acceptodds
Under review as a conference paper at ICLR 2027

Conflux: Self-Refining Unified Diffusion LLM with Cross-Paradigm On-Policy Distillation

Abstract

We propose Conflux, a unified discrete diffusion large language model (DLLM) for visual understanding and generation tasks. Compared with prior unified diffusion models, Conflux introduces two major innovations. First, we propose a unified cross-paradigm on-policy distillation (OPD) framework that successfully transfer the visual understanding ability of autoregressive VLMs and visual generation ability of continuous flow models, to a unified DLLM. Second, Conflux incorporates a novel progressive self-refining diffusion process that gradually corrupts each token across multiple steps during training, as opposed to adopting the conventional masked diffusion formulation that directly corrupts a clean image token to an uninformative mask token. At inference, this design allows Conflux to make soft commit to a token, enable progressive self-refinement for better visual quality. Through extensive experiments, we show that Conflux achieves state-of-the-art performance among unified DLLMs on diverse tasks including visual question-answering, multi-modal reasoning, object grounding, text-to-image generation, and image editing.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Unified Multimodal Models, Diffusion Language Models, On Policy Distillation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_BCxdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Conflux unifies visual understanding and generation in a discrete diffusion model. It progressively refines tokens using codebook geometry and learns from both autoregressive and continuous image teachers. AR teachers supervise its autoregressive view, with consistency training transferring this knowledge to the diffusion view; the image teacher helps improve its editing outputs. Results on 22 benchmarks, together with ablation studies, show gains in reasoning, image editing, and visual quality.

Critical strengths
  • Well-motivated distillation design. The paper identifies a concrete supervision mismatch and supports its proposed AR-to-diffusion transfer with multiple objective ablations and consistent reasoning gains.
  • Effective progressive discrete refinement. The geometry-aware corruption process improves visual quality over retrained diffusion baselines, demonstrating a practical approach to stronger unified discrete models.
Critical weaknesses
  • Incomplete attribution of gains. Large-scale training contributes substantially to the overall results. Image-editing comparisons also conflate teacher refinement with on-policy updating, leaving the latter’s independent contribution unclear.
  • Limited evidence for the magnitude and robustness of refinement gains. The fixed-model gain at 64 forward passes is modest, while additional refinement can hurt at low budgets. Missing uncertainty estimates limit confidence in small visual-quality advantages.
Questions for authors
  1. Were all Figure 5 variants matched in initialization, data, training compute, and inference settings? How should their larger gains be reconciled with the modest fixed-model gains in Table 6?
  2. Does the proposed editing method outperform teacher-refined targets generated from a fixed earlier student checkpoint under matched budgets? This would isolate the benefit of on-policy updating.
  3. Can you provide uncertainty estimates for HPSv3 and ImgEdit, and clarify whether the reported compute includes teacher rollouts and ablations?
1 comment

Q1 Were all Figure 5 variants matched in initialization, data, training compute, and inference settings? How should their larger gains be reconciled with the modest fixed-model gains in Table 6?

Yes, they have matched initialization, data, and training compute. In terms of the gap, there are two interpretations. First, Figure 5 shows ablation runs with a short schedule which is 20k steps, while the full stage 1 training takes 400k steps and is considerably more expensive, hence the setup is different. These results show that self-refinement is more data efficient and converges faster under apples-to-apples comparison. Second, we find that even if refinement is not enabled at inference, it still outperforms vanilla mask diffusion, suggesting self-refining training is helpful even if at inference we set K=1 and never update tokens after initial commits. This is because the training objective and the distribution of training-time corruption is different. Self-refinement task is “easier” than one-step mask-to-token prediction, offering smoother distribution and training dynamics.

Q2 Does the proposed editing method outperform teacher-refined targets generated from a fixed earlier student checkpoint under matched budgets?

Yes. In revision Table 4, we added an off-policy baseline which uses refined targets on frozen students, which result in worse performance.

