Prior work has studied linguistic diversity in generation tasks such as story generation. At the same time, lexical diversity can also be related to uncertainty over plausible model outputs. Instruction tuning has been shown to affect diversity in generative settings, while instruction-tuned models can also exhibit verbalized overconfidence. This raises a question: do changes in model confidence induced by instruction tuning correspond to changes in the lexical diversity of generated outputs? In our paper, we study this connection through the diversity of generated answer rationales.</p>\n","updatedAt":"2026-08-14T08:35:56.353Z","author":{"_id":"629a3dbcd496c6dcdebf41cc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg","fullname":"Irina Proskurina","name":"iproskurina","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":8,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.935979425907135},"editors":["iproskurina"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.13430","authors":[{"_id":"6a7ecf8a42823931a1f177fd","user":{"_id":"629a3dbcd496c6dcdebf41cc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg","isPro":true,"fullname":"Irina Proskurina","user":"iproskurina","type":"user","name":"iproskurina"},"name":"Irina Proskurina","status":"claimed_verified","statusLastChangedAt":"2026-08-14T08:45:04.824Z","hidden":false},{"_id":"6a7ecf8a42823931a1f177fe","name":"Mayank Kumar","hidden":false},{"_id":"6a7ecf8a42823931a1f177ff","name":"Oyindolapo O. Komolafe","hidden":false}],"publishedAt":"2026-08-13T00:00:00.000Z","submittedOnDailyAt":"2026-08-14T00:00:00.000Z","title":"Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity","submittedOnDailyBy":{"_id":"629a3dbcd496c6dcdebf41cc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1655113762275-629a3dbcd496c6dcdebf41cc.jpeg","isPro":true,"fullname":"Irina Proskurina","user":"iproskurina","type":"user","name":"iproskurina"},"summary":"Instruction-tuned language models achieve strong performance across a range of generation tasks, but have also recently been shown to exhibit verbalized overconfidence. In question answering, verbalized model overconfidence may be associated with the consistency of the generated supporting rationales. In this paper, we study whether corresponding changes in the lexical diversity of generated answer rationales accompany changes in model confidence induced by instruction tuning. We evaluate three matched base and instruction-tuned models across question-answering benchmarks and find that instruction tuning consistently alters answer confidence, despite limited changes in predictive accuracy and decreases in likelihood-based calibration. Secondly, we observe a non-uniform effect of instruction tuning on rationale diversity: cross-rationale diversity consistently decreases, whereas surface-level lexical diversity varies in both direction and magnitude across models and benchmarks. Finally, we find that these differences persist after controlling for answer selection and rationale length, confirming that confidence and rationale diversity capture distinct effects of instruction tuning.","upvotes":0,"discussionId":"6a7ecf8a42823931a1f17800","ai_summary":"Instruction tuning changes model confidence and reduces rationale diversity without improving calibration, indicating distinct effects on reasoning and certainty.","ai_keywords":["instruction-tuned language models","verbalized overconfidence","question answering","lexical diversity","rationale diversity","likelihood-based calibration"],"ai_summary_model":"thinkingmachines/Inkling-Small"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[],"acceptLanguages":["en"],"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.13430.md","query":{}}">
Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity
Abstract
Instruction tuning changes model confidence and reduces rationale diversity without improving calibration, indicating distinct effects on reasoning and certainty.
Instruction-tuned language models achieve strong performance across a range of generation tasks, but have also recently been shown to exhibit verbalized overconfidence. In question answering, verbalized model overconfidence may be associated with the consistency of the generated supporting rationales. In this paper, we study whether corresponding changes in the lexical diversity of generated answer rationales accompany changes in model confidence induced by instruction tuning. We evaluate three matched base and instruction-tuned models across question-answering benchmarks and find that instruction tuning consistently alters answer confidence, despite limited changes in predictive accuracy and decreases in likelihood-based calibration. Secondly, we observe a non-uniform effect of instruction tuning on rationale diversity: cross-rationale diversity consistently decreases, whereas surface-level lexical diversity varies in both direction and magnitude across models and benchmarks. Finally, we find that these differences persist after controlling for answer selection and rationale length, confirming that confidence and rationale diversity capture distinct effects of instruction tuning.
Community
Prior work has studied linguistic diversity in generation tasks such as story generation. At the same time, lexical diversity can also be related to uncertainty over plausible model outputs. Instruction tuning has been shown to affect diversity in generative settings, while instruction-tuned models can also exhibit verbalized overconfidence. This raises a question: do changes in model confidence induced by instruction tuning correspond to changes in the lexical diversity of generated outputs? In our paper, we study this connection through the diversity of generated answer rationales.
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/2608.13430 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.13430 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.13430 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.