PaperlayerView exact revision on arXiv

Source-linked AI summary

When Confidence Misleads: Suffix Anchoring and Anchor-Proximity Confidence Modulation for Diffusion Language Models

Jungwon Park, Jimyeong Kim, Jungmin Ko, Nojun Kwak, Wonjong Rhee

arXiv:2605.28181v2cs.CL

TL;DR

Fully non-AR diffusion language-model decoding can misinterpret confidence, causing premature EOT prediction or early commitment near a suffix anchor. The paper proposes training-free Suffix-Anchored Confidence Modulation, which combines suffix anchoring with progress-dependent local confidence adjustment. Across text-only, vision-language, and code-generation benchmarks, it consistently outperforms diverse fully non-AR and semi-AR baselines, with particularly large gains under highly parallel decoding.

  • Problem

    Confidence-based position selection can commit masked positions without sufficient context in fully non-AR decoding, including premature EOT tokens and anchor-adjacent positions.

  • Method

    Suffix-Anchored Confidence Modulation uses a minimal suffix anchor and down-weights nearby confidence early, relaxing the modulation as decoding progress provides more context.

  • Results

    Across text-only reasoning, vision-language reasoning, and code-generation benchmarks, the method consistently outperforms diverse fully non-AR and semi-AR baselines, especially under highly parallel decoding.

  • Takeaways & Limitations

    Improving sequence-wide position selection benefits the fully non-AR regime relative to block-wise semi-AR decoding under highly parallel generation.

  • Takeaways & Limitations

    The method does not address errors from insufficient model knowledge or reasoning ability and may yield smaller gains when incorrect token prediction, rather than position selection, dominates.

Abstract

from arXiv · show

Diffusion language models generate text by iteratively selecting and denoising masked positions, making position selection a central inference-time decision. Most training-free methods rely on model confidence, assuming that high-confidence positions are ready to be decoded. However, this assumption can fail in fully non-autoregressive (fully non-AR) decoding, where selection spans the entire masked response sequence. This can cause incomplete generation, partly due to EOT overconfidence. Existing remedies directly suppress EOT tokens, but provide limited gains or require additional fine-tuning. We show that placing even a semantically minimal suffix anchor near the end of the response region discourages premature EOT prediction and promotes complete generation. However, despite providing little supporting context, the anchor makes nearby positions highly confident and causes them to be decoded too early. We therefore propose Suffix-Anchored Confidence Modulation, a simple training-free method that uses suffix anchoring with progress-dependent confidence modulation near the anchor. Across text-only reasoning, vision-language reasoning, and code-generation benchmarks, our method outperforms diverse fully non-AR and semi-AR baselines. It also achieves particularly large gains under highly parallel decoding, demonstrating the benefit of improving the fully non-AR regime over block-wise semi-AR regime.

1 Introduction

Diffusion language models must jointly predict tokens and choose which masked positions to unmask, but confidence-based selection can commit positions before sufficient context exists. The paper introduces suffix anchoring with progress-dependent confidence modulation to reduce premature EOT prediction and anchor-induced overconfidence, improving fully non-AR decoding across multiple benchmark families.

  • Problem: DLMs iteratively denoise masked positions in parallel, making inference-time position selection central to generation.At each step, the model must determine both token predictions and which masked positions to unmask.
  • Problem: Confidence-based selection can favor positions that lack sufficient supporting context in fully non-AR decoding.Fully non-AR selection spans the entire masked response sequence, unlike strictly left-to-right generation.
  • Problem: Instruction-tuned DLMs may assign high confidence to EOT tokens, producing incomplete or extremely short outputs.Existing remedies include direct EOT suppression, additional training, and semi-AR decoding, but each has stated trade-offs.
  • Approach: A short suffix anchor, such as “The answer is” or “.”, substantially reduces incomplete generation without requiring detailed structural information.The anchor discourages premature EOT prediction but can create high confidence near the anchor despite little supporting context.
  • Approach: Suffix-Anchored Confidence Modulation down-weights confidence near the anchor early and gradually relaxes the modulation as more context is decoded.This aims to reduce premature commitment of anchor-adjacent tokens while retaining suffix anchoring’s completion benefit.
  • Results: Across text-only reasoning, vision-language reasoning, and code-generation benchmarks, the method consistently improves fully non-AR decoding and outperforms diverse baselines.The largest gains occur under highly parallel decoding, including comparisons with semi-AR alternatives.

