TL;DR
Prompting Large Language Models (LLMs) to adopt specific personas or styles (like "be concise") triggers unintended behavioral shifts across other stylistic dimensions (like appearing less expert). This paper systematically maps these "stylistic side effects" and demonstrates that current mitigation strategies suffer from an "alignment tax," often failing to cleanly separate entangled traits without degrading the primary style.
What Problem Does This Paper Address?
When designing conversational agents, developers frequently use system prompts like "be concise," "be professional," or "be empathetic." This practice relies on a fundamental, untested assumption: that style features are orthogonal and can be controlled independently.
In reality, LLMs often bundle traits based on patterns in their training data, social conventions, and latent representations. For instance, human experts often provide detailed, nuanced explanations; if you force an LLM to be concise, does it implicitly stop sounding like an expert? This paper presents the first systematic, statistical investigation into how stylistic controls interfere with one another—a phenomenon the authors term stylistic side effects.
Methods
The authors designed a multi-step pipeline to extract, evaluate, and attempt to mitigate these side effects:
- Survey & Feature Extraction: The authors surveyed 127 conversational agent papers from the ACL Anthology (2023–2025) to identify the 12 most frequently used style features: concise, expert, helpful, empathetic, friendly, detailed, engaging, curious, polite, impartial, outgoing, and efficient.
- Causal Evaluation Framework: They generated responses with and without style modifiers across three models: GPT-5-mini, Qwen3-8B, and Llama-3-8B-Instruct.
- LLM-as-a-Judge: Using a pairwise win-rate comparison protocol (with Qwen3-8B as the judge), they evaluated how a "Main Feature" prompt causally impacts unrelated "Side Features" compared to a neutral baseline.
- Mitigation Strategies: They tested two approaches to fix these side effects:
- Prompt Intervention: Explicitly requesting both traits (e.g., "Please be concise and expert").
- Steering Intervention: Utilizing Contrastive Activation Addition (CAA) to inject activation vectors for the suppressed trait directly into the model's linear layers during inference.
Datasets
The study utilizes two distinct dialogue datasets to cover different conversational domains:
- LMSYS-Chat-1M: Sampled for task-oriented inquiries (e.g., software debugging, text processing).
- DailyDialog: Sampled for open-domain, daily-life interactions (e.g., culture, finance, relationships).
- CASSE (Conversational Agent Stylistic Side Effects): The newly introduced dataset resulting from this work, containing 12,200 synthetically generated messages annotated with win rates across the 12 style features.
Key Insights and Ideas
- Style Entanglement: Style features in high-dimensional space are deeply entangled rather than orthogonal. Prompting for a single feature activates a constellation of associated traits.
- The Length Confound: Response length heavily dictates perceived style. Social traits (friendly, empathetic) strongly correlate with longer responses, while functional traits (concise, efficient) drastically cut length.
- Asymmetry of Influence: The causal relationships between styles are often one-way. For example, prompting an agent to be an Expert naturally makes it more Efficient (66% win rate), but forcing it to be Efficient strips away its perceived Expertise (dropping to 34.6%).
- The Alignment Tax: Attempting to restore a suppressed trait creates a tug-of-war. For example, trying to force a concise agent to maintain expertise often breaks its ability to remain concise.
Key Results
The authors aggregated their causal findings into a comprehensive win-rate matrix:
- Conciseness vs. Expertise: Prompting for Concise severely penalized perceived Expertise (25.6% win rate) and Helpfulness (23.2%), proving the paper's title: a concise agent is indeed judged as less expert.
- Efficiency vs. Helpfulness: Prompting for Efficient caused a drop in Helpfulness to 33.8%.
- Mitigation Performance: Both Prompt and Steering Interventions successfully restored suppressed side features (e.g., prompting for Efficiency + Helpfulness restored helpfulness from a 0.29 win rate to 0.95).
- The Catch: This restoration came at a steep cost to the primary style. For the Concise-Expert pair, steering successfully brought back expertise but caused the Concise win rate to plummet from 81% down to 37%.
Limitations and Future Work
- Ecological Validity: The findings stem entirely from short, synthetic, single-turn responses rather than long, multi-turn human-to-agent dialogues where side effects might naturally attenuate.
- Model Scale: Due to compute constraints, the evaluation did not include the largest frontier models (e.g., full-scale GPT-4 or Claude 3.5 Sonnet), which might exhibit different latent correlations.
- Simplistic Mitigations: The tested mitigation techniques were lightweight (prompt concatenation and linear activation steering). The authors note that achieving a robust balance without collateral drift will likely require more sophisticated approaches, such as multi-objective reinforcement learning or targeted fine-tuning.
Summary updated