@inproceedings{tej-etal-2026-enhancing,
title = "Enhancing {LLM} Performance via {L}o{RA} {SFT} and {DPO} for the {LM} Playschool Challenge",
author = "Tej, Vasamsetti Nihal and
Harshavaishnav, Velineni and
Ghosh, Shreya",
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-1.10/",
pages = "127--131",
ISBN = "979-8-89176-467-5",
abstract = "Large Language Models (LLMs) have achieved SOTA results in many natural language pro- cessing tasks. However, they are primarily trained for static text generation and often strug- gle in interactive, goal-oriented environments that require multi-turn reasoning, state track- ing, and need to give structured output formats. This paper presents our team{'}s submission to the LM Playschool Workshop Shared Task at EMNLP 2026. The LM Playschool Workshop Shared Task at EMNLP 2026 addresses this challenge by evaluating language models as in- teractive agents in structured dialogue games. We hereby explored different fine-tuning strate- gies for Qwen 4B and Qwen 7B models to improve their interactive capabilities and eval- uated them. We perform a comparison of Su- pervised Fine-Tuning (SFT) and Direct Pref- erence Optimization (DPO) using multi-turn training data. Our study analyzes the effective- ness of each optimization strategy for learning robust dialogue behaviors while maintaining models stability towards interactive games and also not to forget its language linguistics by different evaluating metrics. Experimental re- sults show that although supervised fine-tuning provides consistent interactive performance, di- rectly optimizing the foundation model using DPO reduces optimization errors and leads to improved agent behavior. Our best-performing model achieved a clemscore of 38.23 on the official LM Playschool leaderboard, evaluated using the Playpen workbench."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="tej-etal-2026-enhancing">
<titleInfo>
<title>Enhancing LLM Performance via LoRA SFT and DPO for the LM Playschool Challenge</title>
</titleInfo>
<name type="personal">
<namePart type="given">Vasamsetti</namePart>
<namePart type="given">Nihal</namePart>
<namePart type="family">Tej</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Velineni</namePart>
<namePart type="family">Harshavaishnav</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shreya</namePart>
<namePart type="family">Ghosh</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) have achieved SOTA results in many natural language pro- cessing tasks. However, they are primarily trained for static text generation and often strug- gle in interactive, goal-oriented environments that require multi-turn reasoning, state track- ing, and need to give structured output formats. This paper presents our team’s submission to the LM Playschool Workshop Shared Task at EMNLP 2026. The LM Playschool Workshop Shared Task at EMNLP 2026 addresses this challenge by evaluating language models as in- teractive agents in structured dialogue games. We hereby explored different fine-tuning strate- gies for Qwen 4B and Qwen 7B models to improve their interactive capabilities and eval- uated them. We perform a comparison of Su- pervised Fine-Tuning (SFT) and Direct Pref- erence Optimization (DPO) using multi-turn training data. Our study analyzes the effective- ness of each optimization strategy for learning robust dialogue behaviors while maintaining models stability towards interactive games and also not to forget its language linguistics by different evaluating metrics. Experimental re- sults show that although supervised fine-tuning provides consistent interactive performance, di- rectly optimizing the foundation model using DPO reduces optimization errors and leads to improved agent behavior. Our best-performing model achieved a clemscore of 38.23 on the official LM Playschool leaderboard, evaluated using the Playpen workbench.</abstract>
<identifier type="citekey">tej-etal-2026-enhancing</identifier>
<location>
<url>https://aclanthology.org/2026.lmplayschool-1.10/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>127</start>
<end>131</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Enhancing LLM Performance via LoRA SFT and DPO for the LM Playschool Challenge
%A Tej, Vasamsetti Nihal
%A Harshavaishnav, Velineni
%A Ghosh, Shreya
%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 tej-etal-2026-enhancing
%X Large Language Models (LLMs) have achieved SOTA results in many natural language pro- cessing tasks. However, they are primarily trained for static text generation and often strug- gle in interactive, goal-oriented environments that require multi-turn reasoning, state track- ing, and need to give structured output formats. This paper presents our team’s submission to the LM Playschool Workshop Shared Task at EMNLP 2026. The LM Playschool Workshop Shared Task at EMNLP 2026 addresses this challenge by evaluating language models as in- teractive agents in structured dialogue games. We hereby explored different fine-tuning strate- gies for Qwen 4B and Qwen 7B models to improve their interactive capabilities and eval- uated them. We perform a comparison of Su- pervised Fine-Tuning (SFT) and Direct Pref- erence Optimization (DPO) using multi-turn training data. Our study analyzes the effective- ness of each optimization strategy for learning robust dialogue behaviors while maintaining models stability towards interactive games and also not to forget its language linguistics by different evaluating metrics. Experimental re- sults show that although supervised fine-tuning provides consistent interactive performance, di- rectly optimizing the foundation model using DPO reduces optimization errors and leads to improved agent behavior. Our best-performing model achieved a clemscore of 38.23 on the official LM Playschool leaderboard, evaluated using the Playpen workbench.
%U https://aclanthology.org/2026.lmplayschool-1.10/
%P 127-131
Markdown (Informal)
[Enhancing LLM Performance via LoRA SFT and DPO for the LM Playschool Challenge](https://aclanthology.org/2026.lmplayschool-1.10/) (Tej et al., LM Playschool 2026)
ACL