@inproceedings{arun-etal-2026-geometric,
title = "Geometric Separation Without Functional Specialization: Auditing Subspace Disentanglement in Dual-Encoder {LLM} Paraphrase Detection",
author = "Arun, Govind and
Abraham, Ashish and
P, Gadha Lekshmi",
editor = "Sedoc, Jo{\~a}o and
Shu, Raphael and
Drozd, Aleksandr and
Akula, Arjun and
Rogers, Anna and
Giorgi, Salvatore and
Benton, Adrian",
booktitle = "Proceedings of the Seventh Workshop on Insights from Negative Results in {NLP}",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.insights-1.1/",
pages = "1--6",
ISBN = "979-8-89176-480-4",
abstract = "Explicit subspace disentanglement, enforcing orthogonality and simplex constraints on learned projections, is an increasingly common recipe for paraphrase models claimed to separate lexical, syntactic, and semantic information. A dual-encoder hierarchical model trained with these constraints on the DiPPer corpus of LLM-generated paraphrases produces near-perfect geometric separation, a pairwise Centered Kernel Alignment (CKA) of roughly 0.0007, yet that separation is not evidence of functional linguistic specialization. Three results establish this. First, near-zero CKA is largely preordained by an explicit orthogonality penalty, which converges to $4\times10^{-6}$. Second, each of the three subspaces, including the nominally syntactic one for which we identify no functional probe, classifies paraphrase in isolation at $F_1$ above 0.986, so the subspaces are not functionally distinct. Third, removing disentanglement at matched capacity costs only 0.013 $F_1$, leaving most of the reported gain attributable to dual-encoder capacity rather than disentanglement. Capacity-matched ablations and controlled functional probes, not CKA alone, are required to substantiate disentanglement claims; orthogonality penalties otherwise manufacture the very metric used to certify them."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="arun-etal-2026-geometric">
<titleInfo>
<title>Geometric Separation Without Functional Specialization: Auditing Subspace Disentanglement in Dual-Encoder LLM Paraphrase Detection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Govind</namePart>
<namePart type="family">Arun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ashish</namePart>
<namePart type="family">Abraham</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gadha</namePart>
<namePart type="given">Lekshmi</namePart>
<namePart type="family">P</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 Seventh Workshop on Insights from Negative Results in NLP</title>
</titleInfo>
<name type="personal">
<namePart type="given">João</namePart>
<namePart type="family">Sedoc</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Raphael</namePart>
<namePart type="family">Shu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aleksandr</namePart>
<namePart type="family">Drozd</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Arjun</namePart>
<namePart type="family">Akula</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Anna</namePart>
<namePart type="family">Rogers</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Salvatore</namePart>
<namePart type="family">Giorgi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Adrian</namePart>
<namePart type="family">Benton</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">979-8-89176-480-4</identifier>
</relatedItem>
<abstract>Explicit subspace disentanglement, enforcing orthogonality and simplex constraints on learned projections, is an increasingly common recipe for paraphrase models claimed to separate lexical, syntactic, and semantic information. A dual-encoder hierarchical model trained with these constraints on the DiPPer corpus of LLM-generated paraphrases produces near-perfect geometric separation, a pairwise Centered Kernel Alignment (CKA) of roughly 0.0007, yet that separation is not evidence of functional linguistic specialization. Three results establish this. First, near-zero CKA is largely preordained by an explicit orthogonality penalty, which converges to 4\times10⁻6. Second, each of the three subspaces, including the nominally syntactic one for which we identify no functional probe, classifies paraphrase in isolation at F₁ above 0.986, so the subspaces are not functionally distinct. Third, removing disentanglement at matched capacity costs only 0.013 F₁, leaving most of the reported gain attributable to dual-encoder capacity rather than disentanglement. Capacity-matched ablations and controlled functional probes, not CKA alone, are required to substantiate disentanglement claims; orthogonality penalties otherwise manufacture the very metric used to certify them.</abstract>
<identifier type="citekey">arun-etal-2026-geometric</identifier>
<location>
<url>https://aclanthology.org/2026.insights-1.1/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1</start>
<end>6</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Geometric Separation Without Functional Specialization: Auditing Subspace Disentanglement in Dual-Encoder LLM Paraphrase Detection
%A Arun, Govind
%A Abraham, Ashish
%A P, Gadha Lekshmi
%Y Sedoc, João
%Y Shu, Raphael
%Y Drozd, Aleksandr
%Y Akula, Arjun
%Y Rogers, Anna
%Y Giorgi, Salvatore
%Y Benton, Adrian
%S Proceedings of the Seventh Workshop on Insights from Negative Results in NLP
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-480-4
%F arun-etal-2026-geometric
%X Explicit subspace disentanglement, enforcing orthogonality and simplex constraints on learned projections, is an increasingly common recipe for paraphrase models claimed to separate lexical, syntactic, and semantic information. A dual-encoder hierarchical model trained with these constraints on the DiPPer corpus of LLM-generated paraphrases produces near-perfect geometric separation, a pairwise Centered Kernel Alignment (CKA) of roughly 0.0007, yet that separation is not evidence of functional linguistic specialization. Three results establish this. First, near-zero CKA is largely preordained by an explicit orthogonality penalty, which converges to 4\times10⁻6. Second, each of the three subspaces, including the nominally syntactic one for which we identify no functional probe, classifies paraphrase in isolation at F₁ above 0.986, so the subspaces are not functionally distinct. Third, removing disentanglement at matched capacity costs only 0.013 F₁, leaving most of the reported gain attributable to dual-encoder capacity rather than disentanglement. Capacity-matched ablations and controlled functional probes, not CKA alone, are required to substantiate disentanglement claims; orthogonality penalties otherwise manufacture the very metric used to certify them.
%U https://aclanthology.org/2026.insights-1.1/
%P 1-6
Markdown (Informal)
[Geometric Separation Without Functional Specialization: Auditing Subspace Disentanglement in Dual-Encoder LLM Paraphrase Detection](https://aclanthology.org/2026.insights-1.1/) (Arun et al., insights 2026)
ACL