effectiveness of REG.

Q4 Does REPA+REG mainly improve training speed or final model performance? As in Figure 3, the early training improvements look much larger than the final improvement for ELF-B. Could the authors compare models at matched training budgets and show whether your performance gap remains after longer training?

Q5 How fair are the efficiency comparisons? Since the parameter counts and NFE exclude the frozen encoder required during inference, could the authors provide end-to-end inference time and total computational cost? Also, was SCCFG for the code benchmarks selected on a separate validation set or directly on the reported evaluation benchmarks?

9 comments
AuthorsOpenReview ↗

Response to reviewer tfNh (1)

Thank you for the review. We added consistency sampler results and comparisons of accuracy at lower training and inference cost to the updated version of the paper. Below, we clarify the training contribution of REG, present additional teacher ablations, and compare accuracy at lower training and inference cost. We also report controls for prompting, response lengths and code evaluation. Additions to the paper are coloured blue.

To further demonstrate the superior performance of our recipe, we have added consistency-sampler results without additional training. Using the consistency sampler and same NFE, ELF-REG-L outperforms ELF-L baseline across all six benchmarks (Table 4). ELF-REG's highest pass@1 accuracies using the consistency sampler are 61.50% on GSM8K (NFE 128) and 20.15% on MATH-500 (NFE 128), compared with 58.40% and 16.85% for the ELF baseline. On HumanEval and HumanEval+, ELF-REG reaches 33.08% and 31.40% (both NFE 128), versus 28.35% and 26.83%. On MBPP-378 and MBPP-500, ELF-REG reaches 38.76% and 25.60% (both NFE 128), versus 36.51% and 24.55%. Please see Table 4 for more details.

[W1] [Q1] Limited novelty and motivation - We develop an effective & efficient recipe for continuous reasoning.

We thank the reviewer for raising this concern. We clarify the technical motivation below and have edited the paper's introduction to emphasise our contributions.

Our main contribution is building an effective recipe for math reasoning and code generation with fully continuous dLMs, using frozen AR representations, whereas ELF using the bidirectional t5-small encoder did not perform well in our ablations (Figure 4). Our aim is to improve reasoning accuracy with limited training and inference compute to narrow the gap between continuous dLMs and AR LLMs. ELF learns to predict clean text-encoder states and decode them into tokens, and we hypothesise that clean representations from a more expressive AR teacher can provide useful additional supervision for this denoising process, improving reasoning performance over the ELF baseline and prior works with less compute cost. REPA supervises intermediate denoiser features at individual token positions, while REG supplies a global target for the entire prompt and response. We note that representation alignment methods are not, by design, supposed to overcome a hard limitation inherent to the diffusion model. REPA-trained diffusion models for image generation can be matched by non-REPA trained models using 17.5x the training steps on image tasks (Figure 1, [1]); however, alignment can provide significant speedups to training. Furthermore, we test the benefit of combining representational alignment and entanglement targets for continuous dLMs using early-stop during inference, where we examine the effect of REPA+REG on the intermediate clean predictions for early-stop generation.

We emphasise our contributions:

  1. We develop and evaluate representation choices for adapting ELF to math reasoning and code generation. We use Qwen3-0.6B-Base as the text encoder and the larger Qwen3-1.7B-Base as the REPA+REG teacher, which belong to the same family and share a tokeniser. We adapt REG by using the teacher's last content-token state as a global target for the prompt and response. The Qwen3 encoder outperforms the default t5-small encoder when training ELF on GSM8K (Figure 4), and the last content-token target outperforms mean pooling for REG (Table 26). These choices establish an effective configuration that outperforms baselines at lower training and inference compute. With consistency sampling, ELF-REG-B achieves 37.91% GSM8K pass@1 at NFE 16, compared with 33.60% for FMLM+ (Init) at NFE 64, using less than half the training and inference FLOPs (Table 32). Table 31 also shows higher accuracy than ELF baseline with fewer training and inference FLOPs on the selected math and code benchmarks.
AuthorsOpenReview ↗

Response to reviewer tfNh (2)

  1. Regarding the application of REPA+REG, we investigate the contribution of its supervision and its interaction with early-stop:

