acceptodds
Under review as a conference paper at ICLR 2027

TreePC: Posterior-Consistent and Dependency Aware Decoding for Few-Step Diffusion Language Models

Abstract

Diffusion language models (dLLMs) enable parallel generation by unmasking multiple tokens at each denoising step. Existing dLLM efficient decoding methods usually achieve speedups by substantially reducing decoding steps. However, aggressive step reduction forces dLLM to commit far more larger token sets at each step, thereby degrading generation quality. We identify two primary causes of quality degradation in few-step dLLM efficient decoding: (1) ***marginal drift***, where a few-step student’s per-position posteriors deviate from those of a fine-step teacher; and (2) ***dependency error***, where independent sampling ignores compatibility among tokens committed within the same step. To fill the identified research gap, we introduce **TreePC**, a posterior-consistent and dependency-aware decoding framework that first aligns student marginals through Posterior Consistency LoRA (PC-LoRA), then predicts pairwise dependencies from teacher counterfactuals, constructs a counterfactual-dependency maximum spanning tree, and performs parent-conditioned corrections along the tree. A learnable global scale jointly controls the correction budget with edge-specific gates, allowing TreePC to exploit informative dependencies while preserving calibrated posteriors when conditional evidence is weak. During inference, TreePC invokes dLLM backbone model exactly once per diffusion step; its dependency estimator, tree construction, and correction module are lightweight and leave backbone neural function evaluations unchanged. Experimental results on GSM8K, MATH-500, HumanEval, and MBPP demonstrate that TreePC improves task accuracy and reduces marginal and conditional KL divergence while bringing very modest decoding latency. These results establish structured posterior correction as an effective quality-efficiency trade-off for few-step dLLM generation. Code is available at: [TreePC](https://anonymous.4open.science/r/TreePC_efficient_dLLM_decoding-C22F/).

Primary area:
foundation or frontier models, including LLMs
Keywords:
Diffusion Language Models, Efficient Few-Step Decoding, Posterior Consistency, Dependency-Aware Decoding, Parameter-Efficient Fine-Tuning, Conditional Posterior Correction, Maximum Spanning Tree

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

The paper decomposes few-step dLLM decoding loss into marginal drift (per-position posterior gap to a fine-step teacher) and dependency error (incompatibility among tokens committed in the same step under conditional independence). PC-LoRA freezes the backbone and aligns marginals with mixed-budget on-policy teacher KL. Counterfactual parent reveals then supply pairwise dependency targets; a lightweight head predicts them from one student forward pass, the graph is compressed into a maximum spanning tree rooted at the highest-confidence candidate, and a correction head applies gated parent-conditioned logit residuals, with exactly one backbone forward per step. The appendix provides an exact KL decomposition and the Chow-Liu projection. At 64 NFE, TreePC beats vanilla, Fast-dLLM and CD4LM on GSM8K, MATH-500, HumanEval and MBPP.

Critical strengths

The decomposition maps one-to-one onto the method: DKL(P∗∥Pind)=TCT+∑iDKL(qiT∥piS)D_{\mathrm{KL}}(P^*\|P_{\mathrm{ind}})=\mathrm{TC}^T+\sum_i D_{\mathrm{KL}}(q^T_i\|p^S_i) splits marginal calibration from residual dependence, and the Chow-Liu connection justifies a mutual-information-weighted MST rather than an ad hoc graph. Counterfactual parent reveals are a computable supervision signal. Mechanism-level metrics are thorough (marginal KL, top-1 agreement, Spearman, edge overlap, conditional KL reduction) with local-tree and random-tree controls.