MethodImgEdit
Conflux-Base3.81
Off Policy, Independent Teacher Generation3.90
Off Policy, Refined Frozen Student Generation3.99
On Policy, Refined Student Generation4.17

Q3.1 Can you provide uncertainty estimates for HPSv3 and ImgEdit.

For HPSv3, we report 95% confidence interval of different refinement settings in revision Table 7, copied below

Actual NFEN=1N=4N=8N=16
47.3700 ± 0.0396-2.4895 ± 0.0478——
89.6214 ± 0.03219.0038 ± 0.03490.0834 ± 0.0520—
1610.5449 ± 0.027610.8205 ± 0.026910.3044 ± 0.02853.0348 ± 0.0543
3210.8921 ± 0.025511.0137 ± 0.025510.9894 ± 0.025410.7620 ± 0.0261
4810.9378 ± 0.025511.0369 ± 0.025111.1167 ± 0.025011.0861 ± 0.0248
6411.0051 ± 0.025511.0287 ± 0.025211.1050 ± 0.025011.1010 ± 0.0248

For ImgEdit , we report in revision table 6

Uncertainty on ImgEdit. We report 95% confidence intervals computed using 3 independent seeds.

ModelOverall
Conflux-Base3.808 ± 0.016
+ Off-Policy KD (Independent Target)3.902 ± 0.021
+ Off-Policy KD (Refined Target)4.003 ± 0.022
+ Cross-Paradigm OPD (\ours)4.171 ± 0.018

Q3.2 Whether the reported compute includes teacher rollouts and ablations?

Yes

W1 Incomplete attribution of gains.

Given a final checkpoint performance, the exact factorization of improvements of OPD, self-refining diffusion, and increased training data for each task is difficult to separate because they are introduced at different training stages discussed in Section 4.1. In particular, the self-refining diffusion vs vanilla mask diffusion is studied in the pretraining stage 1 whose data is mere image-text pairs. Image editing data is introduced in stage 2 SFT, and OPD is performed in stage 3. Running multiple setups across the entire pipeline is computationally expensive, so we organize our ablation studies into grouops of component matched settings at respective training stages.

Specifically, self-refining diffusion is ablated in the pretraining stage on the text-to-image task (Figure 5), where we match the initialization, data, and training schedule. Image editing is added to stage 2, which already underwent considerable training with self-refining diffusion in stage 1. OPD is further introduced in stage 3, where we compare our proposed OPD with off-policy alternatives (Table 4) under matched compute. These experiments are initialized from the same stage 2 checkpoint. At this stage, it becomes impossible to have a fair, non-self refining baseline unless we conduct another full scale training of stage 1, 2.

While the performance gain cannot be quantitatively factorized, we emphasize that each design is accompanied by an ablation study at respective stages with matched compute and data, which establishes their effectiveness independent of other components and data scaling. In summary, self-refining ablations are conducted in stage 1 with a 20k schedule on matched initialization, schedule, and data. OPD ablations are conducted in stage 3 initialized from the same checkpoint of stage 2 training (Conflux-Base) which has self-refinement enabled, so the 3.81→4.173.81\rightarrow4.17 ImgEdit gain in table 4 completely attributes to cross-paradigm OPD.

W2 Limited evidence for the magnitude and robustness of refinement gains

See our response in Q1,Q2, where we show the improvements are statiscally significant.

Reviewer_5U3SOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Conflux, a unified discrete diffusion language model (DLLM) for multimodal understanding and generation. Its contributions are:

  1. A cross-paradigm on-policy distillation (OPD) framework that distills from an autoregressive (AR) VLM teacher for understanding and a continuous-flow image-editing teacher for generation.

  2. A self-refining masked diffusion process that extends binary mask/clean states with intermediate token-age states. The process uses ball-tree codebook geometry for efficient implementation.