2 Related Work

The supplied passages situate diffusion language modeling within broader diffusion-generation research and depict confidence-selection failure modes relevant to decoding. They also include an unrelated reasoning question and do not provide substantive details about the cited related-work methods.

  • Diffusion modeling: Diffusion models have achieved strong generative performance in continuous domains such as image and video generation.The passage lists prior work extending diffusion models across these continuous domains.
  • Diffusion modeling: Prior work extended diffusion to discrete text generation through categorical corruption, discrete-state Markov chains, and continuous-time variants.Subsequent studies developed masked diffusion language models and clarified their connections, though the supplied passage is truncated.
  • Other supplied material: The supplied passages include a tennis-ball word problem but provide no related-work claim about it.It is not substantively connected to the section’s research discussion.
  • Decoding failure modes: The figure contrasts premature EOT unmasking with anchor-induced high confidence near the suffix anchor.Darker blue boxes indicate positions decoded at later steps, and ∅ denotes the EOT token.
  • Decoding failure modes: Suffix anchoring reduces the EOT token ratio while concentrating positions decoded in the first 15% of steps near the anchor.The response region contains 256 tokens divided into 32 bins; results are computed on GSM8K.

3 When Confidence Misleads Position Selection

Fully non-AR confidence selection can fail in two ways: high-confidence EOT tokens can truncate outputs, while suffix anchoring shifts premature confidence toward nearby positions. The section motivates retaining anchoring’s completion benefit while reducing this local overconfidence.

  • Setup: Fully non-AR decoding can unmask any masked response position, including distant positions whose supporting context has not yet been generated.This broad candidate set creates opportunities for premature high-confidence selection.
  • Failure mode 1: EOT overconfidence: Naive decoding may assign high probability to EOT near the response-region end, causing early unmasking and incomplete outputs.This demonstrates that high confidence does not necessarily mean a position is ready to decode.
  • Suffix anchoring: A short suffix anchor conditions predictions toward non-EOT content and substantially reduces the average EOT ratio.The analysis uses “The answer is” as a fixed anchor near the end of the response region.
  • Failure mode 2: Anchor-induced local overconfidence: Suffix anchoring also makes anchor-adjacent positions highly confident before sufficient context has been decoded, producing inaccurate early tokens.The first 15% of decoding steps contain a disproportionately large fraction of positions near the anchor.
  • Summary: The paper identifies anchor-induced local overconfidence as a second failure mode that motivates confidence modulation near the anchor.The intended strategy preserves completion gains while reducing premature unmasking.

4 Method

The method combines a short suffix anchor with confidence modulation that is strongest near the anchor early in decoding and relaxes as decoding progresses. This preserves the anchor’s completion benefit while reducing premature decoding of anchor-adjacent positions.

  • Method overview: The method inserts a short suffix anchor and reweights confidence scores near it before selecting masked positions for unmasking.The anchor can be a phrase such as “The answer is” or a minimal token such as “.”; reweighted scores rank positions for decoding.
  • Position selection: The reweighted confidence score is used to rank masked positions, while the method remains compatible with different confidence definitions and requires no training or architectural changes.The base score may be maximum predicted-token probability or the gap between the top two probabilities.
  • Suffix anchoring: Suffix anchoring conditions predictions across the masked response region, encouraging non-EOT content up to the anchor.The anchor positions remain fixed throughout decoding and discourage premature EOT prediction.
  • Anchor-proximity weight: Anchor-proximity weights assign larger modulation to positions closer to the anchor and decay continuously with distance.Exponential decay remains positive without an explicit cutoff; κ controls decay rate, β controls modulation strength, and clipping bounds the weight.
  • Progress-dependent modulation: Confidence modulation is stronger early, when little context is decoded, and gradually relaxes as more tokens are unmasked.Decoding progress is based on the number of masked positions, with larger values indicating later stages; γ controls relaxation speed.

