@inproceedings{tharakan-tiku-2026-alignment,
title = "Alignment Sharpens Elicited Label Distributions under Logit and Sampled Elicitation",
author = "Tharakan, Christy Jacob and
Tiku, Kohsheen",
editor = "Xu, Shanshan and
T.Y.S.S., Santosh and
Kurniawan, Kemal and
Peng, Siyao and
Wallbridge, Sarenne and
Koto, Fajri and
Dziri, Nouha and
Plank, Barbara",
booktitle = "Proceedings of the First Workshop on Pluralistic {AI} {NLP}: Diversity-aware, Sociotechnical, Responsible Alignment ({PANDORA} 2026)",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.pandora-1.9/",
pages = "169--190",
ISBN = "979-8-89176-484-2",
abstract = "Alignment may encourage a model to commit to one answer even when human judgments differ. We study this effect by comparing base and aligned checkpoints on ChaosNLI and Measuring Hate Speech using logit, sampled, and verbalized elicitation. Across eight model pairs from 7B to 32B, including a mixture-of-experts model, alignment lowers the entropy of logit-elicited label distributions. Sampled elicitation agrees on the three pairs tested. The effect survives prompt-format, answer-order, word-only, and task-compliance controls. Its implications for human fit are less straightforward: Jensen{--}Shannon distance worsens on ChaosNLI, but changes in both directions on the native MHS hate-speech task. Six of seven aligned checkpoints in the main sweep fit ChaosNLI annotations worse than a uniform predictor, yet temperature scaling to base mean entropy removes much of the base{--}aligned gap. Majority-label agreement also changes in both directions. Verbalized probabilities show a different pattern, with entropy increases and decreases across three families and substantial sensitivity to the prompt{'}s numeric example in some base checkpoints. Staged comparisons of two Ai2 lineages further show that the contributions of SFT and DPO depend on task and prompt format. These results support a consistent sharpening effect under logit and sampled elicitation, while showing why entropy alone cannot establish whether alignment reduces a model{'}s ability to represent human disagreement."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tharakan-tiku-2026-alignment">
<titleInfo>
<title>Alignment Sharpens Elicited Label Distributions under Logit and Sampled Elicitation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Christy</namePart>
<namePart type="given">Jacob</namePart>
<namePart type="family">Tharakan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kohsheen</namePart>
<namePart type="family">Tiku</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the First Workshop on Pluralistic AI NLP: Diversity-aware, Sociotechnical, Responsible Alignment (PANDORA 2026)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shanshan</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Santosh</namePart>
<namePart type="family">T.Y.S.S.</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kemal</namePart>
<namePart type="family">Kurniawan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Siyao</namePart>
<namePart type="family">Peng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sarenne</namePart>
<namePart type="family">Wallbridge</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fajri</namePart>
<namePart type="family">Koto</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nouha</namePart>
<namePart type="family">Dziri</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Barbara</namePart>
<namePart type="family">Plank</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-484-2</identifier>
</relatedItem>
<abstract>Alignment may encourage a model to commit to one answer even when human judgments differ. We study this effect by comparing base and aligned checkpoints on ChaosNLI and Measuring Hate Speech using logit, sampled, and verbalized elicitation. Across eight model pairs from 7B to 32B, including a mixture-of-experts model, alignment lowers the entropy of logit-elicited label distributions. Sampled elicitation agrees on the three pairs tested. The effect survives prompt-format, answer-order, word-only, and task-compliance controls. Its implications for human fit are less straightforward: Jensen–Shannon distance worsens on ChaosNLI, but changes in both directions on the native MHS hate-speech task. Six of seven aligned checkpoints in the main sweep fit ChaosNLI annotations worse than a uniform predictor, yet temperature scaling to base mean entropy removes much of the base–aligned gap. Majority-label agreement also changes in both directions. Verbalized probabilities show a different pattern, with entropy increases and decreases across three families and substantial sensitivity to the prompt’s numeric example in some base checkpoints. Staged comparisons of two Ai2 lineages further show that the contributions of SFT and DPO depend on task and prompt format. These results support a consistent sharpening effect under logit and sampled elicitation, while showing why entropy alone cannot establish whether alignment reduces a model’s ability to represent human disagreement.</abstract>
<identifier type="citekey">tharakan-tiku-2026-alignment</identifier>
<location>
<url>https://aclanthology.org/2026.pandora-1.9/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>169</start>
<end>190</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Alignment Sharpens Elicited Label Distributions under Logit and Sampled Elicitation
%A Tharakan, Christy Jacob
%A Tiku, Kohsheen
%Y Xu, Shanshan
%Y T.Y.S.S., Santosh
%Y Kurniawan, Kemal
%Y Peng, Siyao
%Y Wallbridge, Sarenne
%Y Koto, Fajri
%Y Dziri, Nouha
%Y Plank, Barbara
%S Proceedings of the First Workshop on Pluralistic AI NLP: Diversity-aware, Sociotechnical, Responsible Alignment (PANDORA 2026)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-484-2
%F tharakan-tiku-2026-alignment
%X Alignment may encourage a model to commit to one answer even when human judgments differ. We study this effect by comparing base and aligned checkpoints on ChaosNLI and Measuring Hate Speech using logit, sampled, and verbalized elicitation. Across eight model pairs from 7B to 32B, including a mixture-of-experts model, alignment lowers the entropy of logit-elicited label distributions. Sampled elicitation agrees on the three pairs tested. The effect survives prompt-format, answer-order, word-only, and task-compliance controls. Its implications for human fit are less straightforward: Jensen–Shannon distance worsens on ChaosNLI, but changes in both directions on the native MHS hate-speech task. Six of seven aligned checkpoints in the main sweep fit ChaosNLI annotations worse than a uniform predictor, yet temperature scaling to base mean entropy removes much of the base–aligned gap. Majority-label agreement also changes in both directions. Verbalized probabilities show a different pattern, with entropy increases and decreases across three families and substantial sensitivity to the prompt’s numeric example in some base checkpoints. Staged comparisons of two Ai2 lineages further show that the contributions of SFT and DPO depend on task and prompt format. These results support a consistent sharpening effect under logit and sampled elicitation, while showing why entropy alone cannot establish whether alignment reduces a model’s ability to represent human disagreement.
%U https://aclanthology.org/2026.pandora-1.9/
%P 169-190
Markdown (Informal)
[Alignment Sharpens Elicited Label Distributions under Logit and Sampled Elicitation](https://aclanthology.org/2026.pandora-1.9/) (Tharakan & Tiku, PANDORA 2026)
ACL