@inproceedings{du-etal-2026-paraphrases,
title = "When Paraphrases Break Binding: A Lexical Cliff in Small Heads over Frozen {CLIP} Text Embeddings",
author = "Du, Xinmiao and
Jiang, Zhuoyu and
Wu, Xihong",
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.2262/",
pages = "51437--51453",
ISBN = "979-8-89176-459-0",
abstract = "Small learned heads over frozen CLIP text embeddings are routinely evaluated on held-out examples whose lexical cue variants still come from the training pool, leaving lexical generalization untested. We study this gap on a controlled compositional preference task and identify a lexical cliff: a per-attribute cross-attention head over frozen ViT-B/32 features loses accuracy on held-out cue variants even when they sit inside a narrow CLIP-text neighbourhood (cosine [0.87, 0.97] to training) and the visual input is unchanged. A substitution probe that swaps only the text embedding lifts the worst-failing token from 0.500 to 0.893, and a 2{\texttimes}2 encoder factorial shows that cliff closure tracks the text encoder rather than the visual front-end. The cliff has two distinguishable mechanisms{---}a cosine-correlated decline on motion verbs (r=0.627, n=12) and a class-bimodal pattern on size and speed{---}and replicates across OpenCLIP-LAION-2B and SigLIP-WebLI with encoder-family-dependent token identity. Scaling to ViT-L/14 closes most but not all of the cliff, indicating the threshold is set by text-encoder geometry rather than raw cosine. We release CF-PrefBench v4 and recommend lexical-held-out splits with cosine stratification for evaluating small learned heads over frozen CLIP text."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="du-etal-2026-paraphrases">
<titleInfo>
<title>When Paraphrases Break Binding: A Lexical Cliff in Small Heads over Frozen CLIP Text Embeddings</title>
</titleInfo>
<name type="personal">
<namePart type="given">Xinmiao</namePart>
<namePart type="family">Du</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhuoyu</namePart>
<namePart type="family">Jiang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xihong</namePart>
<namePart type="family">Wu</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>Small learned heads over frozen CLIP text embeddings are routinely evaluated on held-out examples whose lexical cue variants still come from the training pool, leaving lexical generalization untested. We study this gap on a controlled compositional preference task and identify a lexical cliff: a per-attribute cross-attention head over frozen ViT-B/32 features loses accuracy on held-out cue variants even when they sit inside a narrow CLIP-text neighbourhood (cosine [0.87, 0.97] to training) and the visual input is unchanged. A substitution probe that swaps only the text embedding lifts the worst-failing token from 0.500 to 0.893, and a 2×2 encoder factorial shows that cliff closure tracks the text encoder rather than the visual front-end. The cliff has two distinguishable mechanisms—a cosine-correlated decline on motion verbs (r=0.627, n=12) and a class-bimodal pattern on size and speed—and replicates across OpenCLIP-LAION-2B and SigLIP-WebLI with encoder-family-dependent token identity. Scaling to ViT-L/14 closes most but not all of the cliff, indicating the threshold is set by text-encoder geometry rather than raw cosine. We release CF-PrefBench v4 and recommend lexical-held-out splits with cosine stratification for evaluating small learned heads over frozen CLIP text.</abstract>
<identifier type="citekey">du-etal-2026-paraphrases</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2262/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>51437</start>
<end>51453</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When Paraphrases Break Binding: A Lexical Cliff in Small Heads over Frozen CLIP Text Embeddings
%A Du, Xinmiao
%A Jiang, Zhuoyu
%A Wu, Xihong
%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 du-etal-2026-paraphrases
%X Small learned heads over frozen CLIP text embeddings are routinely evaluated on held-out examples whose lexical cue variants still come from the training pool, leaving lexical generalization untested. We study this gap on a controlled compositional preference task and identify a lexical cliff: a per-attribute cross-attention head over frozen ViT-B/32 features loses accuracy on held-out cue variants even when they sit inside a narrow CLIP-text neighbourhood (cosine [0.87, 0.97] to training) and the visual input is unchanged. A substitution probe that swaps only the text embedding lifts the worst-failing token from 0.500 to 0.893, and a 2×2 encoder factorial shows that cliff closure tracks the text encoder rather than the visual front-end. The cliff has two distinguishable mechanisms—a cosine-correlated decline on motion verbs (r=0.627, n=12) and a class-bimodal pattern on size and speed—and replicates across OpenCLIP-LAION-2B and SigLIP-WebLI with encoder-family-dependent token identity. Scaling to ViT-L/14 closes most but not all of the cliff, indicating the threshold is set by text-encoder geometry rather than raw cosine. We release CF-PrefBench v4 and recommend lexical-held-out splits with cosine stratification for evaluating small learned heads over frozen CLIP text.
%U https://aclanthology.org/2026.emnlp-main.2262/
%P 51437-51453
Markdown (Informal)
[When Paraphrases Break Binding: A Lexical Cliff in Small Heads over Frozen CLIP Text Embeddings](https://aclanthology.org/2026.emnlp-main.2262/) (Du et al., EMNLP 2026)
ACL