@inproceedings{yu-etal-2026-composition,
title = "Composition Collapse: Stable Factual Knowledge Does Not Imply Compositional Reasoning",
author = "Yu, Zhe and
Xing, Wenpeng and
Wei, Yunzhao and
Chen, Jie and
Wang, Hongzhi and
Teng, Xuyang and
Han, Meng",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1411/",
pages = "30070--30092",
ISBN = "979-8-89176-462-0",
abstract = "A model can know when two events occurred yet still answer which came first incorrectly. Multi-hop QA accuracy conflates such composition errors with missing facts. We introduce a double-gate protocol for temporal factual chains and evaluate a case for composition only if the model (i) answers every constituent fact correctly across paraphrased probes and (ii) answers every aligned sub-question correctly in isolation. If the original multi-hop answer remains incorrect, the case is a residual composition failure; we call the systematic occurrence of such failures composition collapse. On D4v2, a benchmark of 390 temporal chains, two post-trained models differ by only 0.3 percentage points in atomic stability yet by 47.4 points in depth-2 (two-fact) residual failure. In a controlled same-base comparison, outcome-verified reinforcement learning reduces depth-4 residual failure by 54 points relative to reasoning-trace SFT under matched data and LoRA budget. On 80 examples that pass both gates for every GRPO variant, failure decreases monotonically, showing that model-specific gate subsets are unlikely to drive the trend. Finally, chain-of-thought decoding recovers 70{--}75{\%} of gate-passing failures, consistent with an important role for inference-time computation. These results motivate reporting atomic stability and gate-controlled composition separately when evaluating multi-hop reasoning."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yu-etal-2026-composition">
<titleInfo>
<title>Composition Collapse: Stable Factual Knowledge Does Not Imply Compositional Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhe</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenpeng</namePart>
<namePart type="family">Xing</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yunzhao</namePart>
<namePart type="family">Wei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jie</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hongzhi</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuyang</namePart>
<namePart type="family">Teng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Meng</namePart>
<namePart type="family">Han</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>A model can know when two events occurred yet still answer which came first incorrectly. Multi-hop QA accuracy conflates such composition errors with missing facts. We introduce a double-gate protocol for temporal factual chains and evaluate a case for composition only if the model (i) answers every constituent fact correctly across paraphrased probes and (ii) answers every aligned sub-question correctly in isolation. If the original multi-hop answer remains incorrect, the case is a residual composition failure; we call the systematic occurrence of such failures composition collapse. On D4v2, a benchmark of 390 temporal chains, two post-trained models differ by only 0.3 percentage points in atomic stability yet by 47.4 points in depth-2 (two-fact) residual failure. In a controlled same-base comparison, outcome-verified reinforcement learning reduces depth-4 residual failure by 54 points relative to reasoning-trace SFT under matched data and LoRA budget. On 80 examples that pass both gates for every GRPO variant, failure decreases monotonically, showing that model-specific gate subsets are unlikely to drive the trend. Finally, chain-of-thought decoding recovers 70–75% of gate-passing failures, consistent with an important role for inference-time computation. These results motivate reporting atomic stability and gate-controlled composition separately when evaluating multi-hop reasoning.</abstract>
<identifier type="citekey">yu-etal-2026-composition</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1411/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>30070</start>
<end>30092</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Composition Collapse: Stable Factual Knowledge Does Not Imply Compositional Reasoning
%A Yu, Zhe
%A Xing, Wenpeng
%A Wei, Yunzhao
%A Chen, Jie
%A Wang, Hongzhi
%A Teng, Xuyang
%A Han, Meng
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F yu-etal-2026-composition
%X A model can know when two events occurred yet still answer which came first incorrectly. Multi-hop QA accuracy conflates such composition errors with missing facts. We introduce a double-gate protocol for temporal factual chains and evaluate a case for composition only if the model (i) answers every constituent fact correctly across paraphrased probes and (ii) answers every aligned sub-question correctly in isolation. If the original multi-hop answer remains incorrect, the case is a residual composition failure; we call the systematic occurrence of such failures composition collapse. On D4v2, a benchmark of 390 temporal chains, two post-trained models differ by only 0.3 percentage points in atomic stability yet by 47.4 points in depth-2 (two-fact) residual failure. In a controlled same-base comparison, outcome-verified reinforcement learning reduces depth-4 residual failure by 54 points relative to reasoning-trace SFT under matched data and LoRA budget. On 80 examples that pass both gates for every GRPO variant, failure decreases monotonically, showing that model-specific gate subsets are unlikely to drive the trend. Finally, chain-of-thought decoding recovers 70–75% of gate-passing failures, consistent with an important role for inference-time computation. These results motivate reporting atomic stability and gate-controlled composition separately when evaluating multi-hop reasoning.
%U https://aclanthology.org/2026.findings-emnlp.1411/
%P 30070-30092
Markdown (Informal)
[Composition Collapse: Stable Factual Knowledge Does Not Imply Compositional Reasoning](https://aclanthology.org/2026.findings-emnlp.1411/) (Yu et al., Findings 2026)
ACL