    • Language reasoning allows us to distinguish answer correctness from convergence of the entire response: answers often agree with the fully denoised endpoint even while the decoded responses continue changing (Table 11). We investigate how REPA+REG interacts with early-stop in this setting. Compared with ELF-B baseline, ELF-REG-B exhibits lower acceleration, lower error relative to its trajectory endpoint, and earlier token stability (Figure 2). These observations connect representation supervision to the quality of intermediate clean predictions used for low-NFE reasoning. Our math-reasoning and code-generation results demonstrate useful early predictions without dedicated few-step training. For example, with early-stop ODE, ELF-REG-L achieves 46.39% MBPP-378 pass@10 at NFE 8, compared with 40.49% for PlaidQ-D16 at NFE 17 (Table 16). The original image-generation REPA and REG studies do not report this interaction between representation supervision and early-stop.

    • Adapting REG to an AR teacher requires choosing a global representation, since the teacher lacks the dedicated CLS token used by the image encoders considered in REG. Our ablation shows that the last content-token state works better than mean pooling (Table 26). We also identify a different role for REG in ELF-REG. At epoch 6, REPA improves ELF-B accuracy from 21.74% to 24.56%, and REPA+REG improves it further to 28.92%. Omitting REG at inference retains performance, at 28.66% (Table 5). This supports a benefit from global REG supervision beyond token-level REPA alone: the global REG target improves the shared text-generation weights during training, with little dependence on retaining the token during generation. This differs from the inference-time semantic guidance emphasised in the image-generation REG paper [2].

  2. In our theory contributions, we draw a connection between early-stop and the remaining denoising flow map using decoder-margin results. Our supporting analysis explains the relationship between clean-latent predictions during denoising, continued denoising, and decoded responses. Specifically, we characterise and bound the error incurred when an intermediate clean prediction replaces the remaining denoising flow, relate this error to decoded-output agreement, and separately derive the discrepancy for the ODE sampler.

In addition to the methodological novelty we explained above, we would like to emphasise the empirical benefit to the wider dLM research community in open-sourcing our training/inference code and checkpoints with SOTA performance compared to baselines.

[W2] [Q2] REG - REG contributes mainly during training.

Does REG mainly act as an additional training objective, or does the jointly denoised global representation provide something important during inference?

We agree with the reviewer's interpretation that the main REG benefit comes from additional training supervision, different from the interpretation of REG for image diffusion models [2]. Our results suggest that retaining the REG token during inference is unnecessary for most of the gain.

Table 5 compares the training components directly (epoch 6, early-stop ODE at NFE 64, and 16 generation seeds):

ELF-B configurationGSM8K pass@1 (%)
ELF-B baseline21.74
REPA-only24.56
REPA+REG28.92
Stripped REPA+REG28.66
REPA+OLT25.60

Stripped REPA+REG uses the same model weights but omits REG during generation. REPA+OLT adds one learnable token without the REG target. Adding REG to REPA improves accuracy by 4.36 percentage points. Most of this improvement remains after removing REG at inference, whereas adding an unsupervised learnable token does not reproduce it. This supports a contribution from REG supervision to the shared text-generation weights. REG alone without REPA gives little improvement in the setting reported in Table 26. Our claim concerns its benefit when combined with REPA.

AuthorsOpenReview ↗

Response to reviewer tfNh (3)

[W3] Teacher ablations - All three teacher configurations improve over the baseline.

I wonder whether the improvements would remain similar with a weaker teacher, a stronger teacher, or a teacher from another model family ...

We evaluated three teacher configurations while keeping the frozen text encoder fixed. All rows below use ELF-B at epoch 6, early-stop ODE at NFE 64, and four generation seeds.

Training configuration / teacherGSM8K pass@1 (%)Standard deviation (%)
ELF-B baseline, no teacher21.991.32
REPA+REG, Qwen3-0.6B (chat)25.441.08
REPA+REG, Qwen3-1.7B-Base29.210.96
REPA+REG, Qwen3-4B-Base27.081.62

All three teacher configurations improve over ELF-B baseline. The improvement therefore does not depend exclusively on the default Qwen3-1.7B-Base teacher. The default teacher Qwen3-1.7B-Base gives the highest mean accuracy, exceeding the performance using the larger Qwen3-4B-Base teacher. This is consistent with findings in iREPA [3] that larger teacher encoders can yield worse generation performance under representation alignment. (Note that the baseline and default-teacher rows reuse 4 of the 16 generation seeds in Table 5, which explains the different mean accuracy).

