Manipulating Transformer-Based Models: Controllability, Steerability, and Robust Interventions

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Alpay, Faruk, Alpay, Taylan
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866911138794438656
author Alpay, Faruk
Alpay, Taylan
author_facet Alpay, Faruk
Alpay, Taylan
contents Transformer-based language models excel in NLP tasks, but fine-grained control remains challenging. This paper explores methods for manipulating transformer models through principled interventions at three levels: prompts, activations, and weights. We formalize controllable text generation as an optimization problem addressable via prompt engineering, parameter-efficient fine-tuning, model editing, and reinforcement learning. We introduce a unified framework encompassing prompt-level steering, activation interventions, and weight-space edits. We analyze robustness and safety implications, including adversarial attacks and alignment mitigations. Theoretically, we show minimal weight updates can achieve targeted behavior changes with limited side-effects. Empirically, we demonstrate >90% success in sentiment control and factual edits while preserving base performance, though generalization-specificity trade-offs exist. We discuss ethical dual-use risks and the need for rigorous evaluation. This work lays groundwork for designing controllable and robust language models.
format Preprint
id arxiv_https___arxiv_org_abs_2509_04549
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Manipulating Transformer-Based Models: Controllability, Steerability, and Robust Interventions
Alpay, Faruk
Alpay, Taylan
Computation and Language
Artificial Intelligence
68T50, 68T05
I.2.7; I.2.6; I.2.11
Transformer-based language models excel in NLP tasks, but fine-grained control remains challenging. This paper explores methods for manipulating transformer models through principled interventions at three levels: prompts, activations, and weights. We formalize controllable text generation as an optimization problem addressable via prompt engineering, parameter-efficient fine-tuning, model editing, and reinforcement learning. We introduce a unified framework encompassing prompt-level steering, activation interventions, and weight-space edits. We analyze robustness and safety implications, including adversarial attacks and alignment mitigations. Theoretically, we show minimal weight updates can achieve targeted behavior changes with limited side-effects. Empirically, we demonstrate >90% success in sentiment control and factual edits while preserving base performance, though generalization-specificity trade-offs exist. We discuss ethical dual-use risks and the need for rigorous evaluation. This work lays groundwork for designing controllable and robust language models.
title Manipulating Transformer-Based Models: Controllability, Steerability, and Robust Interventions
topic Computation and Language
Artificial Intelligence
68T50, 68T05
I.2.7; I.2.6; I.2.11
url https://arxiv.org/abs/2509.04549