acceptodds
Under review as a conference paper at ICLR 2027

Cord: Confidence-Anchored Order Residual Decoding for Masked Diffusion Language Models

Abstract

The decoding order of masked diffusion language models (dLLMs) can substantially affect generation quality, yet common decoders typically apply the same reveal rule throughout generation. Confidence provides a strong default, but it measures how decisively a position can be predicted now, not how useful revealing that position will be for subsequent decoding. We propose CORD, a lightweight prompt-conditioned policy that retains confidence decoding as a stable reference and learns when and how to depart from it through geometric residuals over generation direction, proximity to revealed context, and two-sided infilling. The residuals adapt to both the prompt and the evolving partial response. Rather than sampling a new policy action at every decoding step, CORD samples a single low-dimensional coefficient matrix per rollout, which induces an entire state-dependent reveal trajectory. This yields a contextual-bandit formulation trained directly from sequence-level rewards, without step-level credit assignment or backbone updates. Across three dLLMs and five reasoning, code, and constraint-satisfaction benchmarks, CORD improves average task success from under confidence decoding to , and further improves over the best per-domain fixed residual. Behavioral audits show that the learned residuals induce substantial, task-dependent changes in reveal locality and two-sided infilling. CORD adds only M trainable parameters, requires no additional backbone forward passes, and incurs only inference latency overhead.

Primary area:
foundation or frontier models, including LLMs
Keywords:
diffusion language models, masked diffusion models, decoding order

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ig2BOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Cord, a lightweight prompt-conditioned policy that retains confidence decoding as a reference and learns when and how to depart from it through geometric residuals over generation direction. Experiments on several reasoning, code and constraint-satisfaction benchmarks show improved task success and further improves over the best per-domain fixed residual.

Critical strengths
  1. The overall method is simple and computational cheap. Moving exploration from per-step position actions to a single 24D coefficient program per rollout is good idea. It turns a long-horizon credit-assignment problem into a contextual bandit.
Critical weaknesses
  1. The claim that learned prompt and state adaptive residuals are important is not supported. For the Table 1, the gap between CORD and Best per-domain fixed residuals are very close, and some even lower.
  2. The paper did not compare with any learned based policy baseline like [1]. And every experiment uses one token per step, which curernt dllms often uses multi token per step or block decoding. It is unclear whether how CORD would perform in these settings.

[1] Learning Unmasking Policies for Diffusion Language Models, Jazbec et al, 2025.

Questions for authors

Please check the weakness above.

Reviewer_d2fuOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the "reveal order" in masked diffusion language models (dLLMs), noting that the position for which a model currently holds the highest confidence is not necessarily the one most deserving of being revealed first. The authors propose CORD, which augments confidence scores with three types of "geometric residuals" and employs a prompt-conditioned policy to dynamically adjust the weights of these components based on the current generation state. During training, the dLLM backbone remains frozen, and a low-dimensional policy is learned via rewards derived from the complete, ground-truth answers. Across three backbones and five tasks, CORD improves the average performance from 30.6% (the confidence-based baseline) to 50.5%; however, compared to the best-performing fixed-residual strategy for each specific domain, the gain is marginal. Consequently, while the experiments provide strong evidence for the value of optimizing the reveal order, they also suggest that the additional benefit of dynamic weight adjustment over a well-tuned, fixed strategy is relatively small.

Critical strengths

None

Critical weaknesses
  1. The methodology is rather heuristic and lacks a clear motivation. The concepts of "direction/proximity/bridge," and the eight hand-crafted state features are all chosen arbitrarily. While the paper argues that confidence scores fail to capture the downstream value of revealing a specific position, it does not justify why these particular geometric features serve as appropriate proxies.

  2. The performance gains of CORD over the fixed baselines are marginal, and the baselines themselves are not clearly defined. In Table 1, CORD outperforms the best per-domain fixed-residual baselines by an average of only 0.5 percentage points, with no confidence intervals provided for this margin. Furthermore, details regarding fixed weights, the search space, selection data, and selection metrics are opaque, making it difficult to assess the stability or reproducibility of these gains.

  3. There is a lack of direct, controlled comparisons against existing learned ordering policies. The primary experiments focus on the training-free method and fails to benchmark CORD against established learned ordering policies using the same backbone; consequently, it remains unproven whether CORD’s coefficient-space design is superior to other approaches for learning ordering policies.

Questions for authors
  1. How exactly are the "best per-domain fixed residuals" obtained? What search space, validation data, and metric are used to select them?
  2. Could you include additional baselines for the "learning-based order policy"? This would help clarify whether the performance gains stem specifically from CORD’s coefficient-space design or from the general practice of training an order policy.
Reviewer_kBZfOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The author introduces CORD, which improves masked diffusion large language model (dLLM) by learning which tokens to unmask next. It keeps confidence-based decoding as the base, and adds learnable biases for direction, proximity, and filling gaps between revealed regions.

CORD is lightweight at inference and wins the ordinary confidence-based decoding and other baselines.

Critical strengths
  1. The paper reveals the importance of order of decoding for dLLM, and proposes a reasonable design.
  2. The flow of the paper is easy to follow, with sufficient reasoning, ablations, experiments, and discussion.
  3. It's interesting to have analysis of each task on which residual channels contribute the most. Experiment on the overhead introduced is also showing the effectiveness of CORD.
Critical weaknesses
  1. Despite that the number is better than other baselines, it's slightly hard to tell if the method is statistically best since there is no multi-seed mean, standard deviation for the main result (Table 1, 2, 3).
  2. Only single-token decoding is evaluated, where one of the benefits of dLLM is multi-token decoding. It would be better to have some discussion on this topic.
Questions for authors

Suggestion for future work:

Have the authors considered training a separate CORD policy for each downstream task while keeping the dLLM frozen? This could position CORD as a lightweight alternative or complement to SFT/LoRA for task specialization. A comparison among task-specific CORD, jointly trained CORD, and task-specific fixed residuals on held-out examples would clarify its adaptation quality, data efficiency, and susceptibility to overfitting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.