We restrict these ablations to Qwen teachers to keep tokenisation and the token-level alignment procedure fixed. REPA aligns ELF student and teacher representations at corresponding token positions. With a different tokenizer, the same text can be split into different tokens, so these positions would no longer correspond directly. Comparing such teachers would require an additional procedure for aligning representations across token boundaries. We leave cross-family teacher comparisons to future work; the present results establish robustness across the tested Qwen teacher configurations.

[W4] [Q4] Convergence and final performance - We improve ELF-REG's final accuracy & accuracy at lower training cost.

Does REPA+REG mainly improve training speed or final model performance?

The benefit of REPA+REG extends beyond early convergence. We added consistency-sampler evaluations of the existing final checkpoints, without additional training, demonstrating further improvements of ELF-REG compared to ELF baseline at the final checkpoints. Using the consistency sampler and same NFE to both ELF-REG and ELF-baseline:

Benchmark / architectureNFEELF baseline pass@1 (%)ELF-REG pass@1 (%)
GSM8K / ELF-B12841.3244.29
GSM8K / ELF-L12858.4061.50
MATH-500 / ELF-L12816.8520.15
HumanEval / ELF-L12828.3533.08
HumanEval+ / ELF-L12826.8331.40
MBPP-378 / ELF-L12836.5138.76
MBPP-500 / ELF-L12824.5525.60

These are 4-seed mean accuracy reproduced from Table 4. Additionally, the improvement also appears at other inference budgets: for example, GSM8K ELF-REG-B reaches 42.97% versus 40.28% at NFE 64 (Table 30).

Could the authors compare models at matched training budgets and show whether your performance gap remains after longer training?

We also evaluated earlier ELF-REG checkpoints against the final ELF baseline checkpoints, demonstrating ELF-REG's superior accuracy compared to ELF-baseline at less training/inference compute. Both models use consistency sampling and each pair below lists ELF baseline / ELF-REG, bolding better performance or less compute/latency within each pair:

BenchmarkNFEPass@1 (%)Training FLOPs (×1020\times 10^{20})Training GPU-hoursInference TFLOPs/answer
GSM8K / ELF-B64 / 4740.28 / 41.580.9131 / 0.5081170.8 / 117.415.339 / 11.579
MATH-500 / ELF-L128 / 11216.85 / 17.403.961 / 3.205449.6 / 430.2191.466 / 167.856
HumanEval / ELF-L128 / 6428.35 / 29.882.783 / 2.170317.3 / 305.9191.466 / 96.405
MBPP-378 / ELF-L128 / 11236.51 / 36.642.783 / 2.170317.3 / 305.9191.466 / 167.856

These comparisons are also in Table 31. HumanEval+ also improves from 26.83% to 28.20%, using the same training and inference costs as the HumanEval comparison (Table 31). We demonstrate ELF-REG's better performance with fewer FLOPs and lower latency: ELF-REG uses 44.3% fewer training FLOPs on GSM8K, 19.1% fewer on MATH, and 22.0% fewer on code. The MATH checkpoint also uses 430.2 versus 449.6 training GPU-hours, including GSM8K initialisation; the code checkpoint uses 305.9 versus 317.3 GPU-hours.

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

This paper proposes ELF-REG, extending Embedded Language Flows (ELF) to mathematical reasoning and code generation. The method combines representation alignment with a jointly denoised global representation token, supervised by a frozen autoregressive teacher. The authors further investigate early-stop decoding to reduce inference steps without dedicated few-step training.

Critical strengths
  1. ELF-REG improves over the ELF baseline across GSM8K, MATH-500, HumanEval, and MBPP. The component ablations also provide evidence for the benefit of combining REPA and REG.

  2. The paper demonstrates that useful reasoning accuracy can emerge well before the denoising trajectory converges. The analysis of endpoint error and decoder stability provides additional insights into why early-stop decoding can achieve competitive performance with fewer inference steps.

