Steer Model beyond Assistant: Controlling System Prompt Strength via Contrastive Decoding
Not on the map. Why?
Abstract (arXiv)
Large language models excel at complex instructions yet struggle to deviate from their helpful assistant persona, as post-training instills strong priors that resist conflicting instructions. We introduce system prompt strength, a training-free method that treats prompt adherence as a continuous control. By contrasting logits from target and default system prompts, we isolate and amplify the behavioral signal unique to the target persona by a scalar factor alpha. Across five diverse benchmarks spanning constraint satisfaction, behavioral control, pluralistic alignment, capability modulation, and stylistic control, our method yields substantial improvements: up to +8.5 strict accuracy on IFEval, +45pp refusal rate on OffTopicEval, and +13% steerability on Prompt-Steering. Our approach enables practitioners to modulate system prompt strength, providing dynamic control over model behavior without retraining.
Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Builds on 38 of its 54 references that are in the map. First posted on arXiv 10 Jan 2026.
Builds on
38 of 54 references.
References · 38
- 2022Training language models to follow instructions with human feedbackcited by 15,712
- 2022Constitutional AI: Harmlessness from AI Feedbackcited by 2,950
- 2023Instruction-Following Evaluation for Large Language Modelscited by 1,143
- 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
- 2023Steering Language Models With Activation Engineeringcited by 824
- 2019Plug and Play Language Models: A Simple Approach to Controlled Text Generationcited by 793
- 2023Whose Opinions Do Language Models Reflect?cited by 759
- 2022Contrastive Decoding: Open-ended Text Generation as Optimizationcited by 598
- 2021FUDGE: Controlled Text Generation With Future Discriminatorscited by 374
- 2024Quantifying the Persona Effect in LLM Simulationscited by 164
- 2024Can LLM be a Personalized Judge?cited by 99
- 2023Stay on topic with Classifier-Free Guidancecited by 90
26 more
- 2023On the steerability of large language models toward data-driven personascited by 53
- 2025Flipping the Dialogue: Training and Evaluating User Language Modelscited by 50
- 2025SimBench: Benchmarking the Ability of Large Language Models to Simulate Human Behaviorscited by 48
- 2024Towards Reliable Evaluation of Behavior Steering Interventions in LLMscited by 33
- 2024AI PERSONA: Towards Life-long Personalization of LLMscited by 30
- 2024LLMs + Persona-Plug = Personalized LLMscited by 29
- 2024Nudging: Inference-time Alignment of LLMs via Guided Decodingcited by 22
- 2024The Potential and Challenges of Evaluating Attitudes, Opinions, and Values in Large Language Modelscited by 19
- 2024Adaptive Contrastive Decoding in Retrieval-Augmented Generation for Handling Noisy Contextscited by 19
- 2025Localized Cultural Knowledge is Conserved and Controllable in Large Language Modelscited by 19
- 2025Learning to Make MISTAKEs: Modeling Incorrect Student Thinking And Key Errorscited by 19
- 2024ToxiCraft: A Novel Framework for Synthetic Generation of Harmful Informationcited by 17
- 2025Spectrum Tuning: Post-Training for Distributional Coverage and In-Context Steerabilitycited by 17
- 2023Know Your Audience: Do LLMs Adapt to Different Age and Education Levels?cited by 16
- 2024From Tarzan to Tolkien: Controlling the Language Proficiency Level of LLMs for Content Generationcited by 15
- 2025Navigating the Alignment-Calibration Trade-off: A Pareto-Superior Frontier via Model Mergingcited by 10
- 2025When Personalization Meets Reality: A Multi-Faceted Analysis of Personalized Preference Learningcited by 9
- 2025Toward Safe and Human-Aligned Game Conversational Recommendation via Multi-Agent Decompositioncited by 9
- 2025A Course Correction in Steerability Evaluation: Revealing Miscalibration and Side Effects in LLMscited by 7
- 2025Alignment Drift in CEFR-prompted LLMs for Interactive Spanish Tutoringcited by 6
- 2025Privacy-R1: Privacy-Aware Multi-LLM Agent Collaboration via Reinforcement Learningcited by 6
- 2026Decoupling the Effect of Chain-of-Thought Reasoning: A Human Label Variation Perspectivecited by 6
- 2025iNews: A Multimodal Dataset for Modeling Personalized Affective Responses to Newscited by 4
- 2025Personas within Parameters: Fine-Tuning Small Language Models with Low-Rank Adapters to Mimic User Behaviorscited by 4
- 2025OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost Always!cited by 4
- 2025Toward Beginner-Friendly LLMs for Language Learning: Controlling Difficulty in Conversationcited by 3
Built on by
Citing papers per year: 2026: 1.
1 paper.