@article{huang-etal-2026-haring,
title = "{S} {HARING} {B} {EYOND} {D} {ECISION}: Deep Collaboration between Large Language Models via Representation Ensemble",
author = "Huang, Yichong and
Feng, Xiaocheng and
Fu, Jinlan and
Feng, Xiachong and
Li, Baohang and
Ye, Zekai and
Qin, Libo and
Fei, Hao and
Ng, See-Kiong and
Qin, Bing",
journal = "Transactions of the Association for Computational Linguistics",
volume = "14",
year = "2026",
address = "Cambridge, MA",
publisher = "MIT Press",
url = "https://aclanthology.org/2026.tacl-1.79/",
doi = "10.1162/tacl.a.773",
pages = "1765--1786",
abstract = "Large Language Models (LLMs) exhibit unique strengths arising from differences in model architecture, training data, and strategies. Ensemble learning has been explored to leverage these complementary strengths through decision-level sharing (i.e.,Decision Ensemble), which combines the predictions from multiple LLMs. However, such methods integrate only shallow decisions and overlook the exchange of deeper levels of information within the internal representations of LLMs, such as problem understanding, world knowledge, and latent reasoning patterns. In this work, we propose Representation Ensemble (RISE), a novel ensemble framework that enables cross-LLM representation sharing for richer information exchange. To address challenges of representation-level interaction caused by layer misalignment and latent-space incompatibility across LLMs, we introduce a representation alignment method based on relational similarity measures and an orthogonal latent-space transformation. Experimental results show that (1) RISE achieves performance competitive with existing decision ensemble methods, and (2) RISE is strongly complementary to decision ensemble, with their combination boosting collaboration gains by 14{\%}{--}41{\%}. Finally, we further compare ensemble of small LLMs to a single larger LLM and to model merging and composition approaches, and find that ensemble learning consistently generalizes well without additional training."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="huang-etal-2026-haring">
<titleInfo>
<title>S HARING B EYOND D ECISION: Deep Collaboration between Large Language Models via Representation Ensemble</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yichong</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaocheng</namePart>
<namePart type="family">Feng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jinlan</namePart>
<namePart type="family">Fu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiachong</namePart>
<namePart type="family">Feng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Baohang</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zekai</namePart>
<namePart type="family">Ye</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Libo</namePart>
<namePart type="family">Qin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hao</namePart>
<namePart type="family">Fei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">See-Kiong</namePart>
<namePart type="family">Ng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bing</namePart>
<namePart type="family">Qin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<genre authority="bibutilsgt">journal article</genre>
<relatedItem type="host">
<titleInfo>
<title>Transactions of the Association for Computational Linguistics</title>
</titleInfo>
<originInfo>
<issuance>continuing</issuance>
<publisher>MIT Press</publisher>
<place>
<placeTerm type="text">Cambridge, MA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">periodical</genre>
<genre authority="bibutilsgt">academic journal</genre>
</relatedItem>
<abstract>Large Language Models (LLMs) exhibit unique strengths arising from differences in model architecture, training data, and strategies. Ensemble learning has been explored to leverage these complementary strengths through decision-level sharing (i.e.,Decision Ensemble), which combines the predictions from multiple LLMs. However, such methods integrate only shallow decisions and overlook the exchange of deeper levels of information within the internal representations of LLMs, such as problem understanding, world knowledge, and latent reasoning patterns. In this work, we propose Representation Ensemble (RISE), a novel ensemble framework that enables cross-LLM representation sharing for richer information exchange. To address challenges of representation-level interaction caused by layer misalignment and latent-space incompatibility across LLMs, we introduce a representation alignment method based on relational similarity measures and an orthogonal latent-space transformation. Experimental results show that (1) RISE achieves performance competitive with existing decision ensemble methods, and (2) RISE is strongly complementary to decision ensemble, with their combination boosting collaboration gains by 14%–41%. Finally, we further compare ensemble of small LLMs to a single larger LLM and to model merging and composition approaches, and find that ensemble learning consistently generalizes well without additional training.</abstract>
<identifier type="citekey">huang-etal-2026-haring</identifier>
<identifier type="doi">10.1162/tacl.a.773</identifier>
<location>
<url>https://aclanthology.org/2026.tacl-1.79/</url>
</location>
<part>
<date>2026</date>
<detail type="volume"><number>14</number></detail>
<extent unit="page">
<start>1765</start>
<end>1786</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Journal Article
%T S HARING B EYOND D ECISION: Deep Collaboration between Large Language Models via Representation Ensemble
%A Huang, Yichong
%A Feng, Xiaocheng
%A Fu, Jinlan
%A Feng, Xiachong
%A Li, Baohang
%A Ye, Zekai
%A Qin, Libo
%A Fei, Hao
%A Ng, See-Kiong
%A Qin, Bing
%J Transactions of the Association for Computational Linguistics
%D 2026
%V 14
%I MIT Press
%C Cambridge, MA
%F huang-etal-2026-haring
%X Large Language Models (LLMs) exhibit unique strengths arising from differences in model architecture, training data, and strategies. Ensemble learning has been explored to leverage these complementary strengths through decision-level sharing (i.e.,Decision Ensemble), which combines the predictions from multiple LLMs. However, such methods integrate only shallow decisions and overlook the exchange of deeper levels of information within the internal representations of LLMs, such as problem understanding, world knowledge, and latent reasoning patterns. In this work, we propose Representation Ensemble (RISE), a novel ensemble framework that enables cross-LLM representation sharing for richer information exchange. To address challenges of representation-level interaction caused by layer misalignment and latent-space incompatibility across LLMs, we introduce a representation alignment method based on relational similarity measures and an orthogonal latent-space transformation. Experimental results show that (1) RISE achieves performance competitive with existing decision ensemble methods, and (2) RISE is strongly complementary to decision ensemble, with their combination boosting collaboration gains by 14%–41%. Finally, we further compare ensemble of small LLMs to a single larger LLM and to model merging and composition approaches, and find that ensemble learning consistently generalizes well without additional training.
%R 10.1162/tacl.a.773
%U https://aclanthology.org/2026.tacl-1.79/
%U https://doi.org/10.1162/tacl.a.773
%P 1765-1786
Markdown (Informal)
[S HARING B EYOND D ECISION: Deep Collaboration between Large Language Models via Representation Ensemble](https://aclanthology.org/2026.tacl-1.79/) (Huang et al., TACL 2026)
ACL
- Yichong Huang, Xiaocheng Feng, Jinlan Fu, Xiachong Feng, Baohang Li, Zekai Ye, Libo Qin, Hao Fei, See-Kiong Ng, and Bing Qin. 2026. S HARING B EYOND D ECISION: Deep Collaboration between Large Language Models via Representation Ensemble. Transactions of the Association for Computational Linguistics, 14:1765–1786.