Prior work has shown that instruction-tuned large language models (LLMs) are less well calibrated than their base pre-trained counterparts. However, little is known about the frequently used chat template’s effect on the calibration of conversational LLMs. In this work, we investigate the mechanisms driving this miscalibration by decoupling the effects of the post-training algorithm and the chat format. We find that, while instruction tuning fundamentally harms calibration, the chat template aggravates the issue through an “ownership bias” – models are significantly more confident in their <em>own</em> answers than in identical answers provided by a user. Extensive experiments across six recent open-weight LLMs, three benchmarks, and three confidence elicitation methods show that models assign up to 26% higher confidence to their own responses. Leveraging this insight, we propose a simple inference-time strategy: framing the model's answer as user input during confidence elicitation. This approach significantly reduces overconfidence and improves calibration by up to 26% without the need for retraining, narrowing the gap between base and instruction-tuned models.</p>\n","updatedAt":"2026-06-11T07:22:31.583Z","author":{"_id":"64d4b251071f4a335c97264e","avatarUrl":"/avatars/c14ae3eea034c1b39186a194425e9989.svg","fullname":"Mario Sanz","name":"mario-sanz","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9416484832763672},"editors":["mario-sanz"],"editorAvatarUrls":["/avatars/c14ae3eea034c1b39186a194425e9989.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.03437","authors":[{"_id":"6a202b6b15100c5272a84226","user":{"_id":"64d4b251071f4a335c97264e","avatarUrl":"/avatars/c14ae3eea034c1b39186a194425e9989.svg","isPro":false,"fullname":"Mario Sanz","user":"mario-sanz","type":"user","name":"mario-sanz"},"name":"Mario Sanz-Guerrero","status":"claimed_verified","statusLastChangedAt":"2026-06-05T15:09:11.740Z","hidden":false},{"_id":"6a202b6b15100c5272a84227","name":"Manuel Mager","hidden":false},{"_id":"6a202b6b15100c5272a84228","name":"Katharina von der Wense","hidden":false}],"publishedAt":"2026-06-02T00:00:00.000Z","submittedOnDailyAt":"2026-06-11T00:00:00.000Z","title":"Large Language Models Are Overconfident in Their Own Responses","submittedOnDailyBy":{"_id":"64d4b251071f4a335c97264e","avatarUrl":"/avatars/c14ae3eea034c1b39186a194425e9989.svg","isPro":false,"fullname":"Mario Sanz","user":"mario-sanz","type":"user","name":"mario-sanz"},"summary":"Prior work has shown that instruction-tuned large language models (LLMs) are less well calibrated than their base pre-trained counterparts. However, little is known about the frequently used chat template's effect on the calibration of conversational LLMs. In this work, we investigate the mechanisms driving this miscalibration by decoupling the effects of the post-training algorithm and the chat format. We find that, while instruction tuning fundamentally harms calibration, the chat template aggravates the issue through an \"ownership bias\" -- models are significantly more confident in their own answers than in identical answers provided by a user. Extensive experiments across six recent open-weight LLMs, three benchmarks, and three confidence elicitation methods show that models assign up to 26% higher confidence to their own responses. Leveraging this insight, we propose a simple inference-time strategy: framing the model's answer as user input during confidence elicitation. This approach significantly reduces overconfidence and improves calibration by up to 26% without the need for retraining, narrowing the gap between base and instruction-tuned models.","upvotes":3,"discussionId":"6a202b6b15100c5272a8422d","ai_summary":"Instruction tuning degrades calibration in large language models, with chat templates exacerbating overconfidence through ownership bias, which can be mitigated by reframing model responses as user input during confidence assessment.","ai_keywords":["instruction-tuned large language models","calibration","chat template","ownership bias","confidence elicitation","overconfidence","retraining"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64d4b251071f4a335c97264e","avatarUrl":"/avatars/c14ae3eea034c1b39186a194425e9989.svg","isPro":false,"fullname":"Mario Sanz","user":"mario-sanz","type":"user"},{"_id":"65e0e6fa4394fc3d1b59627a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65e0e6fa4394fc3d1b59627a/rlKw_UdH3MpmpgUcchMgB.jpeg","isPro":false,"fullname":"Minh Duc Bui","user":"MinhDucBui","type":"user"},{"_id":"6a2ae6c2e36bc84d91b6e7cc","avatarUrl":"/avatars/abf4b4c0020f9332b6827952cc53163e.svg","isPro":false,"fullname":"mmgood","user":"mmgood","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.03437.md"}">
Large Language Models Are Overconfident in Their Own Responses
Abstract
Instruction tuning degrades calibration in large language models, with chat templates exacerbating overconfidence through ownership bias, which can be mitigated by reframing model responses as user input during confidence assessment.
Prior work has shown that instruction-tuned large language models (LLMs) are less well calibrated than their base pre-trained counterparts. However, little is known about the frequently used chat template's effect on the calibration of conversational LLMs. In this work, we investigate the mechanisms driving this miscalibration by decoupling the effects of the post-training algorithm and the chat format. We find that, while instruction tuning fundamentally harms calibration, the chat template aggravates the issue through an "ownership bias" -- models are significantly more confident in their own answers than in identical answers provided by a user. Extensive experiments across six recent open-weight LLMs, three benchmarks, and three confidence elicitation methods show that models assign up to 26% higher confidence to their own responses. Leveraging this insight, we propose a simple inference-time strategy: framing the model's answer as user input during confidence elicitation. This approach significantly reduces overconfidence and improves calibration by up to 26% without the need for retraining, narrowing the gap between base and instruction-tuned models.
Community
Prior work has shown that instruction-tuned large language models (LLMs) are less well calibrated than their base pre-trained counterparts. However, little is known about the frequently used chat template’s effect on the calibration of conversational LLMs. In this work, we investigate the mechanisms driving this miscalibration by decoupling the effects of the post-training algorithm and the chat format. We find that, while instruction tuning fundamentally harms calibration, the chat template aggravates the issue through an “ownership bias” – models are significantly more confident in their own answers than in identical answers provided by a user. Extensive experiments across six recent open-weight LLMs, three benchmarks, and three confidence elicitation methods show that models assign up to 26% higher confidence to their own responses. Leveraging this insight, we propose a simple inference-time strategy: framing the model's answer as user input during confidence elicitation. This approach significantly reduces overconfidence and improves calibration by up to 26% without the need for retraining, narrowing the gap between base and instruction-tuned models.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2606.03437 in a model README.md to link it from this page.
Cite arxiv.org/abs/2606.03437 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2606.03437 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.