Large Language Models (LLMs) still produce gender-stereotyped language even in occupation-neutral contexts that reflect deep societal biases (Rudinger et al., 2018). To address this, prior work has proposed prompting, constrained decoding (Dathathri et al., 2020; Zhou et al., 2024), post-processing, and fine-tuning-based alignment (Rafailov et al., 2023; Ravfogel et al., 2022). However, the comparative efficacy and learning dynamics remain little understood. We report a comparative analysis of six control techniques for bias mitigation: prompt-only, generate-and-filter, DFA-based Ctrl-G decoding, Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Iterative Nullspace Projection (INLP). We evaluate each method on a compositional constraint task. This task requires generating sentences that contain at least one agentic and one communal descriptor for each of the twenty Winogender-derived occupations. We quantify trade-offs between control strength and naturalness with evaluations of constraint compliance, lexical diversity, and fluency. Our results reveal key contrasts among the methods: SFT achieves 99.87 +- 0.15% compliance and high lexical diversity, while DPO, despite similar training stability, fails at 4.53 +- 0.82%. Ctrl-G guarantees perfect compliance, but at the cost of severely reduced fluency and diversity. Preference-based learning fundamentally differs: it cannot satisfy compositional constraints, as binary preference signals encode ranking, not logical conjunctions. Only explicit positive supervision enables mitigation of compositional biases; preference-based alignment fails to generalize logical structures, underscoring the limitations of preference learning and the necessity of explicit supervision for fair and fluent controlled generation.

This paper introduces an automated framework WSW2.0 for analyzing vocal interactions in preschool classrooms, enhancing both accuracy and scalability through the integration of wav2vec2-based speaker classification and Whisper (large-v2 and large-v3) speech transcription. A total of 235 minutes of audio recordings (160 minutes from 12 children and 75 minutes from 5 teachers), were used to compare system outputs to expert human annotations. WSW2.0 achieves a weighted F1 score of .845, accuracy of .846, and an error-corrected kappa of .672 for speaker classification (child vs. teacher). Transcription quality is moderate to high with word error rates of .119 for teachers and .238 for children. WSW2.0 exhibits relatively high absolute agreement intraclass correlations (ICC) with expert transcriptions for a range of classroom language features. These include teacher and child mean utterance length, lexical diversity, question asking, and responses to questions and other utterances, which show absolute agreement intraclass correlations between .64 and .98. To establish scalability, we apply the framework to an extensive dataset spanning two years and over 1,592 hours of classroom audio recordings, demonstrating the framework's robustness for broad real-world applications. These findings highlight the potential of deep learning and natural language processing techniques to revolutionize educational research by providing accurate measures of key features of preschool classroom speech, ultimately guiding more effective intervention strategies and supporting early childhood language development.

Sentiment Analysis (SA) models harbor inherent social biases that can be harmful in real-world applications. These biases are identified by examining the output of SA models for sentences that only vary in the identity groups of the subjects. Constructing natural, linguistically rich, relevant, and diverse sets of sentences that provide sufficient coverage over the domain is expensive, especially when addressing a wide range of biases: it requires domain experts and/or crowd-sourcing. In this paper, we present a novel bias testing framework, BTC-SAM, which generates high-quality test cases for bias testing in SA models with minimal specification using Large Language Models (LLMs) for the controllable generation of test sentences. Our experiments show that relying on LLMs can provide high linguistic variation and diversity in the test sentences, thereby offering better test coverage compared to base prompting methods even for previously unseen biases.

Language models are increasingly tested on multilingual creativity, demanding culturally grounded, abstract generations. Standard prompting methods often produce repetitive or shallow outputs. We introduce Adaptive Originality Filtering (AOF), a prompting strategy that enforces novelty and cultural fidelity via semantic rejection. To assess quality, we propose RiddleScore, a metric combining novelty, diversity, fluency, and answer alignment. AOF improves Distinct-2 (0.915 in Japanese), reduces Self-BLEU (0.177), and raises RiddleScore (up to +57.1% in Arabic). Human evaluations confirm fluency, creativity, and cultural fit gains. However, improvements vary: Arabic shows greater RiddleScore gains than Distinct-2; Japanese sees similar changes. Though focused on riddles, our method may apply to broader creative tasks. Overall, semantic filtering with composite evaluation offers a lightweight path to culturally rich generation without fine-tuning.

As Large Language Models (LLMs) are increasingly deployed in customer-facing applications, a critical yet underexplored question is how users communicate differently with LLM chatbots compared to human agent. In this study, we present empirical evidence that users adopt distinct communication styles when users interact with chatbots versus human agents. Our analysis reveals significant differences in grammatical fluency, politeness, and lexical diversity in user language between the two settings. These findings suggest that models trained exclusively on human-human interaction data may not adequately accommodate the communication style shift that occurs once an LLM chatbot is deployed. To enhance LLM robustness to post-launch communication style changes, we experimented with two strategies: (1) data augmentation during the post-training phase and (2) inference-time user message reformulation. Our results indicate that models trained on stylistically diverse datasets significantly outperform those trained exclusively on original or stylistically uniform datasets, while inference-time reformulation proved less effective. These insights help us to better adapt our models for improved LLM-user interaction experiences.