Critical weaknesses
  1. Limited methodological novelty. The main method combines REPA and REG, both established techniques from diffusion modeling, with an existing ELF architecture. Early-stop generation has also been investigated in prior work. While the combination is effective, the paper introduces relatively limited new methodological insights beyond adapting these components to language reasoning tasks. The theoretical analysis largely builds on existing flow-map and decoder-margin results.

  2. The paper emphasizes improved efficiency, but NFE does not directly translate into wall-clock latency or FLOPs across different architectures. ELF-REG also relies on a pretrained Qwen3 encoder at inference and an additional AR teacher during training, whose costs are not fully reflected in the headline comparisons.

  3. Table 7 shows that ELF-REG requires more training GPU-hours than ELF on code (407.9 vs. 317.3) and MATH (563.1 vs. 449.6, including initialization). Thus, the efficiency benefits appear task-dependent。

Questions for authors
  1. Can the authors clarify the methodological novelty beyond combining REPA, REG, and early-stop decoding?

  2. Which component contributes most to the improvements at larger model scales?

  3. Could the authors provide end-to-end latency and FLOPs comparisons with competing diffusion models and autoregressive baselines, including the frozen encoder cost?

6 comments
AuthorsOpenReview ↗

Response to reviewer yxQJ (1)

We added consistency-sampler results and comparisons of accuracy at lower training and inference cost to the updated version of the paper. Below, we report encoder-inclusive FLOPs and H200 generation time, together with earlier-checkpoint results that address the training GPU-hour concern. Additions to the paper are coloured blue.

[Strength 1] ELF-REG improves over the ELF baseline

To further demonstrate the superior performance of our recipe, we added consistency-sampler results without additional training. Using the consistency sampler and same NFE, ELF-REG-L outperforms ELF-L baseline across all six benchmarks (Table 4). ELF-REG's highest pass@1 accuracy using the consistency sampler are 61.50% on GSM8K (NFE 128) and 20.15% on MATH-500 (NFE 128), compared with 58.40% and 16.85% for the ELF baseline. On HumanEval and HumanEval+, ELF-REG reaches 33.08% and 31.40% (both NFE 128), versus 28.35% and 26.83%. On MBPP-378 and MBPP-500, ELF-REG reaches 38.76% and 25.60% (both NFE 128), versus 36.51% and 24.55%. Please see Table 4 for more details.

[W1] [Q1] Limited novelty - We develop an effective & efficient recipe for continuous reasoning.

Can the authors clarify the methodological novelty beyond combining REPA, REG, and early-stop decoding?

We thank the reviewer for raising this concern. We address this concern in the rebuttal and edit the paper's introduction to emphasise our novelty.

Our main contribution is building an effective recipe for math reasoning and code generation with fully continuous dLMs, using frozen AR representations, whereas ELF using the bidirectional t5-small encoder did not perform well in our ablations (Figure 4). Our aim is to improve reasoning accuracy with limited training and inference compute to narrow the gap between continuous dLMs and AR LLMs. Empirically, REPA+REG improves reasoning performance over the ELF baseline and prior works with less compute cost, and ELF-REG models support strong low-NFE early-stop generation without dedicated few-step training. We note that representation alignment methods are not, by design, supposed to overcome a hard limitation inherent to the diffusion model. REPA-trained diffusion models for image generation can be matched by non-REPA trained models using 17.5x the training steps on image tasks (Figure 1, [1]); however, alignment can provide significant speedups to training. Furthermore, we investigate how representation supervision affects the intermediate clean predictions used for early-stop in language denoising.

We emphasise our contributions:

  1. We develop and evaluate representation choices for adapting ELF to math reasoning and code generation. We use Qwen3-0.6B-Base as the text encoder and the larger Qwen3-1.7B-Base as the REPA+REG teacher, where both belong to the Qwen3 family and use shared tokenisation. We adapt REG by using the teacher's last content-token state as a global target for the prompt and response. Figure 4 compares the Qwen3 and default t5-small encoder configurations where Qwen3 encoder outperforms t5-small for training ELF on GSM8K, and Table 26 shows that the last content-token target outperforms mean pooling for REG. These choices establish an effective configuration that out-performs baselines at lower training and inference compute budgets. Our central aim is to narrow the reasoning-performance gap between fully continuous dLMs and strong AR models, while improving accuracy at lower compute than prior dLMs. ELF-REG-B achieves 37.91% GSM8K pass@1 at NFE 16, compared with 33.60% for FMLM+ (Init) at NFE 64, where ELF-REG-B uses less than half the training and inference FLOPs (Table 32). Table 31 also shows higher accuracy than ELF baseline with fewer training and inference FLOPs on the selected math and code benchmarks.
AuthorsOpenReview ↗

