@inproceedings{naszadi-monz-2026-role,
title = "Role-Specific Strengths and Alignment in {LLM} Cross-Play and Self-Play",
author = "Naszadi, Kata and
Monz, Christof",
editor = "Suglia, Alessandro and
Giulianelli, Mario and
Schlangen, David and
Bernardi, Raffaella and
Fern{\'a}ndez, Raquel and
Hakimov, Sherzod and
McCallum, Sabrina and
Mazzaccara, Davide and
Le, Dieu-Thu and
Lemon, Oliver and
Koller, Alexander",
booktitle = "The Proceedings of Improving Language Models through Learning from Dialogue Game Interaction",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.lmplayschool-1.4/",
pages = "41--48",
ISBN = "979-8-89176-467-5",
abstract = "Dialogue-game benchmarks usually pair a model with itself, and the resulting self-play score is treated as a measure of that model{'}s ability. We test whether self-play scores measure partner-independent ability by comparing them with cross-play evaluations. Across nine Clembench games, we evaluate every possible seating of three popular open-weight models. The cross-play reading exposes additional model strengths and weaknesses that are not detectable in self-play. In four of nine games, a mixed pair of two different models beats the best self-play score; in another four cases, a model does better paired with itself than paired with the strongest available partner. Looking at the transcripts, two mechanisms explain most of this: some pairs succeed because one partner writes more detailed messages that keep the interaction on track, while some self-play scores are inflated because both copies of a model make the same mistake and agree by accident."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="naszadi-monz-2026-role">
<titleInfo>
<title>Role-Specific Strengths and Alignment in LLM Cross-Play and Self-Play</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kata</namePart>
<namePart type="family">Naszadi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christof</namePart>
<namePart type="family">Monz</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>The Proceedings of Improving Language Models through Learning from Dialogue Game Interaction</title>
</titleInfo>
<name type="personal">
<namePart type="given">Alessandro</namePart>
<namePart type="family">Suglia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mario</namePart>
<namePart type="family">Giulianelli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">David</namePart>
<namePart type="family">Schlangen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Raffaella</namePart>
<namePart type="family">Bernardi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Raquel</namePart>
<namePart type="family">Fernández</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sherzod</namePart>
<namePart type="family">Hakimov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sabrina</namePart>
<namePart type="family">McCallum</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Davide</namePart>
<namePart type="family">Mazzaccara</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dieu-Thu</namePart>
<namePart type="family">Le</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Oliver</namePart>
<namePart type="family">Lemon</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Koller</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">979-8-89176-467-5</identifier>
</relatedItem>
<abstract>Dialogue-game benchmarks usually pair a model with itself, and the resulting self-play score is treated as a measure of that model’s ability. We test whether self-play scores measure partner-independent ability by comparing them with cross-play evaluations. Across nine Clembench games, we evaluate every possible seating of three popular open-weight models. The cross-play reading exposes additional model strengths and weaknesses that are not detectable in self-play. In four of nine games, a mixed pair of two different models beats the best self-play score; in another four cases, a model does better paired with itself than paired with the strongest available partner. Looking at the transcripts, two mechanisms explain most of this: some pairs succeed because one partner writes more detailed messages that keep the interaction on track, while some self-play scores are inflated because both copies of a model make the same mistake and agree by accident.</abstract>
<identifier type="citekey">naszadi-monz-2026-role</identifier>
<location>
<url>https://aclanthology.org/2026.lmplayschool-1.4/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>41</start>
<end>48</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Role-Specific Strengths and Alignment in LLM Cross-Play and Self-Play
%A Naszadi, Kata
%A Monz, Christof
%Y Suglia, Alessandro
%Y Giulianelli, Mario
%Y Schlangen, David
%Y Bernardi, Raffaella
%Y Fernández, Raquel
%Y Hakimov, Sherzod
%Y McCallum, Sabrina
%Y Mazzaccara, Davide
%Y Le, Dieu-Thu
%Y Lemon, Oliver
%Y Koller, Alexander
%S The Proceedings of Improving Language Models through Learning from Dialogue Game Interaction
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-467-5
%F naszadi-monz-2026-role
%X Dialogue-game benchmarks usually pair a model with itself, and the resulting self-play score is treated as a measure of that model’s ability. We test whether self-play scores measure partner-independent ability by comparing them with cross-play evaluations. Across nine Clembench games, we evaluate every possible seating of three popular open-weight models. The cross-play reading exposes additional model strengths and weaknesses that are not detectable in self-play. In four of nine games, a mixed pair of two different models beats the best self-play score; in another four cases, a model does better paired with itself than paired with the strongest available partner. Looking at the transcripts, two mechanisms explain most of this: some pairs succeed because one partner writes more detailed messages that keep the interaction on track, while some self-play scores are inflated because both copies of a model make the same mistake and agree by accident.
%U https://aclanthology.org/2026.lmplayschool-1.4/
%P 41-48
Markdown (Informal)
[Role-Specific Strengths and Alignment in LLM Cross-Play and Self-Play](https://aclanthology.org/2026.lmplayschool-1.4/) (Naszadi & Monz, LM Playschool 2026)
ACL