@inproceedings{savcisens-eliassi-rad-2026-trilemma,
title = "The Trilemma of Truth in Large Language Models",
author = "Savcisens, Germans and
Eliassi-Rad, Tina",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.7/",
pages = "120--194",
ISBN = "979-8-89176-459-0",
abstract = "The public often attributes human-like qualities to large language models (LLMs), assuming that they ``know'' certain things. In reality, LLMs encode information retained during training as internal probabilistic knowledge. This study examines existing methods for probing the veracity of that knowledge and identifies three flawed underlying assumptions. To address these flaws, we introduce `sAwMIL{`} (Sparse-Aware Multiple-Instance Learning), a multiclass probing framework that combines multiple-instance learning with conformal prediction. `sAwMIL{`} leverages LLMs' internal representations to classify statements as {\_}true{\_}, {\_}false{\_}, or {\_}neither{\_}. We evaluate `sAwMIL{`} across 16 open-source LLMs, including default and chat-based variants, using three new curated datasets. Our results show that (1) common probing methods fail to provide a reliable and transferable veracity direction and, in some settings, perform worse than zero-shot prompting; (2) truth and falsehood are not encoded symmetrically; and (3) LLMs encode a third type of signal that is distinct from both true and false."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="savcisens-eliassi-rad-2026-trilemma">
<titleInfo>
<title>The Trilemma of Truth in Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Germans</namePart>
<namePart type="family">Savcisens</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tina</namePart>
<namePart type="family">Eliassi-Rad</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 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>The public often attributes human-like qualities to large language models (LLMs), assuming that they “know” certain things. In reality, LLMs encode information retained during training as internal probabilistic knowledge. This study examines existing methods for probing the veracity of that knowledge and identifies three flawed underlying assumptions. To address these flaws, we introduce ‘sAwMIL‘ (Sparse-Aware Multiple-Instance Learning), a multiclass probing framework that combines multiple-instance learning with conformal prediction. ‘sAwMIL‘ leverages LLMs’ internal representations to classify statements as _true_, _false_, or _neither_. We evaluate ‘sAwMIL‘ across 16 open-source LLMs, including default and chat-based variants, using three new curated datasets. Our results show that (1) common probing methods fail to provide a reliable and transferable veracity direction and, in some settings, perform worse than zero-shot prompting; (2) truth and falsehood are not encoded symmetrically; and (3) LLMs encode a third type of signal that is distinct from both true and false.</abstract>
<identifier type="citekey">savcisens-eliassi-rad-2026-trilemma</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.7/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>120</start>
<end>194</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T The Trilemma of Truth in Large Language Models
%A Savcisens, Germans
%A Eliassi-Rad, Tina
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F savcisens-eliassi-rad-2026-trilemma
%X The public often attributes human-like qualities to large language models (LLMs), assuming that they “know” certain things. In reality, LLMs encode information retained during training as internal probabilistic knowledge. This study examines existing methods for probing the veracity of that knowledge and identifies three flawed underlying assumptions. To address these flaws, we introduce ‘sAwMIL‘ (Sparse-Aware Multiple-Instance Learning), a multiclass probing framework that combines multiple-instance learning with conformal prediction. ‘sAwMIL‘ leverages LLMs’ internal representations to classify statements as _true_, _false_, or _neither_. We evaluate ‘sAwMIL‘ across 16 open-source LLMs, including default and chat-based variants, using three new curated datasets. Our results show that (1) common probing methods fail to provide a reliable and transferable veracity direction and, in some settings, perform worse than zero-shot prompting; (2) truth and falsehood are not encoded symmetrically; and (3) LLMs encode a third type of signal that is distinct from both true and false.
%U https://aclanthology.org/2026.emnlp-main.7/
%P 120-194
Markdown (Informal)
[The Trilemma of Truth in Large Language Models](https://aclanthology.org/2026.emnlp-main.7/) (Savcisens & Eliassi-Rad, EMNLP 2026)
ACL
- Germans Savcisens and Tina Eliassi-Rad. 2026. The Trilemma of Truth in Large Language Models. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 120–194, Budapest, Hungary. Association for Computational Linguistics.