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

2026-05-27

AI Summary

This work addresses the issue of premature decoding in fully non-autoregressive diffusion language models, where high-confidence token selection—particularly at the end-of-text (EOT) position—often leads to incomplete generations. To mitigate this without requiring additional training, the authors propose a suffix-anchor-guided approach that dynamically modulates the confidence scores of tokens near strategically inserted anchor positions. By integrating suffix anchor insertion with a decoding-progress-aware confidence adjustment mechanism, the method effectively balances generation completeness and decoding timing. Experiments across diverse tasks—including plain text generation, vision-language reasoning, and code synthesis—demonstrate consistent performance gains over explicit EOT suppression strategies, while preserving the inherent parallel decoding efficiency of non-autoregressive models.

Abstract

Diffusion language models decode text by iteratively denoising masked token sequences, making the choice of which positions to decode a central inference-time decision. Most training-free decoding strategies use model confidence for position selection, assuming that high-confidence positions are ready to be decoded. In this work, we revisit this assumption by studying when confidence misleads fully non-autoregressive (fully non-AR) decoding. EOT tokens can receive high confidence and cause incomplete generation; inserting a suffix anchor can mitigate this issue but introduces local overconfidence near the anchor, causing anchor-adjacent tokens to be decoded too early. To address these issues, we propose Suffix-Anchored Confidence Modulation, a simple training-free method that inserts a short suffix anchor to encourage response completion and modulates confidence near the anchor according to decoding progress. This preserves the response-completion benefit of suffix anchoring while reducing premature decoding of anchor-adjacent tokens. Across text-only reasoning, vision-language reasoning, and code-generation benchmarks, our method consistently improves confidence-based fully non-AR decoding, outperforms explicit EOT suppression, and preserves the parallel decoding advantage of fully non-AR generation.

Problem

• diffusion language models
• non-autoregressive decoding
• confidence misleading
• suffix anchoring
• premature decoding

Innovation

• Suffix Anchoring
• Confidence Modulation
• Diffusion Language Models
• Fully Non-Autoregressive Decoding
• Training-Free Decoding