@inproceedings{darshan-divekar-2026-gradients,
title = "When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for {LLM} Judges",
author = "Darshan, Parth and
Divekar, Abhishek",
editor = "Mysore, Sheshera and
Kumar, Sachin and
Balachandran, Vidhisha and
Hayati, Shirley Anugrah and
Brahman, Faeze and
Moussa, Hanane Nour and
Salemi, Alireza",
booktitle = "Proceedings of the Second Workshop on Customizable {NLP}: Progress and Challenges in Customizing {NLP} for a Domain, Application, Group, or Individual ({C}ustom{NLP}4{U})",
month = jul,
year = "2026",
address = "San Diego, California, USA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.customnlp4u-1.21/",
pages = "216--228",
ISBN = "979-8-89176-396-8",
abstract = "Customizing an LLM judge to a specific task or domain often involves optimizing its prompt across multiple evaluation criteria simultaneously. Textual gradient methods automate this for a single judge criterion, however they produce natural-language critiques, not numerical vectors. Thus, the conflict-resolution toolkit of multi-task learning (PCGrad, MGDA) doesn{'}t apply to the multi-objective textual gradient setting. We test five decomposition modes of textual gradient optimizers by varying how much cross-task information the loss, gradient and optimizer LLMs share. In 6 of 10 configurations on SummEval, we observe that optimization never improves over the initial prompt. Gradient specificity drops by 59{\%} (from 9.0 to 3.7) when the gradient LLM processes multiple criteria jointly. Separately, we observe that naively combining per-task instructions into a single prompt degrades Spearman{'}s {\ensuremath{\rho}} by -5.3{\%}. These results identify two separable failure modes: optimization-time gradient dilution and inference-time instruction interference, which together constrain the design space for multi-objective judge customization using textual feedback."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="darshan-divekar-2026-gradients">
<titleInfo>
<title>When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges</title>
</titleInfo>
<name type="personal">
<namePart type="given">Parth</namePart>
<namePart type="family">Darshan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Abhishek</namePart>
<namePart type="family">Divekar</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 Second Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual (CustomNLP4U)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sheshera</namePart>
<namePart type="family">Mysore</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sachin</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Vidhisha</namePart>
<namePart type="family">Balachandran</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shirley</namePart>
<namePart type="given">Anugrah</namePart>
<namePart type="family">Hayati</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Faeze</namePart>
<namePart type="family">Brahman</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hanane</namePart>
<namePart type="given">Nour</namePart>
<namePart type="family">Moussa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alireza</namePart>
<namePart type="family">Salemi</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, USA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-396-8</identifier>
</relatedItem>
<abstract>Customizing an LLM judge to a specific task or domain often involves optimizing its prompt across multiple evaluation criteria simultaneously. Textual gradient methods automate this for a single judge criterion, however they produce natural-language critiques, not numerical vectors. Thus, the conflict-resolution toolkit of multi-task learning (PCGrad, MGDA) doesn’t apply to the multi-objective textual gradient setting. We test five decomposition modes of textual gradient optimizers by varying how much cross-task information the loss, gradient and optimizer LLMs share. In 6 of 10 configurations on SummEval, we observe that optimization never improves over the initial prompt. Gradient specificity drops by 59% (from 9.0 to 3.7) when the gradient LLM processes multiple criteria jointly. Separately, we observe that naively combining per-task instructions into a single prompt degrades Spearman’s \ensuremathρ by -5.3%. These results identify two separable failure modes: optimization-time gradient dilution and inference-time instruction interference, which together constrain the design space for multi-objective judge customization using textual feedback.</abstract>
<identifier type="citekey">darshan-divekar-2026-gradients</identifier>
<location>
<url>https://aclanthology.org/2026.customnlp4u-1.21/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>216</start>
<end>228</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges
%A Darshan, Parth
%A Divekar, Abhishek
%Y Mysore, Sheshera
%Y Kumar, Sachin
%Y Balachandran, Vidhisha
%Y Hayati, Shirley Anugrah
%Y Brahman, Faeze
%Y Moussa, Hanane Nour
%Y Salemi, Alireza
%S Proceedings of the Second Workshop on Customizable NLP: Progress and Challenges in Customizing NLP for a Domain, Application, Group, or Individual (CustomNLP4U)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, USA
%@ 979-8-89176-396-8
%F darshan-divekar-2026-gradients
%X Customizing an LLM judge to a specific task or domain often involves optimizing its prompt across multiple evaluation criteria simultaneously. Textual gradient methods automate this for a single judge criterion, however they produce natural-language critiques, not numerical vectors. Thus, the conflict-resolution toolkit of multi-task learning (PCGrad, MGDA) doesn’t apply to the multi-objective textual gradient setting. We test five decomposition modes of textual gradient optimizers by varying how much cross-task information the loss, gradient and optimizer LLMs share. In 6 of 10 configurations on SummEval, we observe that optimization never improves over the initial prompt. Gradient specificity drops by 59% (from 9.0 to 3.7) when the gradient LLM processes multiple criteria jointly. Separately, we observe that naively combining per-task instructions into a single prompt degrades Spearman’s \ensuremathρ by -5.3%. These results identify two separable failure modes: optimization-time gradient dilution and inference-time instruction interference, which together constrain the design space for multi-objective judge customization using textual feedback.
%U https://aclanthology.org/2026.customnlp4u-1.21/
%P 216-228
Markdown (Informal)
[When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges](https://aclanthology.org/2026.customnlp4u-1.21/) (Darshan & Divekar, CustomNLP4U 2026)
ACL