@inproceedings{mishra-roy-2026-spend,
title = "Spend the Wasted Epochs: Preference Post-Training from the Saturation Checkpoint for Sample-Efficient Language Modeling",
author = "Mishra, Ujjwal and
Roy, Sudipta",
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.52/",
pages = "646--658",
ISBN = "979-8-89176-475-0",
abstract = "Sample-efficient pretraining under a developmental data budget, such as the ten million words and ten epochs of the BabyLM Strict-Small track, is usually approached by adding training signal: a new objective, optimizer, curriculum, or teacher. The budget itself receives less scrutiny. We show that much of it is spent after learning has stopped. A 33.0M-parameter GPT-BERT hybrid trained under the track rules saturates on downstream evaluations by roughly 70M words seen, about 70{\%} of its permitted exposure, while training loss continues to fall; the final epochs improve fit, not competence. From this diagnosis we derive three design principles, timing, preservation, and alignment, and instantiate them as Minimal-Pair Preference Optimization (MPO): a short post-training phase, begun at the saturation checkpoint, that teaches the model to prefer real corpus sentences over length-preserving corruptions of them under a frozen-reference penalty. A controlled study of seven in-training interventions supports the principles: none improves on the baseline, and the unanchored ones damage the same cluster of long-distance dependencies first. MPO improves five of the seven official task averages, concedes under a point on the other two, is alone in repairing that cluster, and matches the fully trained baseline with 7.2 rather than 10 epochs."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="mishra-roy-2026-spend">
<titleInfo>
<title>Spend the Wasted Epochs: Preference Post-Training from the Saturation Checkpoint for Sample-Efficient Language Modeling</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ujjwal</namePart>
<namePart type="family">Mishra</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sudipta</namePart>
<namePart type="family">Roy</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>Sample-efficient pretraining under a developmental data budget, such as the ten million words and ten epochs of the BabyLM Strict-Small track, is usually approached by adding training signal: a new objective, optimizer, curriculum, or teacher. The budget itself receives less scrutiny. We show that much of it is spent after learning has stopped. A 33.0M-parameter GPT-BERT hybrid trained under the track rules saturates on downstream evaluations by roughly 70M words seen, about 70% of its permitted exposure, while training loss continues to fall; the final epochs improve fit, not competence. From this diagnosis we derive three design principles, timing, preservation, and alignment, and instantiate them as Minimal-Pair Preference Optimization (MPO): a short post-training phase, begun at the saturation checkpoint, that teaches the model to prefer real corpus sentences over length-preserving corruptions of them under a frozen-reference penalty. A controlled study of seven in-training interventions supports the principles: none improves on the baseline, and the unanchored ones damage the same cluster of long-distance dependencies first. MPO improves five of the seven official task averages, concedes under a point on the other two, is alone in repairing that cluster, and matches the fully trained baseline with 7.2 rather than 10 epochs.</abstract>
<identifier type="citekey">mishra-roy-2026-spend</identifier>
<location>
<url>https://aclanthology.org/2026.babylm-main.52/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>646</start>
<end>658</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Spend the Wasted Epochs: Preference Post-Training from the Saturation Checkpoint for Sample-Efficient Language Modeling
%A Mishra, Ujjwal
%A Roy, Sudipta
%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 mishra-roy-2026-spend
%X Sample-efficient pretraining under a developmental data budget, such as the ten million words and ten epochs of the BabyLM Strict-Small track, is usually approached by adding training signal: a new objective, optimizer, curriculum, or teacher. The budget itself receives less scrutiny. We show that much of it is spent after learning has stopped. A 33.0M-parameter GPT-BERT hybrid trained under the track rules saturates on downstream evaluations by roughly 70M words seen, about 70% of its permitted exposure, while training loss continues to fall; the final epochs improve fit, not competence. From this diagnosis we derive three design principles, timing, preservation, and alignment, and instantiate them as Minimal-Pair Preference Optimization (MPO): a short post-training phase, begun at the saturation checkpoint, that teaches the model to prefer real corpus sentences over length-preserving corruptions of them under a frozen-reference penalty. A controlled study of seven in-training interventions supports the principles: none improves on the baseline, and the unanchored ones damage the same cluster of long-distance dependencies first. MPO improves five of the seven official task averages, concedes under a point on the other two, is alone in repairing that cluster, and matches the fully trained baseline with 7.2 rather than 10 epochs.
%U https://aclanthology.org/2026.babylm-main.52/
%P 646-658
Markdown (Informal)
[Spend the Wasted Epochs: Preference Post-Training from the Saturation Checkpoint for Sample-Efficient Language Modeling](https://aclanthology.org/2026.babylm-main.52/) (Mishra & Roy, BabyLM 2026)
ACL