As large language models (LLMs) begin to saturate existing benchmarks, automated benchmark creation using LLMs (LLM-as-a-benchmark) has emerged as a scalable alternative to slow and costly human curation. While these generated test sets have to potential to cheaply rank models, we demonstrate a critical flaw. LLM-generated benchmarks systematically favor the model that created the benchmark: they exhibit self-bias on low resource languages to English translation tasks. We show three key findings on automatic benchmarking of LLMs for translation: First, this bias originates from two sources: the generated test data (LLM-as-a-testset) and the evaluation method (LLM-as-an-evaluator), with their combination amplifying the effect. Second, self-bias in LLM-as-a-benchmark is heavily influenced by the model's generation capabilities in the source language. For instance, we observe more pronounced bias in into-English translation, where the model's generation system is developed, than in out-of-English translation tasks. Third, we observe that low diversity in source text is one attribution to self-bias. Our results suggest that improving the diversity of these generated source texts can mitigate some of the observed self-bias. The rapid advancements in Large Language Models (LLMs) have led to an unprecedented saturation of existing, meticulously human-curated benchmarks. This phenomenon exposes two critical, intertwined challenges: traditional benchmark creation is too laborious and expensive to keep pace with rapid model development, and this challenge is compounded by the inherent difficulty of constructing high-quality benchmarks for low-resource languages, even with human labor, which further strains existing benchmark resources.

Fine-grained personas have recently been used for generating 'diverse' synthetic data for pre-training and supervised fine-tuning of Large Language Models (LLMs). In this work, we measure the diversity of persona-driven synthetically generated prompts and responses with a suite of lexical diversity and redundancy metrics. First, we find that synthetic prompts/instructions are significantly less diverse than human-written ones. Next, we sample responses from LLMs of different sizes with fine-grained and coarse persona descriptions to investigate how much fine-grained detail in persona descriptions contribute to generated text diversity. Our results indicate that persona prompting produces higher lexical diversity than prompting without personas, particularly in larger models. In contrast, adding fine-grained persona details yields minimal gains in diversity compared to simply specifying a length cutoff in the prompt.

Diverse language model responses are crucial for creative generation, open-ended tasks, and self-improvement training. We show that common diversity metrics, and even reward models used for preference optimization, systematically bias models toward shorter outputs, limiting expressiveness. To address this, we introduce Diverse, not Short (Diverse-NS), a length-controlled data selection strategy that improves response diversity while maintaining length parity. By generating and filtering preference data that balances diversity, quality, and length, Diverse-NS enables effective training using only 3,000 preference pairs. Applied to LLaMA-3.1-8B and the Olmo-2 family, Diverse-NS substantially enhances lexical and semantic diversity. We show consistent improvement in diversity with minor reduction or gains in response quality on four creative generation tasks: Divergent Associations, Persona Generation, Alternate Uses, and Creative Writing. Surprisingly, experiments with the Olmo-2 model family (7B, and 13B) show that smaller models like Olmo-2-7B can serve as effective "diversity teachers" for larger models. By explicitly addressing length bias, our method efficiently pushes models toward more diverse and expressive outputs.

Ensuring user privacy by synthesizing data from large language models (LLMs) tuned under differential privacy (DP) has become popular recently. However, the impact of DP fine-tuned LLMs on the quality of the language and the utility of the texts they produce has not been investigated. In this work, we tune five LLMs with three corpora under four levels of privacy and assess the length, the grammatical correctness, and the lexical diversity of the text outputs they produce. We also probe the utility of the synthetic outputs in downstream classification tasks such as book genre recognition based on book descriptions and cause of death recognition based on verbal autopsies. The results indicate that LLMs tuned under stronger privacy constrains produce texts that are shorter by at least 77 %, that are less grammatically correct by at least 9 %, and are less diverse by at least 10 % in bi-gram diversity. Furthermore, the accuracy they reach in downstream classification tasks decreases, which might be detrimental to the usefulness of the generated synthetic data.

Generative agent models specifically tailored for smart education are critical, yet remain relatively underdeveloped. A key challenge stems from the inherent complexity of educational contexts: learners are human beings with various cognitive behaviors, and pedagogy is fundamentally centered on personalized human-to-human communication. To address this issue, this paper proposes AgentSME, a unified generative agent framework powered by LLM. Three directional communication modes are considered in the models, namely Solo, Mono, and Echo, reflecting different types of agency autonomy and communicative reciprocity. Accuracy is adopted as the primary evaluation metric, complemented by three diversity indices designed to assess the diversity of reasoning contents. Six widely used LLMs are tested to validate the robustness of communication modes across different model tiers, which are equally divided into base-capacity and high-capacity configurations. The results show that generative agents that employ the Echo communication mode achieve the highest accuracy scores, while DeepSeek exhibits the greatest diversity. This study provides valuable information to improve agent learning capabilities and inspire smart education models.