Saved in:
Bibliographic Details
Main Authors: Soor, Sampriti, Ghosh, Suklav, Sur, Arijit
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2512.08131
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908699754233856
author Soor, Sampriti
Ghosh, Suklav
Sur, Arijit
author_facet Soor, Sampriti
Ghosh, Suklav
Sur, Arijit
contents Language models are vulnerable to short adversarial suffixes that can reliably alter predictions. Previous works usually find such suffixes with gradient search or rule-based methods, but these are brittle and often tied to a single task or model. In this paper, a reinforcement learning framework is used where the suffix is treated as a policy and trained with Proximal Policy Optimization against a frozen model as a reward oracle. Rewards are shaped using calibrated cross-entropy, removing label bias and aggregating across surface forms to improve transferability. The proposed method is evaluated on five diverse NLP benchmark datasets, covering sentiment, natural language inference, paraphrase, and commonsense reasoning, using three distinct language models: Qwen2-1.5B Instruct, TinyLlama-1.1B Chat, and Phi-1.5. Results show that RL-trained suffixes consistently degrade accuracy and transfer more effectively across tasks and models than previous adversarial triggers of similar genres.
format Preprint
id arxiv_https___arxiv_org_abs_2512_08131
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Universal Adversarial Suffixes for Language Models Using Reinforcement Learning with Calibrated Reward
Soor, Sampriti
Ghosh, Suklav
Sur, Arijit
Computation and Language
Language models are vulnerable to short adversarial suffixes that can reliably alter predictions. Previous works usually find such suffixes with gradient search or rule-based methods, but these are brittle and often tied to a single task or model. In this paper, a reinforcement learning framework is used where the suffix is treated as a policy and trained with Proximal Policy Optimization against a frozen model as a reward oracle. Rewards are shaped using calibrated cross-entropy, removing label bias and aggregating across surface forms to improve transferability. The proposed method is evaluated on five diverse NLP benchmark datasets, covering sentiment, natural language inference, paraphrase, and commonsense reasoning, using three distinct language models: Qwen2-1.5B Instruct, TinyLlama-1.1B Chat, and Phi-1.5. Results show that RL-trained suffixes consistently degrade accuracy and transfer more effectively across tasks and models than previous adversarial triggers of similar genres.
title Universal Adversarial Suffixes for Language Models Using Reinforcement Learning with Calibrated Reward
topic Computation and Language
url https://arxiv.org/abs/2512.08131