acceptodds
Under review as a conference paper at ICLR 2027

Enhancing Diffusion Language Models with Autoregressive Post-Training Weights

Abstract

Diffusion language models (dLLMs) have emerged as a promising alternative to autoregressive (AR) language models, offering flexible token-update orders and parallel decoding. Recent dLLMs are often initialized from pretrained AR models before diffusion conversion to inherit their learned representations. After the conversion, however, they typically discard the extensive post-training ecosystem of their AR ancestors. In this work, we show that these existing AR post-training weight updates can instead be effectively recycled to enhance diffusion models, despite the changes by AR-to-diffusion conversion. Directly adding an AR post-training weight update to a diffusion base model remains effective, bringing its performance close to that achieved by direct diffusion post-training. Notably, AR and diffusion post-training updates are nearly orthogonal in weight space, yet induce substantially more aligned representation changes in the diffusion model. Their distinct updates are also complementary: composing their weights can retain gains from both and further improve the diffusion-post-trained model. Based on these findings, we propose A2D, a simple training-free framework for enhancing diffusion models with existing AR post-training resources. A2D can transfer capabilities from AR post-trained models to diffusion base models, and further improve already post-trained diffusion models by composing AR and diffusion post-training updates. Across Dream, DreamReasoner, DiffuCoder, Nemotron-Labs-Diffusion, and DiffusionGemma, A2D reliably improves instruction following, mathematical reasoning, and coding with both supervised fine-tuning and reinforcement learning updates, without additional training, or inference-time computation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion Language Models, Model Merging, Post-Training

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xNJ5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether post-training updates learned by an autoregressive (AR) language model remain useful after that model has been converted into a diffusion language model. A2D applies an AR task vector to the diffusion base model and, when a diffusion-post-trained checkpoint is available, interpolates the AR and diffusion updates. Both operations work directly in weight space and require no further training.

Critical strengths
  1. The question is timely and practically motivated. Diffusion models often inherit AR pretrained weights, while useful AR post-trained checkpoints are also available but are seldom reused after conversion. The proposed method is simple and has no additional inference-time computation once the weights are merged.

  2. The controlled experiments train matched AR and diffusion models on the same math or code data, making a reasonable case that the AR update itself can survive diffusion conversion. The finding that the two updates are nearly orthogonal in parameter space but induce more aligned representation changes is also interesting.

Critical weaknesses
  1. The methodological novelty of A2D-Transfer is limited relative to AR-MAP[1], which already applies an AR post-training task vector to diffusion checkpoints. The broader evaluation is valuable, but the paper should distinguish this empirical extension more clearly from the previously established transfer operation.

  2. A2D-Merge is a linear interpolation of two weight updates. Its interesting contribution is the observed compatibility of AR and diffusion updates for the same capability, rather than a new merging algorithm. Comparisons with adding a tuned AR update directly to the post-trained diffusion model across non-anchor benchmarks would better isolate this claim.

Lin, Liang, et al. "AR-MAP: Are Autoregressive Large Language Models Implicit Teachers for Diffusion Large Language Models?." arXiv preprint arXiv:2602.02178 (2026).

Questions for authors
  1. What precisely does A2D-Transfer establish beyond AR-MAP? A comparison on a shared Dream setting or a more explicit discussion of why the settings cannot be matched would help define the incremental contribution.
  2. How does A2D-Merge compare across the non-anchor benchmarks with simply adding a tuned AR vector to the already post-trained diffusion checkpoint? The appendix studies this alternative on anchor tasks, but a full side-by-side comparison would show what replacing part of the diffusion update contributes.
3 comments

We thank the reviewer for their valuable feedback and recognition of our method's simplicity, practical motivation, and interesting findings. Below, we address their comments and questions.

W1/Q1: Relationship and comparisons to AR-MAP.

We thank the reviewer for highlighting this connection. We acknowledge and respect AR-MAP’s contribution to transferring AR updates in the DPO setting and explicitly discussed it in the manuscript. A2D-Transfer and AR-MAP both build on Task Arithmetic and share the same task-vector addition operation. Our main contribution, as correctly summarized by the reviewer, lies in systematically establishing and studying the recycling and compatibility of AR post-training updates to enhance diffusion models, rather than introducing a new merging method.