5 Experiments

The experiments evaluate Suffix-Anchored Confidence Modulation across text-only reasoning, vision-language reasoning, code generation, parallelism, ablations, and efficiency. The method consistently improves fully non-AR decoding, remains robust across response lengths and minimal anchors, and adds negligible inference overhead.

  • Experimental Setup: Experiments cover text-only reasoning, vision-language reasoning, and code generation using LLaDA, LLaDA-1.5, and LaViDa-Instruct.Benchmarks include GSM8K, MATH-500, StrategyQA, MMLU-Pro, MathVista, ChartQA, HumanEval, and MBPP.
  • Text-only reasoning: 53.88 and 54.77 average accuracy are achieved on LLaDA and LLaDA-1.5, respectively, exceeding Rainbow Padding and Fast-dLLM by 4.76 and 2.98 points.The method achieves the best result on every text-only benchmark for both models while remaining training-free and fully non-AR.
  • Vision-language and code generation: 45.67 average accuracy on vision-language reasoning outperforms Fast-dLLM by 6.88 points, with 46.04 accuracy on ChartQA and 45.30 on MathVista.The MathVista result is close to Template Infilling’s 45.60.
  • Fully non-AR advantage under high parallelism: 57.70 accuracy at T = 32 outperforms the best semi-AR result of 36.32 under the most highly parallel setting.The method outperforms every semi-AR configuration across step budgets, while the best semi-AR result falls to 11.37 when block size decreases from 64 to 8.
  • Progress-dependent modulation: 72.25 accuracy without progress dependence is below 76.88 with progress-dependent modulation, supporting gradual relaxation during decoding.The ablation evaluates the factor (1 −p(t))γ in Eq. (4).
  • Inference efficiency: Throughput changes from 25.02 to 24.93 tokens/s and latency from 10.23 to 10.27 s/sample compared with top-probability decoding.The method therefore introduces negligible inference overhead while improving decoding quality.

6 Conclusion

The paper identifies premature EOT prediction and anchor-induced local overconfidence as failures of confidence-based position selection in fully non-AR decoding. Suffix-Anchored Confidence Modulation addresses both with suffix anchoring and progress-dependent modulation, outperforming diverse baselines across tasks and especially under highly parallel decoding.

  • Confidence-based position selection can fail through premature EOT prediction and anchor-induced local overconfidence in fully non-AR DLM decoding.
  • Suffix-Anchored Confidence Modulation combines suffix anchoring with progress-dependent confidence modulation near the anchor.
  • The method consistently outperforms diverse semi-AR and fully non-AR baselines across text-only reasoning, vision-language reasoning, and code-generation benchmarks.
  • The method remains effective under highly parallel decoding, supporting sequence-wide position selection in the fully non-AR regime.

Limitations

The method’s scope is limited by its inference-time confidence reweighting and by fixed experimental choices. Its evaluation also covers representative rather than fully diverse tasks and settings.

  • Because the method does not update model parameters, it does not address errors from insufficient knowledge or reasoning ability.
  • Gains may be smaller when failures primarily reflect incorrect token predictions rather than premature or suboptimal position selection.
  • Main experiments use fixed suffix anchors and predefined anchor positions, although ablations show robustness across tested choices.
  • The optimal anchor, placement, or modulation strength may vary across tasks and output formats, motivating more adaptive strategies.
  • Experiments focus on representative text-only and vision-language DLMs and standard reasoning and code-generation benchmarks, leaving multilingual and more diverse multimodal evaluation open.

B.1 Models and Evaluation Splits

The evaluation uses specified benchmark datasets, model checkpoints, splits, and task-specific protocols. MathVista is evaluated on testmini because its test split lacks answer labels.

