@inproceedings{srivastava-2026-raising,
title = "Raising a Small Language Model: From Imitation to Curiosity-Driven Learning in Dialogue Games",
author = "Srivastava, Varad",
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-1.13/",
pages = "148--159",
ISBN = "979-8-89176-467-5",
abstract = "Children do not learn language from text corpora, rather they learn it in interaction by imitating competent speakers, receiving feedback, practicing, being corrected, and playing. The LM Playschool Challenge asks whether language models can profit from analogous regimes. We take the smallest recommended base model, Qwen3.5-2B (clemscore 13.63 in our evaluation environment), and traverse a sequence of five interaction-derived learning regimes that mirror a developmental progression: imitation of competent others (supervised fine-tuning on success-filtered transcripts), learning from outcome contrasts (direct preference optimisation), self-imitation (fine-tuning on the model{'}s own successes), corrective feedback (contrastive pairs from the model{'}s own failures against its own and expert successes), and intrinsically motivated trial-and-error (online GRPO with a novelty bonus). The offline regimes carry the model to 67.6 clemscore on our validation evaluation ($+54.0$ over the base model). On the organisers' held-out evaluation, our two submitted checkpoints achieve the largest out-of-domain gains of any submission in the challenge ($+11.9$ and $+10.8$ clemscore, against a maximum of $+6.5$ for any other entry), with in-domain gains of $+33.2$ and $+37.3$ and static-benchmark performance unchanged. Online GRPO \textit{regresses} from its initialisation ($-5.0$) unless augmented with a random-network-distillation novelty bonus, which fully recovers the loss ($+5.0$ over the plain-GRPO control), evidence that the bonus primarily counteracts the forgetting that narrow online training induces on untrained games."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="srivastava-2026-raising">
<titleInfo>
<title>Raising a Small Language Model: From Imitation to Curiosity-Driven Learning in Dialogue Games</title>
</titleInfo>
<name type="personal">
<namePart type="given">Varad</namePart>
<namePart type="family">Srivastava</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>Children do not learn language from text corpora, rather they learn it in interaction by imitating competent speakers, receiving feedback, practicing, being corrected, and playing. The LM Playschool Challenge asks whether language models can profit from analogous regimes. We take the smallest recommended base model, Qwen3.5-2B (clemscore 13.63 in our evaluation environment), and traverse a sequence of five interaction-derived learning regimes that mirror a developmental progression: imitation of competent others (supervised fine-tuning on success-filtered transcripts), learning from outcome contrasts (direct preference optimisation), self-imitation (fine-tuning on the model’s own successes), corrective feedback (contrastive pairs from the model’s own failures against its own and expert successes), and intrinsically motivated trial-and-error (online GRPO with a novelty bonus). The offline regimes carry the model to 67.6 clemscore on our validation evaluation (+54.0 over the base model). On the organisers’ held-out evaluation, our two submitted checkpoints achieve the largest out-of-domain gains of any submission in the challenge (+11.9 and +10.8 clemscore, against a maximum of +6.5 for any other entry), with in-domain gains of +33.2 and +37.3 and static-benchmark performance unchanged. Online GRPO regresses from its initialisation (-5.0) unless augmented with a random-network-distillation novelty bonus, which fully recovers the loss (+5.0 over the plain-GRPO control), evidence that the bonus primarily counteracts the forgetting that narrow online training induces on untrained games.</abstract>
<identifier type="citekey">srivastava-2026-raising</identifier>
<location>
<url>https://aclanthology.org/2026.lmplayschool-1.13/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>148</start>
<end>159</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Raising a Small Language Model: From Imitation to Curiosity-Driven Learning in Dialogue Games
%A Srivastava, Varad
%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 srivastava-2026-raising
%X Children do not learn language from text corpora, rather they learn it in interaction by imitating competent speakers, receiving feedback, practicing, being corrected, and playing. The LM Playschool Challenge asks whether language models can profit from analogous regimes. We take the smallest recommended base model, Qwen3.5-2B (clemscore 13.63 in our evaluation environment), and traverse a sequence of five interaction-derived learning regimes that mirror a developmental progression: imitation of competent others (supervised fine-tuning on success-filtered transcripts), learning from outcome contrasts (direct preference optimisation), self-imitation (fine-tuning on the model’s own successes), corrective feedback (contrastive pairs from the model’s own failures against its own and expert successes), and intrinsically motivated trial-and-error (online GRPO with a novelty bonus). The offline regimes carry the model to 67.6 clemscore on our validation evaluation (+54.0 over the base model). On the organisers’ held-out evaluation, our two submitted checkpoints achieve the largest out-of-domain gains of any submission in the challenge (+11.9 and +10.8 clemscore, against a maximum of +6.5 for any other entry), with in-domain gains of +33.2 and +37.3 and static-benchmark performance unchanged. Online GRPO regresses from its initialisation (-5.0) unless augmented with a random-network-distillation novelty bonus, which fully recovers the loss (+5.0 over the plain-GRPO control), evidence that the bonus primarily counteracts the forgetting that narrow online training induces on untrained games.
%U https://aclanthology.org/2026.lmplayschool-1.13/
%P 148-159
Markdown (Informal)
[Raising a Small Language Model: From Imitation to Curiosity-Driven Learning in Dialogue Games](https://aclanthology.org/2026.lmplayschool-1.13/) (Srivastava, LM Playschool 2026)
ACL