  3. A two-tower Elastic-MoT architecture trained in three stages.

The paper evaluates Conflux on visual QA, grounding, multimodal reasoning, text-to-image generation, and image editing. It reports state-of-the-art results among unified DLLMs and reduced gaps to specialized models, and includes ablations of OPD design and self-refinement. Some claims appear overstated or inconsistent with the tables. Reproducibility is limited by non-public teachers and data, as well as incomplete implementation details.

Critical strengths
  1. Cross-paradigm OPD with a transitive distillation objective. The authors address the scarcity of strong MDM teachers by first supervising the student’s AR view with an AR teacher, then using the student’s AR view to supervise its diffusion view. This design avoids the context mismatch that can arise when AR teacher scores are applied directly to masked contexts. The ablation reports that the naive OPD-LM-style baseline degrades performance.

  2. Self-refining masked diffusion with a codebook-geometry implementation. Intermediate token-age states support progressive coarse-to-fine refinement instead of one-step mask-to-clean transitions. The authors implement the process with ball-tree codebook geometry to reduce memory use. Ablations show consistent gains, particularly on HPSv3.

Critical weaknesses

None

Questions for authors
  1. In continuous-flow OPD, the objective is SFT on teacher-refined student rollouts rather than a KL-style on-policy loss. What distinguishes this method from teacher-refined self-training, and why is “on-policy distillation” the appropriate label? If feasible, can the authors compare it with a controlled, KL-based OPD variant?

  2. The self-refining process uses ball-tree depth D, token age K, and interval length ϵ. What values are used in the main experiments? How sensitive are the results to K, D, and ϵ beyond the NFE/refinement-step ablation in Table 6?

  3. Because T2I does not use OPD, how much of the visual-generation improvement is attributable to OPD, self-refining diffusion, and increased training data for each task? Please provide a factorial ablation that separates these factors for T2I and editing.

3 comments

Q1 Why is “on-policy distillation” the appropriate label? Can you provide KL-style variant?

The key distinction for “on-policy distillation” is that we use on-policy samples to refine as opposed to pre-generated student samples. In revision, we added an off-policy baseline which uses refined targets on frozen students, which result in worse performance.

MethodImgEdit
Conflux-Base3.81
Off Policy, Independent Teacher Generation3.90
Off Policy, Refined Frozen Student Generation3.99
On Policy, Refined Student Generation4.17

The KL-based OPD variant is not feasible because KL computation is not tractable in this setting. Specifically, the teacher operates in continuous latent space while the student operates on discrete codebooks. Student predicts each individual tokens over a 2D grid,each of which can be independently corrupted, while the teacher distribution only defines a single global noise level across the image and cannot independently corrupt individual patches.

Q2 Configuration and senstitive of ball-tree depth D, token age K, and interval length ϵ

These details are in Appendix A.2.1 L1133. Since Ball tree is a binary tree, the depth D is completely determined by the tokenizer’s vocabulary size. The tokenizer of Conflux contains 131k discrete codes, which corresponds to a ball tree with D=18. Hence, the vanilla setup is K=17. To enable flexible inference, we in practice randomize K for each training step among the candidates {1,2,..17}, where for K<17, we skip intermediate ages uniformly. For example, K=4 corresponds to age progression 0,4,8,12,17 in the original K=17 notation. \epsilon is randomly sampled from the interval [0,1]uniformly (L1210).

In terms of sensitivity of K, D, and ϵ, they are already fully covered in Table 6 (now table 7 in revision). First, D is fixed since it depends on the tokenizer's vocabulary size. K, D only materially affect the sampling process in the form of changing the number of refinement steps N in our sampler implementation.

Specifically, For a fixed NN, the sampler selects the corresponding effective levels from this underlying tree. For example, when N=4, nominal settings K=4,8 or 17 are mapped to the same levels 0,4,8,12,17 in the original K=17 notation, traversed in reverse during generation. Intermediate levels that are not visited do not introduce additional sampling operations. Likewise, ϵ\epsilon is discretized into the effective update schedule, where values within the same discretization interval produce identical schedules.