Response to reviewer yxQJ (2)

  1. Regarding the application of REPA+REG, we believe that the combination solves important issues:
    • Language reasoning allows us to distinguish the answer's correctness from convergence of the entire response: answers often agree with the fully-denoised endpoint even while the decoded responses continue changing (Table 11). We investigate how REPA+REG interacts with early-stop in this setting. Compared with ELF-B baseline, ELF-REG-B exhibits lower acceleration, lower error relative to its trajectory endpoint, and earlier token stability (Figure 2). These observations connect representation supervision to the quality of intermediate clean predictions used for low-NFE reasoning. Our math-reasoning and code-generation results demonstrate useful early predictions without dedicated few-step training. For example, with early-stop ODE, ELF-REG-L achieves 46.39% MBPP-378 pass@10 at NFE 8, compared with 40.49% for PlaidQ-D16 at NFE 17 (Table 16). The original image-generation REPA and REG studies do not report this interaction between representation supervision and early-stop.
    • Adapting REG to an AR teacher also requires choosing a global representation, since the teacher lacks the dedicated CLS token used by the image encoders considered in REG. Our ablation shows that the last content-token state works better than mean pooling (Table 26). We also identify a different role for the REG token in ELF-REG. The image-generation REG paper [2] emphasises semantic guidance during inference, whereas in our setting, REPA+REG outperforms REPA-only, and most of the additional gain from REPA+REG training is maintained even if we omit REG at inference (Table 5). Thus, the global REG target improves the shared text-generation weights during training, with little dependence on retaining the token during generation. This provides a methodological insight into using global representation supervision for continuous dLMs.
  2. Regarding theory: we draw a theoretical connection between early-stop and flow map using decoder-margin results, which is a novel contribution for continuous dLMs. In our supporting theoretical explanations, we explain the relationship between clean-latent predictions during denoising, continued denoising, and decoded responses, which has not been explored in prior literature. Specifically, our analysis characterises the error incurred when an intermediate clean prediction replaces the remaining denoising flow, and we bound this error, relate this error to decoded-output agreement, and separately derive the discrepancy for the ODE sampler.

In addition to the methodological novelty we explained above, we would like to emphasise the empirical benefit to the wider dLM research community in open-sourcing our training/inference code and checkpoints with SOTA performance compared to baselines.

[W2] [Q3] Inference costs - We include encoder costs and measure FLOPs & generation time.

... provide end-to-end latency and FLOPs comparisons ... including the frozen encoder cost?

We agree that NFE alone does not establish efficiency. We present results that show ELF-REG performing better than other baselines at lower inference/training FLOPs and lower wall-clock latency. The tables below report accuracy, FLOPs and generation time when comparing ELF-REG to baselines; we have also added the accuracy/FLOP comparisons to the paper Table 32. Our inference FLOPs include the frozen encoder, denoising, and decoding, whereas inference timings also include tokenisation. Model loading, disk IO warmup, file writes, and answer scoring are not included in the timed interval. The REPA/REG teacher supplies training targets and is not used at inference.

We measure latency as seconds per answer on one H200, using three batches of 32 questions per batch after one warmup batch. Each method uses its native inference setup and ELF-REG uses the consistency sampler.

