This paper presents the CriticalBrew submission to the CQs-Gen 2025 shared task, which focuses on generating critical questions (CQs) for a given argument. Our approach employs a multi-agent framework containing two sequential components: 1) Generation: machine society simulation for generating CQs and 2) Evaluation: LLM-based evaluation for selecting the top three questions. The first models collaboration as a sequence of thinking patterns (e.g., debate → reflect). The second assesses the generated questions using zero-shot prompting, evaluating them against several criteria (e.g., depth). Experiments with different open-weight LLMs (small vs. large) consistently outperformed the baseline, a single LLM with zero-shot prompting. Two configurations, agent count and thinking patterns, significantly impacted the performance in the shared task’s CQ-usefulness evaluation, whereas different LLM-based evaluation strategies (e.g., scoring) had no impact. Our code is available on GitHub.

This research explores strategies for steering the output of large language models (LLMs) towards specific styles, such as sentiment, emotion, or writing style, by adding style vectors to the activations of hidden layers during text generation. We show that style vectors can be simply computed from recorded layer activations for input texts in a specific style in contrast to more complex training-based approaches. Through a series of experiments, we demonstrate the effectiveness of activation engineering using such style vectors to influence the style of generated text in a nuanced and parameterisable way, distinguishing it from prompt engineering. The presented research constitutes a significant step towards developing more adaptive and effective AI-empowered interactive systems.