This paper studies when confidence-based decoding can mislead fully non-autoregressive diffusion language models. One known failure case is EOT overconfidence, where the model assigns high confidence to end-of-text tokens too early, leading to extremely short or incomplete generations. The paper shows that inserting a short suffix anchor near the end of the response region can encourage the model to generate complete, meaningful content. However, this anchor can also make nearby positions overconfident, causing them to be decoded prematurely. To address this issue, the paper proposes Suffix-Anchored Confidence Modulation, a training-free decoding method that preserves the response-completion benefit of suffix anchoring while mitigating premature decoding near the anchor.</p>\n","updatedAt":"2026-06-01T11:27:01.984Z","author":{"_id":"63873047a616fb0ef6d71732","avatarUrl":"/avatars/a166f0bc4a0b72fc3bbd659ca5a4b63e.svg","fullname":"Jungwon Park","name":"JwonP","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8993055820465088},"editors":["JwonP"],"editorAvatarUrls":["/avatars/a166f0bc4a0b72fc3bbd659ca5a4b63e.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.28181","authors":[{"_id":"6a17b535b4840f905a674326","name":"Jungwon Park","hidden":false},{"_id":"6a17b535b4840f905a674327","name":"Jimyeong Kim","hidden":false},{"_id":"6a17b535b4840f905a674328","name":"Jungmin Ko","hidden":false},{"_id":"6a17b535b4840f905a674329","name":"Nojun Kwak","hidden":false},{"_id":"6a17b535b4840f905a67432a","name":"Wonjong Rhee","hidden":false}],"publishedAt":"2026-05-27T00:00:00.000Z","submittedOnDailyAt":"2026-06-01T00:00:00.000Z","title":"When Confidence Misleads: Suffix Anchoring and Anchor-Proximity Confidence Modulation for Diffusion Language Models","submittedOnDailyBy":{"_id":"63873047a616fb0ef6d71732","avatarUrl":"/avatars/a166f0bc4a0b72fc3bbd659ca5a4b63e.svg","isPro":false,"fullname":"Jungwon Park","user":"JwonP","type":"user","name":"JwonP"},"summary":"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.","upvotes":2,"discussionId":"6a17b535b4840f905a67432b","ai_summary":"Researchers investigate how confidence-based decoding in fully non-autoregressive models can be improved by addressing issues with EOT tokens and premature decoding through suffix-anchored confidence modulation.","ai_keywords":["diffusion language models","masked token sequences","fully non-autoregressive decoding","model confidence","EOT tokens","suffix anchor","confidence modulation","text reasoning","vision-language reasoning","code generation"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"63873047a616fb0ef6d71732","avatarUrl":"/avatars/a166f0bc4a0b72fc3bbd659ca5a4b63e.svg","isPro":false,"fullname":"Jungwon Park","user":"JwonP","type":"user"},{"_id":"661ab1f1fa3b144a381fa454","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/661ab1f1fa3b144a381fa454/IlpZBb9NCjo7ntFwMIH53.png","isPro":false,"fullname":"Urro","user":"urroxyz","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.28181.md"}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2605.28181 in a model README.md to link it from this page.

Cite arxiv.org/abs/2605.28181 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2605.28181 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.