However, the two works address different research questions and differ in their scope. AR-MAP targets the difficulty of estimating likelihoods for diffusion during specifically the DPO stage: its workflow trains an AR donor with DPO and transfers the update as an alternative to diffusion-specific preference training. In contrast, the main motivation for A2D is to leverage the rich ecosystem of existing AR post-training checkpoints to further enhance diffusion models across post-training stages without training. Our extensive evaluations demonstrate the broad applicability of A2D across model scales, diverse post-training paradigms including instruction tuning, reasoning, and RL, and specialized domains such as coding and medical VQA.

Importantly, A2D goes beyond empirical demonstration of transferring by systematically investigating the relationship and complementarity between the two post-training processes. As the reviewer also appreciates, our controlled experiments reveal that the two updates are nearly orthogonal in parameter space yet induce more aligned representation changes. These findings motivate A2D-Merge, which combines independently acquired post-training gains to further improve diffusion-post-trained models, rather than merely replacing diffusion-specific training. This progression from transfer to complementary post-training composition is central to A2D's contribution.

We illustrate the distinctions in the shared Dream model family. AR-MAP trains an AR DPO update and transfers it to Dream-7B-Instruct for preference alignment. In contrast, A2D aims to leverage the broader existing AR post-training ecosystem, directly reusing released checkpoints such as Qwen2.5-7B-Instruct to transfer instruction-tuning capabilities to Dream-7B-Base. Going one step further, A2D-Merge combines this AR update with Dream's own instruction-tuning update, achieving further improvements over both A2D-Transfer and Dream-7B-Instruct.

Following the reviewer’s suggestion, we will add a dedicated section in the revised manuscript to more clearly compare and distinguish A2D from AR-MAP.

AuthorsOpenReview ↗

Rebuttal - part 2

W2/Q2: Direct addition versus A2D-Merge.

We thank the reviewer for this insightful suggestion.

Comparison with direct addition. Following the notation in our manuscript, we refer to the baseline for adding an AR update to the post-trained diffusion model as “+τ_A only”. Based on the reviewer’s suggestion, we extended these ablations to non-anchor benchmarks, including diffusion-only downscaling as a control (denoted as “−τ_D only”). The results are summarized below. Results cover 7 anchor and 21 non-anchor model–benchmark evaluations with complete ablations. Wins indicate strict improvements over the diffusion parent.

MethodAnchor avg.Non-anchor avg.Overall avg.Wins vs. parent
Diffusion parent72.352.257.3—
+τ_A only75.855.160.319/28 (67.9%)
−τ_D only74.152.958.220/28 (71.4%)
A2D-Merge76.855.360.727/28 (96.4%)

On non-anchor benchmarks, A2D-Merge achieves higher average performance than +τ_A only (55.3 vs. 55.1). More importantly, its gains are substantially more consistent: A2D-Merge improves upon the diffusion parent in 20/21 evaluations (95%), tying it in the remaining one, whereas +τ_A only improves performance in only 14/21 (67%).

Across all 28 evaluations, both +τ_A only and A2D-Merge provide larger average gains over the parent than diffusion-only downscaling (+3.0 and +3.4 vs. +0.9 percentage points), highlighting the importance of incorporating AR updates. Globally, A2D-Merge achieves the highest overall average performance and improves upon the parent in 27/28 evaluations, compared with 19/28 improvements for +τ_A only and 20/28 for downscaling alone.

Notably, the +τ_A only baseline also falls within our framework of AR-to-diffusion post-training merging. It combines the same two updates but adopts a different merging strategy that fixes the diffusion update coefficient at 1. These results support our design choice of jointly rebalancing AR and diffusion updates for more consistent gains. We will include these expanded anchor and non-anchor ablations in the revised manuscript.

Choice of linear interpolation. While A2D uses simple and direct linear interpolation, we have additionally experimented with more advanced model merging techniques including TIES [1] and DARE [2], and the results demonstrate that these methods do not further enhance the performance of A2D, partially supporting the observed compatibility of AR and diffusion updates, and our choice for linear interpolation.