This equivalence follows directly from the sampling rule: at each scheduled update, the model receives the same current state, predicts a clean token, and applies the same corruption kernel associated with the selected tree level. Neither the nominal K nor the undiscretized ϵ∈[0,1]\epsilon\in[0,1] enters this operation independently. Consequently, identical effective schedules induce identical transition kernels and therefore the same output distribution. Given a fixed seed, they produce identical trajectories. The concrete correspondence is shown below

NFENHPSv3Kε
417.37001–17(0, 1/4]
44−2.48954–17(3/4, 1]
819.62141–17(0, 1/8]
849.00384–17(3/8, 1/2]
880.08348–17(7/8, 1]
16110.54491–17(0, 1/16]
16410.82054–17(3/16, 1/4]
16810.30448–17(7/16, 1/2]
16163.034816–17(15/16, 1]
32110.89211–17(0, 1/32]
32411.01374–17(3/32, 1/8]
32810.98948–17(7/32, 1/4]
321610.762016–17(15/32, 1/2]
48110.93781–17(0, 1/48]
48411.03694–17(1/16, 1/12]
48811.11678–17(7/48, 1/6]
481611.086116–17(5/16, 1/3]
64111.00511–17(0, 1/64]
64411.02874–17(3/64, 1/16]
64811.10508–17(7/64, 1/8]
641611.101016–17(15/64, 1/4]

In summary, sensitivity to D is not well-defined since D is not a variable. Sensitivity to K, ϵ is already covered in refinement step study.

Q3 Performance Gain BreakDown

Given a final checkpoint performance, the exact factorization of improvements of OPD, self-refining diffusion, and increased training data for each task is difficult to separate because they are introduced at different training stages discussed in Section 4.1. In particular, the self-refining diffusion vs vanilla mask diffusion is studied in the pretraining stage 1 whose data is mere image-text pairs. Image editing data is introduced in stage 2 SFT, and OPD is performed in stage 3. Running multiple setups across the entire pipeline is computationally expensive, so we organize our ablation studies into grouops of component matched settings at respective training stages.

Specifically, self-refining diffusion is ablated in the pretraining stage on the text-to-image task (Figure 5), where we match the initialization, data, and training schedule. Image editing is added to stage 2, which already underwent considerable training with self-refining diffusion in stage 1. OPD is further introduced in stage 3, where we compare our proposed OPD with off-policy alternatives (Table 4) under matched compute. These experiments are initialized from the same stage 2 checkpoint. At this stage, it becomes impossible to have a fair, non-self refining baseline unless we conduct another full scale training of stage 1, 2.

While the performance gain cannot be quantitatively factorized, we emphasize that each design is accompanied by an ablation study at respective stages with matched compute and data, which establishes their effectiveness independent of other components and data scaling. In summary, self-refining ablations are conducted in stage 1 with a 20k schedule on matched initialization, schedule, and data. OPD ablations are conducted in stage 3 initialized from the same checkpoint of stage 2 training (Conflux-Base) which has self-refinement enabled, so the 3.81→4.173.81\rightarrow4.17 ImgEdit gain in table 4 completely attributes to cross-paradigm OPD.

Reviewer_5U3SOpenReview ↗

Official Comment

I appreciate the author’s detailed rebuttal. They conducted numerous experiments that addressed my concerns, so I have decided to maintain my score.

Reviewer_Pp7COpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Conflux unifies multimodal understanding and image generation through block diffusion. A hierarchy over VQ codes and token-age states enables gradual image refinement. Understanding transfers from an autoregressive teacher through the student's AR view to its diffusion view; editing uses teacher-refined student outputs as supervision, rather than conventional KL-style on-policy distillation. Benchmarks, checkpoint comparisons and ablations evaluate understanding, reasoning, generation and editing.

