@inproceedings{foroutan-meshram-2026-adapting,
title = "Adapting {SC}o{R}e to Single-Player Dialogue Games: Initial Experiments on {PLAYPEN}",
author = "Foroutan, Neda and
Meshram, Valhari",
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-1.11/",
pages = "132--139",
ISBN = "979-8-89176-467-5",
abstract = "Large Language Models (LLMs) are increasingly deployed as autonomous agents that solve tasks through multi-step interactions. However, most post-training methods are designed for single-turn reasoning or instruction-following tasks and provide limited supervision for long-horizon conversational decision making. In this work, we investigate whether Self-Correction via Reinforcement Learning (SCoRe), originally proposed for single-step mathematical reasoning and code generation, can improve agentic performance in multi-turn dialogue games. We adapt SCoRe to the PLAYPEN benchmark by fine-tuning Qwen3.5-9B on five single-player dialogue games using interaction-level rewards and iterative self-correction. Experimental results show that the resulting model substantially improves performance on the trained single-player games and further generalizes to unseen multi-player dialogue games, increasing the overall ClemScore from 41.18 to 43.31. Although a slight decrease is observed on the aggregate StatScore, the model largely retains its general language understanding and reasoning capabilities. These findings demonstrate that self-correction reinforcement learning is an effective approach for improving long-horizon conversational agent behavior beyond single-step reasoning tasks."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="foroutan-meshram-2026-adapting">
<titleInfo>
<title>Adapting SCoRe to Single-Player Dialogue Games: Initial Experiments on PLAYPEN</title>
</titleInfo>
<name type="personal">
<namePart type="given">Neda</namePart>
<namePart type="family">Foroutan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Valhari</namePart>
<namePart type="family">Meshram</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>Large Language Models (LLMs) are increasingly deployed as autonomous agents that solve tasks through multi-step interactions. However, most post-training methods are designed for single-turn reasoning or instruction-following tasks and provide limited supervision for long-horizon conversational decision making. In this work, we investigate whether Self-Correction via Reinforcement Learning (SCoRe), originally proposed for single-step mathematical reasoning and code generation, can improve agentic performance in multi-turn dialogue games. We adapt SCoRe to the PLAYPEN benchmark by fine-tuning Qwen3.5-9B on five single-player dialogue games using interaction-level rewards and iterative self-correction. Experimental results show that the resulting model substantially improves performance on the trained single-player games and further generalizes to unseen multi-player dialogue games, increasing the overall ClemScore from 41.18 to 43.31. Although a slight decrease is observed on the aggregate StatScore, the model largely retains its general language understanding and reasoning capabilities. These findings demonstrate that self-correction reinforcement learning is an effective approach for improving long-horizon conversational agent behavior beyond single-step reasoning tasks.</abstract>
<identifier type="citekey">foroutan-meshram-2026-adapting</identifier>
<location>
<url>https://aclanthology.org/2026.lmplayschool-1.11/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>132</start>
<end>139</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Adapting SCoRe to Single-Player Dialogue Games: Initial Experiments on PLAYPEN
%A Foroutan, Neda
%A Meshram, Valhari
%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 foroutan-meshram-2026-adapting
%X Large Language Models (LLMs) are increasingly deployed as autonomous agents that solve tasks through multi-step interactions. However, most post-training methods are designed for single-turn reasoning or instruction-following tasks and provide limited supervision for long-horizon conversational decision making. In this work, we investigate whether Self-Correction via Reinforcement Learning (SCoRe), originally proposed for single-step mathematical reasoning and code generation, can improve agentic performance in multi-turn dialogue games. We adapt SCoRe to the PLAYPEN benchmark by fine-tuning Qwen3.5-9B on five single-player dialogue games using interaction-level rewards and iterative self-correction. Experimental results show that the resulting model substantially improves performance on the trained single-player games and further generalizes to unseen multi-player dialogue games, increasing the overall ClemScore from 41.18 to 43.31. Although a slight decrease is observed on the aggregate StatScore, the model largely retains its general language understanding and reasoning capabilities. These findings demonstrate that self-correction reinforcement learning is an effective approach for improving long-horizon conversational agent behavior beyond single-step reasoning tasks.
%U https://aclanthology.org/2026.lmplayschool-1.11/
%P 132-139
Markdown (Informal)
[Adapting SCoRe to Single-Player Dialogue Games: Initial Experiments on PLAYPEN](https://aclanthology.org/2026.lmplayschool-1.11/) (Foroutan & Meshram, LM Playschool 2026)
ACL