Modelτ_D / τ_A weightsIFEvalMATH-500MBPP+Average
Transfer
Dream-7B Base0 / 049.720.859.343.3
TIES0 / 0.955.536.461.951.3
DARE0 / 0.860.335.460.652.1
A2D (transfer)0 / 0.959.240.262.754.0
Merge
Dream-7B Instruct1 / 063.038.457.152.8
TIES0.8 / 0.263.643.857.955.1
DARE0.1 / 0.947.335.861.648.2
A2D (merge)0.4 / 0.672.845.058.758.8

We hope this effectively addresses the reviewer’s concerns and remain available for any further discussion or clarification.

[1] TIES-Merging: Resolving Interference When Merging Models, Yadav et al., NeurIPS 2023.

[2] Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch, Yu et al., ICML 2024.

AuthorsOpenReview ↗

Gentle Follow-up

Dear Reviewer,

We wanted to follow up on whether our responses have addressed your concerns. We would be happy to address any remaining questions before the discussion period ends.

Thank you again for your valuable feedback and for your time.

Best regards, The Authors

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

This work proposes A2D, a training-free framework that transfers autoregressive (AR) post-training updates to compatible diffusion language models (DLLMs) or interpolates them with diffusion post-training updates. Experiments across six model families demonstrate gains in instruction following, mathematical reasoning, coding, and medical VQA without changing the diffusion architecture or decoding procedure. Further analyses show that nearly orthogonal parameter updates can induce related representational changes and complementary behavioral gains.

Critical strengths
  1. The paper addresses an interesting research topic.

  2. Although AR-MAP has used model merging to align diffusion language models with AR models, this work presents several empirical findings that help us better understand the relationship between AR models and DLLMs.

  3. The method is simple and effective, and the paper is well written and easy to understand.

Critical weaknesses
  1. The performance improvement cannot be predicted in advance. The method is largely heuristic, and there is no accuracy-based indicator to guide how to merge an AR model with a DLLM to achieve the best performance.

  2. Model merging is currently limited to models within the same model series. If the DLLM has not undergone continued pretraining from the AR checkpoint, merging the models may not improve performance.

Questions for authors
  1. What happens if we merge a DLLM with an AR model when the DLLM has not undergone continued pretraining from the AR checkpoint? The merge may fail because of differences in their architectures, training corpora, and tokenizers. How can we avoid such failures, or what fully cross-model method could we propose to address this problem?

  2. Dream is evaluated using one token per decoding step. Will merging harm the model's parallel decoding capability? What about its bidirectional reasoning ability?

3 comments

We thank the reviewer for their valuable feedback and recognition of our research topic, our method's simplicity and effectiveness, the clarity of our writing, and the useful empirical insights into the relationship between AR models and dLLMs. Below, we address their comments and questions.

W1: Accuracy-based indicator and predicting the gain in advance.

Accuracy-based indicator. We would like to clarify that A2D already employs an accuracy-based criterion for selecting the merging coefficients. Specifically, for each model pair, we select the coefficients using a single anchor benchmark and keep them fixed across all other evaluated benchmarks. We would appreciate the reviewer's further clarification if they are indicating a different question.

Predicting the gain in advance. Our evaluation follows the established practice in model merging literature [1-3], where merging performance is validated empirically using accuracy-based metrics, rather than predicted in advance. Furthermore, our controlled experiments reveal nearly orthogonal AR and diffusion task vectors, which is commonly considered a useful indicator for potentially favorable merging performance. Our empirical results are consistent with this indication: A2D-Merge improves performance in all 7 evaluated settings and on 27 out of 28 benchmarks (the remaining benchmark results in a tie).

W2/Q1: Generalization to independently pretrained models and different architectures.

We thank the reviewer for raising this question. We agree that directly merging independently pretrained models may fail due to misaligned parameter spaces. Indeed, traditional model merging methods typically assume a shared initialization [1-3]. A2D specifically targets the practically important setting where diffusion models inherit pretrained AR weights, as is the case for many recent diffusion models, such as Dream, DreamReasoner, DiffuLLaMA, DiffuCoder, LLaDA2.0, SDAR, dQWEN3.5, and Nemotron-Labs-Diffusion. Their AR donor models have rich post-training ecosystems, providing extensive resources that A2D can directly reuse.

