Steer Model beyond Assistant: Controlling System Prompt Strength via Contrastive Decoding

TL;DR

System prompt strength introduces a decoding-time knob $\alpha$ that amplifies the behavioral delta between a target persona and the default assistant by applying contrastive decoding within a single model. By contrasting logits from the target and default system prompts, the method sharpens adherence to non-standard instructions without retraining, yielding measurable gains in strict instruction adherence, refusal behavior, and in-context steerability across diverse model families and datasets. The approach reveals a trade-off at higher $\alpha$, where stronger steering can impair general capabilities, but with practical defaults (e.g., $\alpha$ in [0.5,1]) practitioners can achieve robust control while preserving core performance. Overall, system prompt strength offers dynamic, training-free controllability for pluralistic alignment, domain-specific agents, and pedagogy-oriented roles, while remaining compatible with existing prompt engineering and safety frameworks.

Abstract

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.

Figures (2)

• Figure 1: System Prompt Strength via Contrastive Decoding.Left: Standard decoding fails to follow system prompts ($\times$), while our method successfully steers the model ($\checkmark$). Middle: Our method uses contrastive decoding to shift generation from the "Default Assistant Space" towards a "Steered Output Space" controlled by strength $\alpha$. Right: by contrasting target and default prompt logits, we suppress generic assistant tokens and boost persona-aligned tokens, with $\alpha$ controlling the amplification strength.
• Figure 2: Effect of $\alpha$ on OffTopicEval OS (left y-axis, blue) and IFEval Strict accuracy (right y-axis, orange).