Critical strengths
  1. The transfer design gives a unified diffusion student access to stronger teachers using different generation paradigms. Reasoning improves over the base checkpoint across all seven benchmarks, while overall editing performance improves over both the base and independent-teacher-target control. This supports adapting editing targets to student outputs rather than using independent teacher generations. Separate single-task ablations support the understanding losses and rollout choices in a setting distinct from joint training.

  2. The hierarchy enables tractable refinement of intermediate token commitments through forward marginals and reverse transitions on the stated support. Its practical value is supported by separately trained controls: self-refinement improves HPSv3 over uniform and masked diffusion, while its GenEval performance is comparable to uniform diffusion and higher than masked alternatives overall. The benefit is budget-dependent: multiple refinements help at larger NFE but can cause severe HPSv3 deterioration at low NFE.

Critical weaknesses
  1. Incomplete implementation reporting. Reproducing the corruption process requires knowing how the formal code-and-age state is used in the neural implementation and how ε is set during training. The state itself is defined, but its actual neural representation or use is unexplained. The training value or sampling distribution of ε, each token's transition duration, is unspecified; randomizing K does not determine it. The inference sweep does not resolve these training choices.

  2. Uncertainty in cross-model generation comparisons. The small HPSv3 margin over Qwen-Image is difficult to interpret without reported per-model sampling settings, result provenance and uncertainty information. It is unclear which entries were rerun or drawn from reports; the refinement study's protocol does not establish these settings. This leaves the specific cross-family ranking less secure than the controlled refinement and editing benefits supporting the broader contribution.

Questions for authors
  1. Implementation clarification: How is the formal age state represented or used in the neural implementation? What value or sampling distribution of the transition duration ε was used during training?

  2. Evaluation clarification: Which entries in the cross-model generation comparison were rerun? Please give each model's steps, guidance, resolution, samples per prompt and benchmark implementation, plus any existing uncertainty evidence for the Qwen-Image HPSv3 comparison.

2 comments

W1 Q1 Incomplete implementation reporting.

Thank you for your suggestion, and we now include more details in Appendix A.2.1 L1133. On the network level, the new state space is implemented by two separate embedding layers. Token states are represented as the sum of token-age embedding and standard token class embedding. During training, is randomly sampled from the interval [0,1] at uniform (L1210). As a special case, when ϵ→0\epsilon\rightarrow0 is very small, the training input is dominated by clean tokens and fully masked tokens, and the process reduces to vanilla MDM.

W2 Q2 Evaluation clarification

For HPSv3, we evaluated all models ourselves, with 3 samples per prompt. All models use the official default setting, which we assume would be already optimized. For Qwen-Image specifically, we use true_cfg_scale=4.0, num_inference_steps=50 with a resolution 1328x1328. In terms of uncertainty, we report 95% confidence interval of different refinement settings in revision Table 7, which is also copied below.

Actual NFEN=1N=4N=8N=16
47.3700 ± 0.0396-2.4895 ± 0.0478——
89.6214 ± 0.03219.0038 ± 0.03490.0834 ± 0.0520—
1610.5449 ± 0.027610.8205 ± 0.026910.3044 ± 0.02853.0348 ± 0.0543
3210.8921 ± 0.025511.0137 ± 0.025510.9894 ± 0.025410.7620 ± 0.0261
4810.9378 ± 0.025511.0369 ± 0.025111.1167 ± 0.025011.0861 ± 0.0248
6411.0051 ± 0.025511.0287 ± 0.025211.1050 ± 0.025011.1010 ± 0.0248

These results are sampled at 1024px resolution. The entry NFE=48, N=8 is selected for main tables. The direct comparison with Qwen is

ModelHPSv3
Qwen-Image10.9052±0.0159
Conflux11.1167±0.0250

Thank you for the response and clarifications. They address most of my concerns, and I will maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.