Extending model merging to independently pretrained models presents substantial challenges and remains an open problem in the broader model merging literature, rather than a limitation specific to A2D. Recent work for model merging has specifically investigated this challenge with dedicated efforts [4]. However, motivated by the reviewer's suggestion, we revisit our existing Nemotron-LD experiments and conduct new experiments in two increasingly challenging settings to systematically investigate the generalization boundaries of A2D.

Slightly different architectures, shared pretraining. We first highlight that A2D has already been evaluated on models with architectural differences, as demonstrated by our Nemotron-Labs-Diffusion experiments (Appendix B.2). Compared with its Ministral 3 AR donor, Nemotron differs in its diffusion-specific output head, RoPE configuration, and special-token assignments. We handle these differences by mapping compatible Transformer parameters, preserving diffusion-specific components, and excluding mismatched embedding rows from the transferred updates. With this mapping, A2D improves 9/10 benchmarks with Transfer and improves or matches all 10 with Merge.

Same architecture, independent pretraining. We then investigate transfer between AR and diffusion models with the same architecture but independently pretrained from scratch, using the ARM and SMDM models from [5]. This setting isolates the effect of different pretraining initializations. By transferring an AR instruction-following update to SMDM, we improve IFEval accuracy from 7.0% to 11.0%. Although the absolute performance remains limited, this result provides encouraging evidence that AR-to-diffusion transfer can potentially extend beyond shared pretrained initialization.

AuthorsOpenReview ↗

Rebuttal - part 2

Different architectures, independent pretraining. We further investigate a substantially more challenging setting where AR and diffusion models are independently pretrained with different architectures and model sizes. To explore the generalization boundaries of A2D under these conditions, we incorporate THESEUS [4], a dedicated cross-architecture task-vector transport method through representation alignment, to transfer the instruction-tuning update from Qwen2.5-1.5B to Dream-7B-Base. Despite these challenges, we observe improvements in both instruction following (IFEval: 49.9% → 52.0%) and mathematical reasoning (MATH-500: 21.4% → 23.4%), with the average performance increasing from 43.6% to 44.9%. These results provide initial evidence that A2D can be potentially combined with cross-architecture task-vector transport techniques to enable broader AR-to-diffusion transfer, even without shared initialization or architecture.

AR donorIFEvalMATH-500MBPP+Average
Dream-7B-Base49.921.459.543.6
+Qwen2.5-1.5B-Instruct update (THESEUS)52.023.459.344.9

We will include all additional experiments, along with their implementation details, in the revised manuscript.

W2/Q2: Parallel decoding and bidirectional reasoning.

We thank the reviewer for raising these important questions. Our existing analyses already provide evidence that A2D preserves diffusion-specific capabilities. As shown in Fig. 5(a), A2D does not systematically shift the decoding order toward autoregressive generation. Furthermore, Table 20 demonstrates that A2D remains effective under self-speculative decoding on Nemotron-LD, which drafts 32-token blocks in parallel before autoregressive verification, with A2D consistently increasing tokens per forward evaluation.

To further address the reviewer's concerns, we conducted two additional experiments directly evaluating parallel decoding and bidirectional reasoning.

Parallel decoding. We evaluate Dream-7B on MATH-500 with fewer diffusion decoding steps, allowing multiple tokens to be decoded in parallel. As shown in the table below, A2D-Merge continues to outperform the diffusion parent with 2 tokens per step (32.8 vs. 28.8), demonstrating that its gains extend to parallel decoding. At higher parallelism, performance degrades sharply for all models, including the original Dream. These results suggest that A2D preserves parallel decoding benefits at moderate parallelism.

Model512 steps (1 tok/step)256 (2)128 (4)64 (8)
Parent38.428.810.20.8
A2D-Transfer40.232.210.20.0
A2D-Merge45.032.89.20.2

Bidirectional reasoning. We further evaluate A2D on Dream's 4×4 Sudoku planning task, which requires reasoning over bidirectional context. Following the original Dream setup, we use the official evaluation script with 8-shot prompting, 24 denoising steps, greedy decoding, 10 given cells, and 100 puzzles.

