Instruction-tuned large language models (IT-LLMs) exhibit strong zero-shot reasoning, yet their ability to execute simple, self-contained instructions remains underexplored, despite this being foundational to complex instruction-following. We evaluate 20 IT-LLMs on modified MMLU and MMLU-Pro benchmarks, by systematically varying the format of option labels (alphabetic, numeric, Roman) while keeping their meaning identical under four paradigms, namely: (1) With explicit instructions, label changes cause large performance shifts (e.g., -30.45\% for Roman vs. numeric), revealing instruction-format bias. (2) Without instructions, performance drops further (up to -10.84\%) and label sensitivity intensifies, underscoring the role of explicit guidance. (3) When option contents are removed, models fail random-choice baselines except with numeric labels, suggesting weak adherence to atomic directives. (4) Three-shot exemplars yield no significant gains in robustness or fidelity, and generation analyses show persistent label errors, especially for non-numeric formats. Across model sizes, larger LLMs achieve higher accuracy but remain inconsistent in instruction adherence. These results expose the insufficiencies of current instruction-tuning paradigms and highlight the need for evaluation methods and training strategies that explicitly target atomic instruction-following.

Human emotion is expressed in many communication modalities and media formats and so their computational study is equally diversified into natural language processing, audio signal analysis, computer vision, etc. Similarly, the large variety of representation formats used in previous research to describe emotions (polarity scales, basic emotion categories, dimensional approaches, appraisal theory, etc.) have led to an ever proliferating diversity of datasets, predictive models, and software tools for emotion analysis. Because of these two distinct types of heterogeneity, at the expressional and representational level, there is a dire need to unify previous work on increasingly diverging data and label types. This article presents such a unifying computational model. We propose a training procedure that learns a shared latent representation for emotions, so-called emotion embeddings, independent of different natural languages, communication modalities, media or representation label formats, and even disparate model architectures. Experiments on a wide range of heterogeneous affective datasets indicate that this approach yields the desired interoperability for the sake of reusability, interpretability and flexibility, without penalizing prediction quality. Code and data are archived under https://doi.org/10.5281/zenodo.7405327 .

We present EmoCoder, a modular encoder-decoder architecture that generalizes emotion analysis over different tasks (sentence-level, word-level, label-to-label mapping), domains (natural languages and their registers), and label formats (e.g., polarity classes, basic emotions, and affective dimensions). Experiments on 14 datasets indicate that EmoCoder learns an interpretable language-independent representation of emotions, allows seamless absorption of state-of-the-art models, and maintains strong prediction quality, even when tested on unseen combinations of domains and label formats.