GSM8K:

ModelNFEPass@1 (%)Training FLOPs (×1019\times 10^{19})TFLOPs/answerSeconds/answer
ELF-REG-B, epoch 91637.915.0814.6910.0347
FMLM+ (Init)6433.60≈\approx 13.99311.6160.1180
MDLM1024≈\approx 33.905.196138.5551.3544
Duo1024≈\approx 36.025.196138.5541.3919
S-FLM (top-1)512≈\approx 18.185.19569.2721.1395
AuthorsOpenReview ↗

Response to reviewer yxQJ (3)

Code (HE = HumanEval; HE+ = HumanEval+):

ModelNFEHE (%)HE+ (%)MBPP-378 (%)TFLOPs/HETFLOPs/MBPP-378Seconds/HESeconds/MBPP-378
ELF-REG-L, epoch 121624.2423.1727.1824.95424.9540.17660.1756
Open-dCoder12820.8017.6023.9033.94631.7190.49010.5056
oDLM12817.8716.4022.9942.12739.3230.65610.6748
PlaidQ12919.0218.0522.18≈\approx 42.654≈\approx 39.816----
PlaidQ+CFG25722.0419.7024.58≈\approx 84.977≈\approx 79.324----

Bold marks the highest accuracy or lowest latency/FLOPs per benchmark. We take performance values of baselines from their papers and ≈\approx denotes estimated values or accuracy read from figures. ELF performance is mean over 4 generation seeds, using EvalPlus for code. FMLM+ training FLOPs includes MDLM initialization. PlaidQ FLOPs are estimated from the architecture due to the code being not accessible (returns 404). Unmeasured timings for PlaidQ are marked -- due to PlaidQ not publishing code where the github link returns HTTP 404.

Shown in the table above, ELF-REG has the highest accuracy while using less inference FLOPs and latency against FMLM+ (Init), MDLM and Duo on GSM8K. On code, ELF-REG also shows best accuracy at lowest inference FLOPs and latency.

For ELF-REG-B at NFE 64, the overall 15.356 TFLOPs per answer comprise 0.816 for Qwen encoding, 13.998 for denoising and 0.542 for the final transformer+decoder call. Encoding accounts for 5.3% of the total at this setting.

... an additional AR teacher during training, whose costs are not fully reflected ...

Our reference training FLOPs count all ELF/ELF-REG training computation plus one full-dataset encoder/teacher feature-preparation pass per training history. This assumes that frozen features are computed once and cached for reuse. Assuming batch size 32, GSM8K feature preparation using Qwen costs 4.786×10184.786 \times 10^{18} FLOPs for ELF-REG, including both encoder and REPA/REG teacher, versus 1.962×10181.962 \times 10^{18} for ELF baseline which does not use the REPA/REG teacher. The epoch 9 ELF-REG-B checkpoint requires 5.081×10195.081 \times 10^{19} training FLOPs, compared with 5.196×10195.196 \times 10^{19} for MDLM/Duo and approximately 13.99×101913.99 \times 10^{19} for FMLM+ (Init). We note that these comparisons may use different training datasets.

We report autoregressive baselines separately:

BenchmarkAR baselinePass@1 (%)TFLOPs/answerSeconds/answer
GSM8KQwen3-0.6B-Base39.200.22910.3842
GSM8KQwen3-1.7B-Base56.250.55830.1120
HumanEvalQwen3-0.6B-Base21.341.06850.5654
HumanEvalQwen3-1.7B-Base38.412.45730.5743
MBPP-378Qwen3-0.6B-Base33.330.21060.0931
MBPP-378Qwen3-1.7B-Base59.790.41470.0626

We evaluated Qwen locally using one seed, temperature 0.8 and top-p 0.95, with EvalPlus for code evaluation. Latency uses KV caching and the same settings as the dLM measurements (batch size 32, measured over 3 batches). On HumanEval, ELF-REG-L exceeds Qwen3-0.6B-Base in accuracy (24.24% versus 21.34%) and has lower latency (0.1766 versus 0.5654 seconds/answer).

[W3] Training efficiency - We improve accuracy with fewer training GPU-hours and FLOPs.

Table 7 shows that ELF-REG requires more training GPU-hours than ELF on code ... and MATH ...

We present results that show ELF-REG performing better than ELF baseline at lower inference/training FLOPs and lower training GPU-hours. We evaluated earlier ELF-REG-L checkpoints, comparing against the ELF-L baseline to compare performance where ELF-REG uses fewer training FLOPs, which we report in the table below (this corresponds to Table 31 in the updated draft). Both models use consistency sampling and we compute mean accuracy over 4 generation seeds. For code evaluation, we use our headline settings.