Critical weaknesses
  1. Every efficiency number is an estimate, and latency is never compared against competing methods. All three RQ4 subfigures are labeled "Estimated"; the text states that per-step cost was measured only on GSM8K and HumanEval and multiplied by NFE, with task-family proxies for MATH-500 and MBPP, so four benchmarks of bars rest on two measurements. The extrapolation ignores EOS early stopping, the dependence of auxiliary cost on mtm_t, and whether the vanilla side uses caching. The per-step latency, throughput and peak memory promised in Section 4.2 appear nowhere. The latency comparison is against vanilla-256/512 only; Fast-dLLM and CD4LM never enter a latency figure, although the former carries a block KV cache so equal NFE does not mean equal wall-clock, and the dLLM-Cache line appears only in related work.
  2. The end-to-end evidence is a single operating point, with no ablation. Of nine nominal budgets only B=64B=64 has a numeric table; B=4B=4, 8 and 16 appear only as axis ticks, although the conclusion itself notes that tree construction and correction add measurable latency at very small budgets. The word "ablation" never appears: the tree against a full graph, the root rule, the gate and global scale, the ranking loss, S=3S=3 counterfactual samples and top-K=16K=16 are none of them isolated, even though the paper calls the ranking term especially important. The generation length LL is never stated, so NFE cannot be converted into the per-step commit count mtm_t, on which both tree coverage and the O(mt2)O(m_t^2) auxiliary cost depend. There are also no seeds or confidence intervals, while several Table 2 gaps are 1-3 points.
  3. Most of the gain comes from standard distillation. From RQ1 and Table 2, Dream GSM8K at 64 NFE gives vanilla 49.7, PC-only about 67.4, TreePC 72.5, so the titular tree correction adds about 5.1 points, and 1.7 on LLaDA. Neither abstract nor conclusion conveys this.
  4. Training availability is unmatched, its cost is unreported, and the closest competitors are absent. TreePC collects trajectories on each benchmark's training distribution and trains three parameter groups, while Fast-dLLM is training-free and it is unstated whether CD4LM was retrained under matched data and compute. Counterfactual labeling runs one teacher forward per sampled parent for every candidate pair, and the four-stage pipeline regenerates on-policy states, yet no figure is given for that total cost. CoDD and Adaptive Parallel Decoding, which the paper lists as the same family, are never compared, and two controls are missing: PC-LoRA plus independent decoding plus a confidence threshold, and a plain LoRA SFT at the same data budget.
  5. The fixed-budget setting is never justified. Per-position confidence is free in text decoding, so discarding adaptivity needs an argument, and two obvious middle grounds go undiscussed: a confidence threshold with a step cap, and attaching the dependency correction to an adaptive scheduler, which the paper's own description allows. This also affects how Table 2 reads: forced to 64 NFE, Fast-dLLM falls below plain vanilla few-step on LLaDA MATH-500 and HumanEval (21.7 vs 26.3, 28.2 vs 34.0), which indicates it is being run outside its design point; the table supports "a confidence threshold fails at 64 NFE" rather than "TreePC beats Fast-dLLM". Separately, a tree keeps m−1m-1 of the m(m−1)/2m(m-1)/2 pairs, half at m=4m=4 but about 3% at m=64m=64, so coverage is lowest exactly where the budget is tightest, and Figure 3(b)'s 75-92% does not say at which mm it was measured.
Questions for authors
  1. Can you provide measured end-to-end wall-clock on all four benchmarks (with the actual step counts under EOS early stopping), broken down into backbone, Dependency Head, MST and Correction Head, and place Fast-dLLM and CD4LM in the same latency figure?
  2. What is the generation length LL? Please report quality, realized NFE, latency and the auxiliary share at B=4B=4, 8 and 16.
  3. Please add component ablations (tree versus full graph, root rule, gate and global scale, ranking loss, SS, top-KK), and place vanilla, PC-only and full TreePC in one table so the net contribution of the tree correction is directly readable.
  4. What are Fast-dLLM's full quality-versus-realized-NFE and wall-clock curves at its own thresholds (0.9, 0.8, 0.7)? What does a fixed budget buy over a threshold with a step cap, and what happens if the dependency correction is attached to an adaptive scheduler instead?
  5. What is the total cost of counterfactual labeling and the four-stage pipeline, in teacher forward passes and GPU hours? With "PC-LoRA + independent decoding + confidence threshold" and a same-data LoRA SFT as controls, how much lead remains, and how does TreePC compare with CoDD and Adaptive Parallel Decoding?
  6. What are the distributions of mtm_t and tree depth, and the value and trigger rate of the fallback threshold τ\tau? How does the fraction of pairs covered by the tree vary with mtm_t, and at which mtm_t was Figure 3(b) measured?
Reviewer_MDqbOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TreePC, a decoding framework for few-step diffusion language models. It first calibrates per-position marginal distributions using Posterior Consistency LoRA, then trains a dependency head via teacher counterfactuals to build a maximum spanning tree over simultaneously committed tokens, and finally applies parent-conditioned logit corrections from root to leaves. Experiments on LLaDA-8B and Dream-7B across math and code benchmarks show accuracy improvements at matched backbone evaluations.

