@inproceedings{kulikov-etal-2026-factrot,
title = "{F}act{R}ot: Auditing Typed Fact Survival Under Iterative {LLM} Rewriting",
author = "Kulikov, Vladislav and
Mozikov, Mikhail and
Stepin, Aleksei and
Tolstokulakov, Boris and
Kabanov, Artyom and
Ivan, Nasonov and
Glazkov, Nikita and
Makarov, Ilya",
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.2472/",
pages = "56120--56129",
ISBN = "979-8-89176-459-0",
abstract = "FactRot measures whether typed factual mentions survive iterative LLM rewriting. We follow each ledger fact through every hop of a rewrite chain and score it separately, instead of measuring aggregate drift in the rewritten text. We score 1,113 ledger facts from 25 Wikipedia passages across 4 models, 4 rewrite operations, and 3 prompt conditions, with chains reaching depth 12. A GEE with passage clusters finds per-hop decay (OR=0.743 per additional hop, $p<.001$). Type-specific faithfulness instructions do not measurably outperform generic faithfulness at hop 5 ($\Delta=0.026$, 95{\%} CI [0.022, 0.073]); both improve survival over neutral prompting by roughly 0.13-0.16 absolute. On the depth-12 neutral subset, answer-restate rewriting is much more destructive than direct paraphrasing, and the frontier preview model has the highest hop-12 survival among the four models. Exact fact-type rankings, however, are unstable across models (Kendall $W=0.082$ at hop 5)."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kulikov-etal-2026-factrot">
<titleInfo>
<title>FactRot: Auditing Typed Fact Survival Under Iterative LLM Rewriting</title>
</titleInfo>
<name type="personal">
<namePart type="given">Vladislav</namePart>
<namePart type="family">Kulikov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mikhail</namePart>
<namePart type="family">Mozikov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aleksei</namePart>
<namePart type="family">Stepin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Boris</namePart>
<namePart type="family">Tolstokulakov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Artyom</namePart>
<namePart type="family">Kabanov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nasonov</namePart>
<namePart type="family">Ivan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nikita</namePart>
<namePart type="family">Glazkov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ilya</namePart>
<namePart type="family">Makarov</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>FactRot measures whether typed factual mentions survive iterative LLM rewriting. We follow each ledger fact through every hop of a rewrite chain and score it separately, instead of measuring aggregate drift in the rewritten text. We score 1,113 ledger facts from 25 Wikipedia passages across 4 models, 4 rewrite operations, and 3 prompt conditions, with chains reaching depth 12. A GEE with passage clusters finds per-hop decay (OR=0.743 per additional hop, p<.001). Type-specific faithfulness instructions do not measurably outperform generic faithfulness at hop 5 (Δ=0.026, 95% CI [0.022, 0.073]); both improve survival over neutral prompting by roughly 0.13-0.16 absolute. On the depth-12 neutral subset, answer-restate rewriting is much more destructive than direct paraphrasing, and the frontier preview model has the highest hop-12 survival among the four models. Exact fact-type rankings, however, are unstable across models (Kendall W=0.082 at hop 5).</abstract>
<identifier type="citekey">kulikov-etal-2026-factrot</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2472/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>56120</start>
<end>56129</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T FactRot: Auditing Typed Fact Survival Under Iterative LLM Rewriting
%A Kulikov, Vladislav
%A Mozikov, Mikhail
%A Stepin, Aleksei
%A Tolstokulakov, Boris
%A Kabanov, Artyom
%A Ivan, Nasonov
%A Glazkov, Nikita
%A Makarov, Ilya
%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 kulikov-etal-2026-factrot
%X FactRot measures whether typed factual mentions survive iterative LLM rewriting. We follow each ledger fact through every hop of a rewrite chain and score it separately, instead of measuring aggregate drift in the rewritten text. We score 1,113 ledger facts from 25 Wikipedia passages across 4 models, 4 rewrite operations, and 3 prompt conditions, with chains reaching depth 12. A GEE with passage clusters finds per-hop decay (OR=0.743 per additional hop, p<.001). Type-specific faithfulness instructions do not measurably outperform generic faithfulness at hop 5 (Δ=0.026, 95% CI [0.022, 0.073]); both improve survival over neutral prompting by roughly 0.13-0.16 absolute. On the depth-12 neutral subset, answer-restate rewriting is much more destructive than direct paraphrasing, and the frontier preview model has the highest hop-12 survival among the four models. Exact fact-type rankings, however, are unstable across models (Kendall W=0.082 at hop 5).
%U https://aclanthology.org/2026.emnlp-main.2472/
%P 56120-56129
Markdown (Informal)
[FactRot: Auditing Typed Fact Survival Under Iterative LLM Rewriting](https://aclanthology.org/2026.emnlp-main.2472/) (Kulikov et al., EMNLP 2026)
ACL
- Vladislav Kulikov, Mikhail Mozikov, Aleksei Stepin, Boris Tolstokulakov, Artyom Kabanov, Nasonov Ivan, Nikita Glazkov, and Ilya Makarov. 2026. FactRot: Auditing Typed Fact Survival Under Iterative LLM Rewriting. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 56120–56129, Budapest, Hungary. Association for Computational Linguistics.