@inproceedings{li-etal-2026-llms-learn,
title = "What Do {LLM}s Learn First? Asymmetric Learning Dynamics of Input Complexity and Output Ambiguity in Preference Alignment",
author = "Li, Mengyang and
Wang, Jingwen and
Zhao, Pinlong",
editor = "Liakata, Maria and
Moreira, Viviane P. and
Zhang, Jiajun and
Jurgens, David",
booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.acl-long.789/",
pages = "17373--17388",
ISBN = "979-8-89176-390-6",
abstract = "Direct Preference Optimization (DPO) has become a standard approach for aligning large language models with human preferences, yet existing methods treat all preference pairs uniformly during training. We identify two distinct sources of learning difficulty: Input Complexity (IC), capturing prompt understanding challenges, and Output Ambiguity (OA), measuring preference discrimination difficulty. Through systematic analysis, we demonstrate that these dimensions induce asymmetric learning dynamics, with IC-related competencies developing rapidly in early training while OA-related competencies emerge more gradually. Building on this observation, we propose DECOPO, a training framework that maintains separate, adaptive pacing schedules for each dimension. Experiments on UltraFeedback show that DECOPO achieves 42.3{\%} length-controlled win rate on AlpacaEval 2.0 and 7.66 on MT-Bench, outperforming curriculum baselines by 2.1{\%} and 0.21 points respectively, while matching full-data baseline performance with only 75{\%} of training samples."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-etal-2026-llms-learn">
<titleInfo>
<title>What Do LLMs Learn First? Asymmetric Learning Dynamics of Input Complexity and Output Ambiguity in Preference Alignment</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mengyang</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingwen</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pinlong</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Maria</namePart>
<namePart type="family">Liakata</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Viviane</namePart>
<namePart type="given">P</namePart>
<namePart type="family">Moreira</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiajun</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">David</namePart>
<namePart type="family">Jurgens</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, California, United States</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-390-6</identifier>
</relatedItem>
<abstract>Direct Preference Optimization (DPO) has become a standard approach for aligning large language models with human preferences, yet existing methods treat all preference pairs uniformly during training. We identify two distinct sources of learning difficulty: Input Complexity (IC), capturing prompt understanding challenges, and Output Ambiguity (OA), measuring preference discrimination difficulty. Through systematic analysis, we demonstrate that these dimensions induce asymmetric learning dynamics, with IC-related competencies developing rapidly in early training while OA-related competencies emerge more gradually. Building on this observation, we propose DECOPO, a training framework that maintains separate, adaptive pacing schedules for each dimension. Experiments on UltraFeedback show that DECOPO achieves 42.3% length-controlled win rate on AlpacaEval 2.0 and 7.66 on MT-Bench, outperforming curriculum baselines by 2.1% and 0.21 points respectively, while matching full-data baseline performance with only 75% of training samples.</abstract>
<identifier type="citekey">li-etal-2026-llms-learn</identifier>
<location>
<url>https://aclanthology.org/2026.acl-long.789/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>17373</start>
<end>17388</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T What Do LLMs Learn First? Asymmetric Learning Dynamics of Input Complexity and Output Ambiguity in Preference Alignment
%A Li, Mengyang
%A Wang, Jingwen
%A Zhao, Pinlong
%Y Liakata, Maria
%Y Moreira, Viviane P.
%Y Zhang, Jiajun
%Y Jurgens, David
%S Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, United States
%@ 979-8-89176-390-6
%F li-etal-2026-llms-learn
%X Direct Preference Optimization (DPO) has become a standard approach for aligning large language models with human preferences, yet existing methods treat all preference pairs uniformly during training. We identify two distinct sources of learning difficulty: Input Complexity (IC), capturing prompt understanding challenges, and Output Ambiguity (OA), measuring preference discrimination difficulty. Through systematic analysis, we demonstrate that these dimensions induce asymmetric learning dynamics, with IC-related competencies developing rapidly in early training while OA-related competencies emerge more gradually. Building on this observation, we propose DECOPO, a training framework that maintains separate, adaptive pacing schedules for each dimension. Experiments on UltraFeedback show that DECOPO achieves 42.3% length-controlled win rate on AlpacaEval 2.0 and 7.66 on MT-Bench, outperforming curriculum baselines by 2.1% and 0.21 points respectively, while matching full-data baseline performance with only 75% of training samples.
%U https://aclanthology.org/2026.acl-long.789/
%P 17373-17388
Markdown (Informal)
[What Do LLMs Learn First? Asymmetric Learning Dynamics of Input Complexity and Output Ambiguity in Preference Alignment](https://aclanthology.org/2026.acl-long.789/) (Li et al., ACL 2026)
ACL