@inproceedings{salhan-buttery-2026-protocol,
title = "Protocol Competence of Small Language Models in Interactive Language Game Environments",
author = "Salhan, Suchir and
Buttery, Paula",
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-1.6/",
pages = "57--79",
ISBN = "979-8-89176-467-5",
abstract = "Learning from interaction requires more than producing appropriate responses: learners must maintain a shared interaction with a partner, following its rules while adapting to feedback and contingencies. We argue that this depends on a prerequisite we call **protocol competence**: the ability to produce valid actions, maintain interaction structure, and respond appropriately across turns. In this paper, we investigate protocol competence in Small Language Models (SLMs) from the Qwen family across five interactive dialogue games from clembench, finding 51.3{\%} of episodes from Qwen3.5-0.8B terminate in protocol violations. However, teacher demonstrations from a larger language model can scaffold participation, raising Wordle play from 0/30 to 29/30. Once models can reliably participate, we find self-improving reinforcement learning becomes effective: a single round of self-generated reinforcement learning (self-ReST) yields large gains from a small number of successful trajectories. Comparing adaptive learning strategies, we find that credit-weighted training provides the most reliable improvements, reducing variance relative to uniform supervision. Together, these results suggest that interactive learning follows a staged trajectory: SLMs must first acquire the competence to participate before they can exploit the contingencies generated by interaction."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="salhan-buttery-2026-protocol">
<titleInfo>
<title>Protocol Competence of Small Language Models in Interactive Language Game Environments</title>
</titleInfo>
<name type="personal">
<namePart type="given">Suchir</namePart>
<namePart type="family">Salhan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Paula</namePart>
<namePart type="family">Buttery</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>Learning from interaction requires more than producing appropriate responses: learners must maintain a shared interaction with a partner, following its rules while adapting to feedback and contingencies. We argue that this depends on a prerequisite we call **protocol competence**: the ability to produce valid actions, maintain interaction structure, and respond appropriately across turns. In this paper, we investigate protocol competence in Small Language Models (SLMs) from the Qwen family across five interactive dialogue games from clembench, finding 51.3% of episodes from Qwen3.5-0.8B terminate in protocol violations. However, teacher demonstrations from a larger language model can scaffold participation, raising Wordle play from 0/30 to 29/30. Once models can reliably participate, we find self-improving reinforcement learning becomes effective: a single round of self-generated reinforcement learning (self-ReST) yields large gains from a small number of successful trajectories. Comparing adaptive learning strategies, we find that credit-weighted training provides the most reliable improvements, reducing variance relative to uniform supervision. Together, these results suggest that interactive learning follows a staged trajectory: SLMs must first acquire the competence to participate before they can exploit the contingencies generated by interaction.</abstract>
<identifier type="citekey">salhan-buttery-2026-protocol</identifier>
<location>
<url>https://aclanthology.org/2026.lmplayschool-1.6/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>57</start>
<end>79</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Protocol Competence of Small Language Models in Interactive Language Game Environments
%A Salhan, Suchir
%A Buttery, Paula
%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 salhan-buttery-2026-protocol
%X Learning from interaction requires more than producing appropriate responses: learners must maintain a shared interaction with a partner, following its rules while adapting to feedback and contingencies. We argue that this depends on a prerequisite we call **protocol competence**: the ability to produce valid actions, maintain interaction structure, and respond appropriately across turns. In this paper, we investigate protocol competence in Small Language Models (SLMs) from the Qwen family across five interactive dialogue games from clembench, finding 51.3% of episodes from Qwen3.5-0.8B terminate in protocol violations. However, teacher demonstrations from a larger language model can scaffold participation, raising Wordle play from 0/30 to 29/30. Once models can reliably participate, we find self-improving reinforcement learning becomes effective: a single round of self-generated reinforcement learning (self-ReST) yields large gains from a small number of successful trajectories. Comparing adaptive learning strategies, we find that credit-weighted training provides the most reliable improvements, reducing variance relative to uniform supervision. Together, these results suggest that interactive learning follows a staged trajectory: SLMs must first acquire the competence to participate before they can exploit the contingencies generated by interaction.
%U https://aclanthology.org/2026.lmplayschool-1.6/
%P 57-79
Markdown (Informal)
[Protocol Competence of Small Language Models in Interactive Language Game Environments](https://aclanthology.org/2026.lmplayschool-1.6/) (Salhan & Buttery, LM Playschool 2026)
ACL