ModelSudoku
Qwen2.5-7B (AR)18.0
Qwen2.5-7B-Instruct (AR donor)23.0
Transfer
Dream-7B Base72.0
A2D (transfer)70.0
Merge
Dream-7B Instruct (parent)89.0
A2D (merge)82.0

Although A2D exhibits some performance decreases relative to its diffusion parents (70.0 vs. 72.0 for Transfer; 82.0 vs. 89.0 for Merge), both models retain strong Sudoku performance, substantially outperforming the AR models (18.0–23.0). Notably, Sudoku represents a specialized constraint-solving task that was neither targeted by the AR post-training updates nor used for merging coefficient selection. These results, together with our decoding-order analysis (Fig. 5(a)), suggest that A2D largely preserves Dream's bidirectional reasoning capabilities, despite some task-specific performance trade-offs.

We hope this effectively addresses the reviewer’s concerns and remain available for any further discussion or clarification.

[1] Editing Models with Task Arithmetic, Ilharco et al., ICLR 2023.

[2] TIES-Merging: Resolving Interference When Merging Models, Yadav et al., NeurIPS 2023.

[3] Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch, Yu et al., ICML 2024.

[4] Transporting Task Vectors across Different Architectures without Training, Rinaldi et al., ICML 2026.

[5] Scaling up Masked Diffusion Models on Text, Nie et al., ICLR 2025

AuthorsOpenReview ↗

Gentle Follow-up

Dear Reviewer,

We wanted to follow up on whether our responses have addressed your concerns. We would be happy to address any remaining questions before the discussion period ends.

Thank you again for your valuable feedback and for your time.

Best regards, The Authors

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

This paper studies whether post-training updates from AR models can be reused in diffusion language models converted from the same AR base. It proposes A2D, a training-free method with two modes. (1) Transfer adds the AR post-training vector to the diffusion base, and (2) Merge interpolates between the AR and diffusion post-training vectors on top of the diffusion post-trained model. Across six diffusion model families and several post-training settings, A2D improves instruction following, math and code, and the authors attribute the gains to complementary AR and diffusion post-training.

Critical strengths

The paper addresses a practical question with a simple and training-free method. The evaluation is broad. It covers six diffusion model families, both text and multimodal tasks, models up to 26B, and a cross-architecture case that requires parameter mapping. The analyses are also informative, including the asymmetry of transfer, the failure of continued-pretraining vectors, and the finding that A2D mainly changes low-confidence tokens.

Critical weaknesses
  • Missing some baselines discussed by the authors. The related work discusses merging methods that reduce interference, such as TIES [1] and DARE [2], and names AR-MAP [3] as the closest prior work, but none of them appear in the experiments. A2D-Merge is a linear interpolation, and the paper’s main claim is that τA\tau_A and τD\tau_D are complementary with limited interference. A comparison with TIES or DARE would directly test this claim. AR-MAP also transfers AR weight updates to diffusion models through weight scaling, which is very close to A2D-Transfer. I would ask the authors to add these comparisons, or to explain why they are not included and how A2D-Transfer differs from AR-MAP.

  • The gains may come from a stronger donor rather than complementarity. In most main experiments, the AR donor is much stronger than the diffusion parent (e.g., IFEval 75.3 vs. 63.0 for Dream-7B). Table 4 and Table 12 also show that, for some families, downscaling τD\tau_D alone gives most of the merge gain. These results suggest that much of the improvement may come from importing a stronger post-training update. I would ask the authors to report the “+τA+\tau_A only” and “−τD-\tau_D only” variants on more benchmarks in the main table.

[1] https://arxiv.org/abs/2306.01708
[2] https://arxiv.org/abs/2311.03099
[3] https://arxiv.org/abs/2602.02178

Questions for authors
  • How does A2D compare with TIES, DARE and AR-MAP under the same settings? What is the methodological difference between A2D-Transfer and AR-MAP?

