Recent developments in AI governance and safety research have called for red-teaming methods that can effectively surface potential risks posed by AI models. Many of these calls have emphasized how the identities and backgrounds of red-teamers can shape their red-teaming strategies, and thus the kinds of risks they are likely to uncover. While automated red-teaming approaches promise to complement human red-teaming by enabling larger-scale exploration of model behavior, current approaches do not consider the role of identity. As an initial step towards incorporating people's background and identities in automated red-teaming, we develop and evaluate a novel method, PersonaTeaming, that introduces personas in the adversarial prompt generation process to explore a wider spectrum of adversarial strategies. In particular, we first introduce a methodology for mutating prompts based on either "red-teaming expert" personas or "regular AI user" personas. We then develop a dynamic persona-generating algorithm that automatically generates various persona types adaptive to different seed prompts. In addition, we develop a set of new metrics to explicitly measure the "mutation distance" to complement existing diversity measurements of adversarial prompts. Our experiments show promising improvements (up to 144.1%) in the attack success rates of adversarial prompts through persona mutation, while maintaining prompt diversity, compared to RainbowPlus, a state-of-the-art automated red-teaming method. We discuss the strengths and limitations of different persona types and mutation methods, shedding light on future opportunities to explore complementarities between automated and human red-teaming approaches.

A prevailing view holds that supervised fine-tuning (SFT) memorizes training data and fails to generalize, whereas reinforcement learning (RL) attains broader robustness. We revisit this claim through a systematic evaluation on two decision-making benchmarks, Sokoban and General Points, and arrive at a different conclusion. We show that much of SFT's perceived failure stems from frozen-prompt artifacts: when trained on fixed instruction templates, SFT models cling to training semantics rather than adapting to new ones. Introducing prompt diversity during training breaks this shortcut and yields strong generalization to unseen instruction variants without harming in-distribution performance. Beyond instruction shifts, we ask whether SFT can generalize to strictly harder tasks. Here, chain-of-thought (CoT) supervision provides an algorithmic scaffold that markedly improves transfer to more difficult regimes, such as larger Sokoban grids with additional boxes and arithmetic with out-of-distribution values or five-card compositions that increase combinatorial complexity. Finally, combining prompt diversity with CoT achieves the best of both worlds: robust generalization across both instruction-variant and difficulty-variant settings, matching or surpassing RL baselines on our benchmarks while retaining SFT's simplicity and stability. These findings challenge the narrative that SFT is inherently inferior to RL and support a data-centric perspective: with appropriately curated demonstrations, vanilla SFT can generalize as strongly as RL.

Personal use of this material is permitted. Abstract --Prompt engineering is essential for optimizing large language models (LLMs), yet the link between prompt structures and task performance remains underexplored. This work introduces an evolutionary approach that combines context-free grammar (CFG) with the MAP-Elites algorithm to systematically explore the prompt space. Our method prioritizes quality and diversity, generating high-performing and structurally varied prompts while analyzing their alignment with diverse tasks by varying traits such as the number of examples (shots) and reasoning depth. By systematically mapping the phenotypic space, we reveal how structural variations influence LLM performance, offering actionable insights for task-specific and adaptable prompt design. Evaluated on seven BigBench Lite tasks across multiple LLMs, our results underscore the critical interplay of quality and diversity, advancing the effectiveness and versatility of LLMs. The rapid advancement of Generative Pre-Trained Transformer (GPT)-based Large Language Models (LLMs), such as ChatGPT, has revolutionized the field of Natural Language Processing (NLP) [1]. These models excel across domains, emphasizing the importance of prompt engineering to optimize performance and bridge user intent with model output [2], [3].

Large Language Models still encounter substantial challenges in reasoning tasks, especially for smaller models, which many users may be restricted to due to resource constraints (e.g. GPU memory restrictions). Inference-time methods to boost LLM performance, such as prompting methods to invoke certain reasoning pathways in responses, have been shown effective in past works, though they largely rely on sequential queries. The ensemble method, which consists of multiple constituent models running in parallel, is a promising approach to achieving better inference-time performance, especially given recent developments that enabled significant speed-ups in LLM batch inference. In this work, we propose a novel, training-free LLM ensemble framework where a single LLM model is fed an optimized, diverse set of prompts in parallel, effectively producing an ensemble at inference time to achieve performance improvement in reasoning tasks. We empirically demonstrate that our method leads to significant gains on math reasoning tasks, e.g., on MATH, where our ensemble consisting of a few small models (e.g., three Qwen2-MATH-1.5B-it models) can outperform a larger model (e.g., Qwen2-MATH-7B-it).

Alignment with human preference prevents large language models (LLMs) from generating misleading or toxic content while requiring high-cost human feedback. Assuming resources of human annotation are limited, there are two different ways of allocating considered: more diverse PROMPTS or more diverse RESPONSES to be labeled. Nonetheless, a straightforward comparison between their impact is absent. In this work, we first control the diversity of both sides according to the number of samples for fine-tuning, which can directly reflect their influence. We find that instead of numerous prompts, more responses but fewer prompts better trigger LLMs for human alignment. Additionally, the concept of diversity for prompts can be more complex than responses that are typically quantified by single digits. Consequently, a new formulation of prompt diversity is proposed, further implying a linear correlation with the final performance of LLMs after fine-tuning. We also leverage it on data augmentation and conduct experiments to show its effect on different algorithms.