Critical strengths
  • The paper cleanly decomposes few-step parallel decoding degradation into marginal drift and dependency error, grounding its tree approximation in Chow-Liu mutual information theory.
  • Using counterfactual parent reveals to supervise pairwise token dependencies allows the model to capture same-step interactions without extra backbone forward passes during inference.
  • Evaluations across two diffusion backbones and four math and coding benchmarks show consistent accuracy gains over vanilla few-step decoding and adaptive baselines.
Critical weaknesses
  • The four-stage training pipeline is overly complex for the marginal gain of the tree module, as text on page 7 and Table 2 reveal that PC-LoRA alone improves Dream-7B HumanEval from 36.0% to 55.0% while adding the entire tree correction machinery only reaches 55.2%.
  • Table 2 omits the PC-only baseline, obscuring the fact that most of the advantage over external baselines comes from LoRA distillation rather than dependency-aware tree decoding.
  • The latency results in Figure 5 rely on analytical proxies rather than direct wall-clock measurements for MATH-500 and MBPP, and do not compare real wall-clock speed against Fast-dLLM with its KV cache enabled.
  • Experiments omit direct comparisons with existing dependency-aware parallel decoding baselines such as CoDD and APD discussed in Appendix B.2.
  • Constructing counterfactual supervision requires multiple teacher evaluations per candidate position, and evaluation is restricted to short math and code tasks rather than longer sequences.
Questions for authors
  1. On Dream-7B at 64 NFE, PC-LoRA alone already achieves 67.4% on GSM8K and 55.0% on HumanEval, whereas full TreePC reaches 72.5% and 55.2%. Can you include the PC-only row in Table 2 across all benchmarks to clarify how much gain comes strictly from tree correction?
  2. Since Figure 5 uses proxy latency estimates, what is the actual measured wall-clock speedup of TreePC compared to Fast-dLLM with KV caching enabled and dependency-aware baselines like CoDD?
  3. How do the offline counterfactual labeling cost and online tree traversal latency scale when generating longer sequences where the number of tokens committed per step grows large?
Reviewer_vCRWOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper targets on the training of few-step DLMs. Importantly, the author mainly argues two limitation of existing few-step DLMs, including marginal drift and dependency error. To this end, the authors propose Posterior Consistency LoRA (PC-LoRA) aligns the few-step student’s per-position distributions with those of a fine-step teacher and a lightweight Dependency Head learns pairwise influences from teacher-generated counterfactuals that reveal a candidate parent token while leaving the child position masked. Experimental results show the effectiveness of this method.

Critical strengths
  • The topic is interesting and training a better few-step DLM is necessary.

  • The authors provide many detailed explaination about their method.

  • The experimental results look good.

Critical weaknesses
  • This paper is very difficult to follow and the presentation needs to be improved. For example, 1) the introducation section involves many mathmatical representations. I suggest that the authors can move many mathmatical representations to the method section and remain the clear and easily understanding. 2) In the second paragraph of the Introducation section, What is the per-position posterior od the student model? What is the a fine-step teacher? What is the marginal posteriors? I DO NOT find any definations about these contents before.

  • The content of method section is very dense. It seems that the posterior consistency LORA is not novel. For dLLM, the distribution alignment between the few-step predictions and fine-step ones is very common such as dInfer. The contribution of introducing the LORA is trival.

  • Figure 1 is very difficult to understand. I suggest the author can explain it more detailed.

  • I think the overall pipeline is very sophisticated, including many stages and components. I dislike such a work which hinders the reproduction and it may be a great job to scale such a method to a larger model.

Questions for authors

I think the author can address my comments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.