  • How do "+τA+\tau_A only" and "−τD−\tau_D only" perform on more benchmarks and in the controlled setting? Does A2D-Merge still give clear gains when the AR donor is not stronger than the diffusion parent?

5 comments

We thank the reviewer for their valuable feedback and recognition of our method's simplicity and training-free nature, the breadth of our evaluation, and our informative analyses. Below, we address their comments and questions.

W1/Q1: Model merging baselines.

As a framework for enhancing diffusion models with AR post-training weights, A2D can incorporate different merging techniques, including TIES and DARE. In our manuscript, we chose task arithmetic (TA) for its simplicity and generality. This allows us to focus on their transferability and complementarity without introducing additional merging heuristics or hyperparameters.

Following the reviewer’s suggestion, we conducted additional experiments integrating these methods into A2D, in the table below. We independently tuned the merging coefficients for each method on the anchor task IFEval and report their best-performing configurations below.

Modelτ_D / τ_A weightsIFEvalMATH-500MBPP+Average
Transfer
Dream-7B Base0 / 049.720.859.343.3
TIES0 / 0.955.536.461.951.3
DARE0 / 0.860.335.460.652.1
A2D (transfer)0 / 0.959.240.262.754.0
Merge
Dream-7B Instruct1 / 063.038.457.152.8
TIES0.8 / 0.263.643.857.955.1
DARE0.1 / 0.947.335.861.648.2
A2D (merge)0.4 / 0.672.845.058.758.8

A2D with standard TA (our default setting) achieves the highest average performance in both Transfer and Merge, outperforming TIES and DARE by at least 1.9 in Transfer and 3.7 percentage points in Merge, respectively. The limited benefits of interference-mitigation techniques, such as TIES and DARE, are consistent with our finding of limited interference between AR and diffusion updates, supporting our choice of TA as a simple and effective merging strategy. We will include these comparisons in the revised manuscript.

W1/Q1: Comparisons with AR-MAP.

We fully acknowledge AR-MAP’s contribution to transferring AR updates in the DPO setting and explicitly discuss it in the manuscript. A2D-Transfer and AR-MAP both build on Task Arithmetic and share the same task-vector addition operation. Our main contribution lies in systematically establishing and studying the reuse of AR post-training updates to enhance diffusion models, rather than introducing a new merging method.

However, the two works address different research questions and differ in their scope. AR-MAP addresses specifically likelihood-estimation difficulties in diffusion DPO by training an AR DPO donor and transferring its update as an alternative to diffusion-specific preference optimization. A2D instead investigates how the existing AR post-training ecosystem can enhance diffusion models, demonstrating its effectiveness across post-training stages including instruction tuning, reasoning, RL, and other task-specialized domains. Beyond transfer, A2D studies the relationships between independently learned AR and diffusion updates for the same capability. Our controlled experiments reveal nearly orthogonal parameter updates with more aligned representation changes. These findings motivate A2D-Merge, which combines both updates to further improve diffusion models that have already completed the corresponding post-training.

W2/Q2: A2D-Merge with a weaker AR donor

We thank the reviewer for this insightful suggestion. While leveraging stronger AR donors is itself a practical advantage of A2D, we agree that it is important to investigate whether the gains also arise from complementary AR and diffusion updates. To this end, we conducted an additional experiment by training an AR donor instruction-tuned on 200K general-purpose examples from Tulu3 and SmolTalk (the same data sources used by Dream’s instruction-tuning), rather than the math-specific NuminaMath dataset. This produces an AR donor that performs worse on math benchmarks than the diffusion-post-trained model (57.3% vs. 60.5% average accuracy), while potentially providing more complementary knowledge from general instruction tuning. We then provide A2D-merge with this weaker AR donor, and present the results below:

ModelGSM8KMATH-500Average
Tulu3+SmolTalk AR donor (weaker)77.537.057.3
Dream NuminaMath-SFT (stronger)74.047.060.5
A2D-Merge86.050.068.0

Despite the weaker AR donor, A2D-Merge improves the diffusion model on both GSM8K and MATH-500, raising the average accuracy from 60.5% to 68.0%. These results suggest that A2D-Merge's gains come not only from stronger AR donors, but also from the complementarity between AR and diffusion post-training updates.

