Pref-CTRL: Preference Driven LLM Alignment using Representation Editing

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ashrafi, Imranul, Unanue, Inigo Jauregi, Piccardi, Massimo
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918468355358720
author Ashrafi, Imranul
Unanue, Inigo Jauregi
Piccardi, Massimo
author_facet Ashrafi, Imranul
Unanue, Inigo Jauregi
Piccardi, Massimo
contents Test-time alignment methods offer a promising alternative to fine-tuning by steering the outputs of large language models (LLMs) at inference time with lightweight interventions on their internal representations. Recently, a prominent and effective approach, RE-Control (Kong et al., 2024), has proposed leveraging an external value function trained over the LLM's hidden states to guide generation via gradient-based editing. While effective, this method overlooks a key characteristic of alignment tasks, i.e. that they are typically formulated as learning from human preferences between candidate responses. To address this, in this paper we propose a novel preference-based training framework, Pref-CTRL, that uses a multi-objective value function to better reflect the structure of preference data. Our approach has outperformed RE-Control on two benchmark datasets and showed greater generalization on out-of-domain datasets. Our source code is available at https://github.com/UTS-nlPUG/pref-ctrl.
format Preprint
id arxiv_https___arxiv_org_abs_2604_23543
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Pref-CTRL: Preference Driven LLM Alignment using Representation Editing
Ashrafi, Imranul
Unanue, Inigo Jauregi
Piccardi, Massimo
Computation and Language
Artificial Intelligence
Test-time alignment methods offer a promising alternative to fine-tuning by steering the outputs of large language models (LLMs) at inference time with lightweight interventions on their internal representations. Recently, a prominent and effective approach, RE-Control (Kong et al., 2024), has proposed leveraging an external value function trained over the LLM's hidden states to guide generation via gradient-based editing. While effective, this method overlooks a key characteristic of alignment tasks, i.e. that they are typically formulated as learning from human preferences between candidate responses. To address this, in this paper we propose a novel preference-based training framework, Pref-CTRL, that uses a multi-objective value function to better reflect the structure of preference data. Our approach has outperformed RE-Control on two benchmark datasets and showed greater generalization on out-of-domain datasets. Our source code is available at https://github.com/UTS-nlPUG/pref-ctrl.
title Pref-CTRL: Preference Driven LLM Alignment using Representation Editing
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2604.23543