@inproceedings{islam-etal-2026-learning,
title = "Learning from Correction: Data-Efficient Language Model Training Through Adaptive Teacher-Model Feedback",
author = "Islam, Sadri and
Kosireddy, Tagore Rao and
Li, Kevin and
Gaines, Dylan",
editor = "Choshen, Leshem and
Cotterell, Ryan and
Gul, Mustafa Omer and
Jumelet, Jaap and
Linzen, Tal and
Mueller, Aaron and
Salhan, Suchir and
Shah, Raj Sanjay and
Warstadt, Alex and
Wilcox, Ethan Gotlieb",
booktitle = "Proceedings of the Second {B}aby{LM} Workshop",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.babylm-main.55/",
pages = "689--701",
ISBN = "979-8-89176-475-0",
abstract = "We present our submission to the Strict-Small track of the 2026 BabyLM Challenge. We study whether a developmentally-inspired curriculum, particularly an adaptive teacher-feedback curriculum based on the student{'}s own productions, can improve language model pre-training under a 10M word budget. We trained a 58M-parameter student from scratch under five separate conditions that shared the same architecture and tokenizer, isolating different aspects of the training data and curriculum structure. We built an official baseline trained on a cleaned version of the official Strict-Small corpus and also a synthetic baseline trained on teacher model generated data. Then, we developed a fixed curriculum to train a model on the same synthetic child-directed text in three developmental stages: infant, toddler, and child. Finally, we introduced an adaptive feedback curriculum in which the teachers generate additional training data by correcting the student{'}s productions during training, which adapts the training data according to the learner{'}s level of understanding. The fixed curriculum demonstrated superior performance on the BLiMP-Supplement, EWoK, and COMPS benchmarks, while the adaptive feedback curriculum substantially improved the Entity Tracking performance in one configuration. However, our official baseline model achieved the highest BLiMP, SuperGLUE, and overall macro-average scores, indicating that our curriculum methods did not produce the overall improvements that we had hypothesized. Our code is available at https://github.com/dcgaines/babylm-2026."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="islam-etal-2026-learning">
<titleInfo>
<title>Learning from Correction: Data-Efficient Language Model Training Through Adaptive Teacher-Model Feedback</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sadri</namePart>
<namePart type="family">Islam</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tagore</namePart>
<namePart type="given">Rao</namePart>
<namePart type="family">Kosireddy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kevin</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dylan</namePart>
<namePart type="family">Gaines</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>Proceedings of the Second BabyLM Workshop</title>
</titleInfo>
<name type="personal">
<namePart type="given">Leshem</namePart>
<namePart type="family">Choshen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ryan</namePart>
<namePart type="family">Cotterell</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mustafa</namePart>
<namePart type="given">Omer</namePart>
<namePart type="family">Gul</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jaap</namePart>
<namePart type="family">Jumelet</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tal</namePart>
<namePart type="family">Linzen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aaron</namePart>
<namePart type="family">Mueller</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Suchir</namePart>
<namePart type="family">Salhan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Raj</namePart>
<namePart type="given">Sanjay</namePart>
<namePart type="family">Shah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alex</namePart>
<namePart type="family">Warstadt</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ethan</namePart>
<namePart type="given">Gotlieb</namePart>
<namePart type="family">Wilcox</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">979-8-89176-475-0</identifier>
</relatedItem>
<abstract>We present our submission to the Strict-Small track of the 2026 BabyLM Challenge. We study whether a developmentally-inspired curriculum, particularly an adaptive teacher-feedback curriculum based on the student’s own productions, can improve language model pre-training under a 10M word budget. We trained a 58M-parameter student from scratch under five separate conditions that shared the same architecture and tokenizer, isolating different aspects of the training data and curriculum structure. We built an official baseline trained on a cleaned version of the official Strict-Small corpus and also a synthetic baseline trained on teacher model generated data. Then, we developed a fixed curriculum to train a model on the same synthetic child-directed text in three developmental stages: infant, toddler, and child. Finally, we introduced an adaptive feedback curriculum in which the teachers generate additional training data by correcting the student’s productions during training, which adapts the training data according to the learner’s level of understanding. The fixed curriculum demonstrated superior performance on the BLiMP-Supplement, EWoK, and COMPS benchmarks, while the adaptive feedback curriculum substantially improved the Entity Tracking performance in one configuration. However, our official baseline model achieved the highest BLiMP, SuperGLUE, and overall macro-average scores, indicating that our curriculum methods did not produce the overall improvements that we had hypothesized. Our code is available at https://github.com/dcgaines/babylm-2026.</abstract>
<identifier type="citekey">islam-etal-2026-learning</identifier>
<location>
<url>https://aclanthology.org/2026.babylm-main.55/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>689</start>
<end>701</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Learning from Correction: Data-Efficient Language Model Training Through Adaptive Teacher-Model Feedback
%A Islam, Sadri
%A Kosireddy, Tagore Rao
%A Li, Kevin
%A Gaines, Dylan
%Y Choshen, Leshem
%Y Cotterell, Ryan
%Y Gul, Mustafa Omer
%Y Jumelet, Jaap
%Y Linzen, Tal
%Y Mueller, Aaron
%Y Salhan, Suchir
%Y Shah, Raj Sanjay
%Y Warstadt, Alex
%Y Wilcox, Ethan Gotlieb
%S Proceedings of the Second BabyLM Workshop
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-475-0
%F islam-etal-2026-learning
%X We present our submission to the Strict-Small track of the 2026 BabyLM Challenge. We study whether a developmentally-inspired curriculum, particularly an adaptive teacher-feedback curriculum based on the student’s own productions, can improve language model pre-training under a 10M word budget. We trained a 58M-parameter student from scratch under five separate conditions that shared the same architecture and tokenizer, isolating different aspects of the training data and curriculum structure. We built an official baseline trained on a cleaned version of the official Strict-Small corpus and also a synthetic baseline trained on teacher model generated data. Then, we developed a fixed curriculum to train a model on the same synthetic child-directed text in three developmental stages: infant, toddler, and child. Finally, we introduced an adaptive feedback curriculum in which the teachers generate additional training data by correcting the student’s productions during training, which adapts the training data according to the learner’s level of understanding. The fixed curriculum demonstrated superior performance on the BLiMP-Supplement, EWoK, and COMPS benchmarks, while the adaptive feedback curriculum substantially improved the Entity Tracking performance in one configuration. However, our official baseline model achieved the highest BLiMP, SuperGLUE, and overall macro-average scores, indicating that our curriculum methods did not produce the overall improvements that we had hypothesized. Our code is available at https://github.com/dcgaines/babylm-2026.
%U https://aclanthology.org/2026.babylm-main.55/
%P 689-701
Markdown (Informal)
[Learning from Correction: Data-Efficient Language Model Training Through Adaptive Teacher-Model Feedback](https://aclanthology.org/2026.babylm-main.55/) (Islam et al., BabyLM 2026)
ACL