AuthorsOpenReview ↗

Rebuttal - part 2

W2/Q2: Contributions of individual updates.

Following the reviewer's suggestion, we extended the +τ_A only and −τ_D only ablations to additional benchmarks, including non-anchor evaluations. The results are summarized below. Results are averaged over 7 anchor and 21 non-anchor benchmarks with complete ablation results. Wins indicate strict improvements over the diffusion parent.

MethodAnchor avg.Non-anchor avg.Overall avg.Wins vs. parent
Diffusion parent72.352.257.3—
+τ_A only75.855.160.319/28 (67.9%)
−τ_D only74.152.958.220/28 (71.4%)
A2D-Merge76.855.360.727/28 (96.4%)

As shown in the table above, both +τ_A only and A2D-Merge achieve notably higher average performance than −τ_D only, highlighting the importance of incorporating AR updates rather than merely downscaling diffusion updates. While A2D-Merge achieves the highest average performance, its more notable advantage is the consistency of its gains: it improves upon the diffusion parent on 27/28 benchmarks (tying on the remaining one), compared with 19/28 and 20/28 for the two ablations. A2D-Merge also outperforms +τ_A only on 17/28 benchmarks and −τ_D only on 24/28 benchmarks. Similar trends are observed in our controlled experiments, as also illustrated in Fig. 4(a), where A2D-Merge achieves 64.3% average accuracy, outperforming both +τ_A only (60.9%) and −τ_D only (63.4%).

We would like to also clarify that the +τ_A only baseline also falls within our framework of AR-to-diffusion post-training merging, but adopts a different merging strategy that fixes the diffusion update coefficient at 1. This comparison allows us to investigate whether jointly rebalancing AR and diffusion updates brings additional benefits, rather than to validate their complementarity.

We hope this effectively addresses the reviewer’s concerns and remain available for any further discussion or clarification.

AuthorsOpenReview ↗

Gentle Follow-up

Dear Reviewer,

We wanted to follow up on whether our responses have addressed your concerns. We would be happy to address any remaining questions before the discussion period ends.

Thank you again for your valuable feedback and for your time.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Manuscript Revisions

Dear Area Chair and Reviewers,

We thank the reviewers for their valuable feedback and suggestions. We have updated the manuscript PDF to incorporate the additional experiments and clarifications from our rebuttal, with all changes highlighted in red. Below, we summarize the main revisions.

Comparison with TIES and DARE (owgH, xNJ5). We added comparisons with TIES and DARE, showing that simple task arithmetic achieves the highest average performance in both Transfer and Merge (Sec. 4.2, Appendix C.6).

Relationship with AR-MAP (owgH, xNJ5). We expanded our discussion of AR-MAP to clarify the relationship between the two approaches and highlight their distinct research questions and contributions. In particular, AR-MAP focuses on diffusion DPO, while A2D investigates the broader reuse of existing AR post-training updates and their complementarity with diffusion updates through A2D-Merge (Sec. 5).

Weaker AR donor (owgH). We added experiments demonstrating that A2D-Merge remains effective even when the AR donor is weaker than the diffusion parent, improving average accuracy from 60.5% to 68.0% (Sec. 4.2, Appendix C.7).

Extended ablations (owgH, xNJ5). We extended the comparisons with adding the AR update only and downscaling the diffusion update only to 28 benchmarks. A2D-Merge improves 27/28 benchmarks, compared with 19/28 and 20/28 for the two ablations (Sec. 4.2, Appendix C.1).

Generalization across architectures (xaMH). We added experiments investigating transfer between independently pretrained models, including cross-architecture transfer using THESEUS (Sec. 4.2, Appendix C.8).

Parallel decoding and bidirectional reasoning (xaMH). We added experiments on Dream's parallel decoding and Sudoku task, showing that A2D preserves diffusion-specific capabilities (Sec. 4.2, Appendix C.9).

We thank the reviewers again for their constructive feedback, which has helped strengthen our manuscript. We remain available to address any further questions or clarifications the reviewers may need.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.