paper-with-me

Papers

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

2025-09-04 · Faruk Alpay, Taylan Alpay arxiv

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.

📄 PDF Abstract BibTeX arXiv:2509.04549

Code (0)

등록된 구현이 없습니다.

Tasks

parameter-efficient fine-tuningReinforcement LearningPrompt EngineeringText Generation

Similar Papers 제목 키워드 기반

How Does Controllability Emerge In Language Models During Pretraining?

2025-08-03 · Jianshu She, Xinyue Li, Eric Xing, Zhengzhong Liu 외 arxiv

Language models can be steered by modifying their internal representations to control concepts such as emotion, style, or truthfulness in generation. However, the conditions for an effective intervention remain unclear a…

SteerEval: A Framework for Evaluating Steerability with Natural Language Profiles for Recommendation

2026-01-28 · Joyce Zhou, Weijie Zhou, Doug Turnbull, Thorsten Joachims arxiv

Natural-language user profiles have recently attracted attention not only for improved interpretability, but also for their potential to make recommender systems more steerable. By enabling direct editing, natural-langua…

A Course Correction in Steerability Evaluation: Revealing Miscalibration and Side Effects in LLMs

2025-05-27 · Trenton Chang, Tobias Schnabel, Adith Swaminathan, Jenna Wiens

Despite advances in large language models (LLMs) on reasoning and instruction-following benchmarks, it remains unclear whether they can reliably produce outputs aligned with a broad variety of user goals, a concept we re…

Instruction FollowingPrompt Engineering

Steered Generation via Gradient-Based Optimization on Sparse Query Features

2026-05-21 · Sumanta Bhattacharyya, Pedram Rooshenas arxiv

Latent steering exploits internal representations of Large Language Models (LLMs) to guide generation, yet interventions on dense states can entangle distinct semantic features. In this paper, we investigate attention qu…

The Geometric Canary: Predicting Steerability and Detecting Drift via Representational Stability

2026-04-20 · Prashant C. Raju arxiv

Reliable deployment of language models requires two capabilities that appear distinct but share a common geometric foundation: predicting whether a model will accept targeted behavioral control, and detecting when its in…