When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Darshan, Parth, Divekar, Abhishek
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911716734926848
author Darshan, Parth
Divekar, Abhishek
author_facet Darshan, Parth
Divekar, Abhishek
contents 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, 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 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.
format Preprint
id arxiv_https___arxiv_org_abs_2605_26046
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges
Darshan, Parth
Divekar, Abhishek
Computation and Language
Artificial Intelligence
Machine Learning
Multiagent Systems
Software Engineering
I.2.7; I.2.6; I.2.4; I.2.8
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, 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 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.
title When Gradients Collide: Failure Modes of Multi-Objective Prompt Optimization for LLM Judges
topic Computation and Language
Artificial Intelligence
Machine Learning
Multiagent Systems
Software Engineering
I.2.7; I.2.6; I.2.4; I.2.8
url https://arxiv.org/abs/2605.26046