@inproceedings{nandi-chakraborty-2026-plausieve,
title = "{PLAUSIEVE}:A Distribution Based Framework for Measuring Model Belief Beyond Logits",
author = "Nandi, Arghodeep and
Chakraborty, Tanmoy",
editor = "Zhang, Yue and
Xu, Ruifeng and
Purwarianti, Ayu and
Rong, Tong and
Suzuki, Jun",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {IJCNLP}-{AACL} 2026",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-aacl.17/",
pages = "300--328",
ISBN = "979-8-89176-444-6",
abstract = "Prompting has transformed large language models into general-purpose decision systems, where token-level probabilities are often treated as indicators of model preference and correctness. However, directly interpreting these probabilities is challenging due to calibration disparities, tokenizer variability, and inconsistent output distributions across models, limiting their statistical comparability. This raises a fundamental question: *when can a model be said to be confident in its decision or preference?* Addressing this question remains an open and important area of research. We introduce **PLAUSIEVE**, a novel distribution-based statistical framework that estimates model preference through a forward pass without token generation. **PLAUSIEVE** provides a calibration-robust and model-independent mechanism for quantifying preference without relying on fixed thresholds or individual tokens. We apply **PLAUSIEVE** in estimating model belief, across two prompt settings (direct and epistemic), two confidence levels, and five models from four families to study prompt{--}architecture interactions. Our analysis shows significant improvements under epistemic prompting, with nearly 100{\%} gains in certain model combinations. Notably, **Llama-3.2-3B** outperforms the larger **Gemma-7B** under epistemic prompting in terms of both decisiveness and veracity, demonstrating that model architecture plays a more significant role than scale in shaping belief formation."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="nandi-chakraborty-2026-plausieve">
<titleInfo>
<title>PLAUSIEVE:A Distribution Based Framework for Measuring Model Belief Beyond Logits</title>
</titleInfo>
<name type="personal">
<namePart type="given">Arghodeep</namePart>
<namePart type="family">Nandi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tanmoy</namePart>
<namePart type="family">Chakraborty</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruifeng</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ayu</namePart>
<namePart type="family">Purwarianti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tong</namePart>
<namePart type="family">Rong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jun</namePart>
<namePart type="family">Suzuki</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-444-6</identifier>
</relatedItem>
<abstract>Prompting has transformed large language models into general-purpose decision systems, where token-level probabilities are often treated as indicators of model preference and correctness. However, directly interpreting these probabilities is challenging due to calibration disparities, tokenizer variability, and inconsistent output distributions across models, limiting their statistical comparability. This raises a fundamental question: *when can a model be said to be confident in its decision or preference?* Addressing this question remains an open and important area of research. We introduce **PLAUSIEVE**, a novel distribution-based statistical framework that estimates model preference through a forward pass without token generation. **PLAUSIEVE** provides a calibration-robust and model-independent mechanism for quantifying preference without relying on fixed thresholds or individual tokens. We apply **PLAUSIEVE** in estimating model belief, across two prompt settings (direct and epistemic), two confidence levels, and five models from four families to study prompt–architecture interactions. Our analysis shows significant improvements under epistemic prompting, with nearly 100% gains in certain model combinations. Notably, **Llama-3.2-3B** outperforms the larger **Gemma-7B** under epistemic prompting in terms of both decisiveness and veracity, demonstrating that model architecture plays a more significant role than scale in shaping belief formation.</abstract>
<identifier type="citekey">nandi-chakraborty-2026-plausieve</identifier>
<location>
<url>https://aclanthology.org/2026.findings-aacl.17/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>300</start>
<end>328</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T PLAUSIEVE:A Distribution Based Framework for Measuring Model Belief Beyond Logits
%A Nandi, Arghodeep
%A Chakraborty, Tanmoy
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F nandi-chakraborty-2026-plausieve
%X Prompting has transformed large language models into general-purpose decision systems, where token-level probabilities are often treated as indicators of model preference and correctness. However, directly interpreting these probabilities is challenging due to calibration disparities, tokenizer variability, and inconsistent output distributions across models, limiting their statistical comparability. This raises a fundamental question: *when can a model be said to be confident in its decision or preference?* Addressing this question remains an open and important area of research. We introduce **PLAUSIEVE**, a novel distribution-based statistical framework that estimates model preference through a forward pass without token generation. **PLAUSIEVE** provides a calibration-robust and model-independent mechanism for quantifying preference without relying on fixed thresholds or individual tokens. We apply **PLAUSIEVE** in estimating model belief, across two prompt settings (direct and epistemic), two confidence levels, and five models from four families to study prompt–architecture interactions. Our analysis shows significant improvements under epistemic prompting, with nearly 100% gains in certain model combinations. Notably, **Llama-3.2-3B** outperforms the larger **Gemma-7B** under epistemic prompting in terms of both decisiveness and veracity, demonstrating that model architecture plays a more significant role than scale in shaping belief formation.
%U https://aclanthology.org/2026.findings-aacl.17/
%P 300-328
Markdown (Informal)
[PLAUSIEVE:A Distribution Based Framework for Measuring Model Belief Beyond Logits](https://aclanthology.org/2026.findings-aacl.17/) (Nandi & Chakraborty, Findings 2026)
ACL