BenchmarkModel, epochNFEPass@1 (%)Training FLOPs (×1020\times 10^{20})Training GPU-hoursInference TFLOPs/answer
MATH-500ELF-L baseline, 812816.853.961449.6191.466
MATH-500ELF-REG-L, 511217.403.205430.2167.856
HumanEvalELF-L baseline, 1212828.352.783317.3191.466
HumanEvalELF-REG-L, 96429.882.170305.996.405
MBPP-378ELF-L baseline, 1212836.512.783317.3191.466
MBPP-378ELF-REG-L, 911236.642.170305.9167.856

Within each benchmark, we bold the highest accuracy, lowest training FLOPs/GPU-hours, and lowest inference FLOPs.

Comments on the paper

1 comment

We thank the reviewers and AC for considering our submission.

Strengths noted by reviewers

Reviewers highlighted ELF-REG's gains over ELF baseline on math and code, and its ability to produce useful answers before denoising converges without dedicated few-step training with early stop.

Reviewer yxQJ highlighted the gains over ELF baseline across math and code. Reviewer Xn8R described ELF-REG as a "hero run" for bringing fully continuous dLMs closer to autoregressive LLM performance. Without additional training and with consistency sampling, ELF-REG-L reaches 20.15% pass@1 on MATH-500 and 33.08% on HumanEval at 128 NFE (Table 4).

Reviewers yxQJ and tfNh highlighted our analysis of endpoint error and decoder stability. On GSM8K, ELF-REG-B reaches near-final accuracy at 64 NFE even though only 8.61% of complete responses match those at 512 NFE (Table 11).

Reviewer yxQJ raised their rating after confirming that the rebuttal addressed their concerns.

Novelty

ELF-REG develops a SOTA recipe involving token-level and global teacher supervision to train fully continuous dLMs for math reasoning and code generation, surpassing baselines at lower training/inference FLOPs. REPA supervises individual token positions and for REG, the denoiser learns to reconstruct the teacher's last content-token state alongside the text. Our ablation shows that most of REG's gain over REPA alone persists when we omit the REG token at inference (Table 5).

As far as we know, we are the first work demonstrating early-stop performance on reasoning tasks with fully continuous dLMs. In early-stop, reasoning responses reach near-final accuracy before the full response stabilises. Early-stop achieves this at one eighth of the full inference time schedule (Table 11). Without dedicated few-step training, ELF-REG-L reaches 46.39% pass@10 on MBPP-378 at 8 network function evaluations (NFE). PlaidQ-D16 reaches 40.49% at 17 NFE (Table 16). All three reviewers noted early-stop for our reasoning scenario as a strength.

Fair comparisons with baselines

ELF-REG-B outperforms other GSM8K baselines with fewer training & inference FLOPs, outperforming FMLM+ (Init), MDLM, Duo and S-FLM (top-1). For example, ELF-REG-B reaches 37.91% pass@1 versus 33.60% for FMLM+ (Init), using less than half the training and inference FLOPs (Table 32).

ELF-REG-L outperforms baseline dLMs on coding benchmarks with fewer inference FLOPs. On HumanEval, ELF-REG-L reaches 24.24% pass@1 versus 22.04% for PlaidQ+CFG, where ELF-REG-L uses less than a third of PlaidQ+CFG's estimated inference FLOPs (Table 32). Under the same code evaluation settings, ELF-REG-L scores 25.79% pass@1 versus Open-dCoder's reported 23.90% for MBPP-378 (response to tfNh).

ELF-REG improves GSM8K pass@1 at matched training FLOPs compared to ELF-baseline. In response to Xn8R and tfNh, we held the training data, architecture and training FLOPs fixed within each isoFLOP comparison. At the largest training FLOP budget, ELF-REG-L reaches 52.18% versus 43.33% for ELF-L baseline (Figure 10).

Plaid [1] uses isoFLOP analysis to study language-model likelihood and [2] show that lower perplexity can correspond to worse downstream performance. We therefore measure GSM8K accuracy directly and do not perform scaling law experiments using perplexity with generic corpuses.

[1] Likelihood-Based Diffusion Language Models, https://arxiv.org/abs/2305.18619

[2] Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check, https://arxiv.org/abs/2507.00885

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.