  • MathVista uses the testmini split because answer labels are unavailable for its test split.
  • Experiments use publicly available LLaDA, LLaDA-1.5, and LaViDa-Instruct checkpoints with reported evaluation splits and set sizes.
  • The evaluation covers text-only, vision-language, reasoning, and code-generation benchmarks using generative answer extraction and execution-based code evaluation.
  • Reasoning prompts include step-by-step instructions, while code-generation prompts follow benchmark-specific protocols.

B.3 Hyperparameter Selection and Sensitivity Analysis

Hyperparameters are selected with lightweight benchmark-specific procedures and reused when possible, while sensitivity analysis tests robustness across broader values and decoding settings.

  • Selected hyperparameters are documented per benchmark, with the GSM8K setting used when no training or validation split is available.
  • The sweep selects κ, β, and γ from predefined ranges using 128 randomly sampled examples, then reuses values across related decoding experiments when possible.
  • Sensitivity analysis varies each hyperparameter around the GSM8K setting on 256 training examples and tests wider ranges for κ and γ.
  • Performance remains stable across a wide range of hyperparameter values, while unmodified top-probability decoding and suffix anchoring alone score 15.63 and 55.86.
  • The algorithm inserts suffix anchors, computes anchor-proximity weights, reweights confidence scores during decoding, and returns the final response segment.
  • The reported gains can substantially exceed selected comparison scores, suggesting they do not rely on a narrowly tuned hyperparameter choice.
  • With top-margin decoding, the full method raises average accuracy from 23.98 to 51.07 for LLaDA and from 34.60 to 51.18 for LLaDA-1.5.
  • For code generation, average pass@1 rises from 27.23 to 30.67 on LLaDA and from 28.94 to 33.29 on LLaDA-1.5.

C.3 Ablation Over Suffix Anchors

The ablation tests whether suffix anchoring depends on a detailed phrase or a particular placement. Minimal anchors perform comparably to the default, and confidence modulation adds further gains.

  • Ablation Over Suffix Anchors: Across both LLaDA models, tested suffix anchors substantially improve decoding without an anchor, while confidence modulation consistently provides additional gains.
  • Ablation Over Suffix Anchors: The semantically minimal anchor “.” achieves 74.68 on LLaDA and 75.36 on LLaDA-1.5 with the full method, closely matching the default anchor.
  • Ablation Over Suffix Anchors: These results indicate that detailed response templates are unnecessary; a lightweight continuation cue can discourage premature EOT prediction.
  • Ablation Over Anchor Positions: Anchor-position ablations vary “The answer is” and “.” within the later response region, including a default position 20 positions before the end.
  • Ablation Over Anchor Positions: Position-dependent changes in EOT ratio support interpreting the suffix anchor as a local cue for non-EOT generation rather than a fixed response template.

D Qualitative Analysis

The qualitative analysis compares unmodified confidence decoding, suffix anchoring, and confidence modulation across reasoning examples, while visualizing how the proposed method resolves responses over time. It also evaluates applicability across decoding strategies, code generation, suffix anchors, and anchor positions.

  • Qualitative comparisons: The qualitative examples illustrate that suffix anchoring mitigates incomplete generation, while the full method additionally modulates confidence near the anchor.
  • Decoding dynamics: Figure 11 visualizes progressive response resolution while confidence modulation near the suffix anchor is gradually relaxed, allowing nearby positions to decode later.The example uses LLaDA 8B-Instruct on GSM8K with top-probability decoding.
  • Quantitative evaluations: Table 11 evaluates suffix anchoring and confidence modulation with both top-probability and top-margin decoding on four text-only reasoning benchmarks.The comparison includes the corresponding baseline, suffix anchoring without confidence modulation, and the full method.
  • Quantitative evaluations: Tables 12–14 extend evaluation to code generation and ablations over suffix anchors and anchor positions.Table 12 reports Pass@1 on HumanEval and MBPP; Tables 13–14 report GSM8K accuracy, with Table 14 also reporting EOT ratio and varying anchor positions.
Loading 2605.28181v2…