Research on cross-dialectal transfer from a standard to a non-standard dialect variety has typically focused on text data. However, dialects are primarily spoken, and non-standard spellings are known to cause issues in text processing. We compare standard-to-dialect transfer in three settings: text models, speech models, and cascaded systems where speech first gets automatically transcribed and then further processed by a text model. In our experiments, we focus on German and multiple German dialects in the context of written and spoken intent and topic classification. To that end, we release the first dialectal audio intent classification dataset. We find that the speech-only setup provides the best results on the dialect data while the text-only setup works best on the standard data. While the cascaded systems lag behind the text-only models for German, they perform relatively well on the dialectal data if the transcription system generates normalized, standard-like output.

We study LLM routing, the problem of selecting the best model for each query while balancing user satisfaction, model expertise, and inference cost. We formulate routing as contextual dueling bandits, learning from pairwise preference feedback rather than absolute scores, thereby yielding label-efficient and dynamic adaptation. Building on this formulation, we introduce Category-Calibrated Fine-Tuning (CCFT), a representation-learning method that derives model embeddings from offline data using contrastive fine-tuning with categorical weighting. These embeddings enable the practical instantiation of Feel-Good Thompson Sampling for Contextual Dueling Bandits (FGTS.CDB), a theoretically grounded posterior-sampling algorithm. We propose four variants of the categorical weighting that explicitly integrate model quality and cost, and we empirically evaluate the proposed methods on the RouterBench and MixInstruct datasets. Across both benchmarks, our methods achieve lower cumulative regret and faster convergence, with better robustness and performance-cost balance than strong baselines built with a general-purpose OpenAI embedding model. The potential of large language models (LLMs) is so great that they have become a necessary part of daily life, with applications ranging from office assistance and fashion/dining suggestions to entertainment. LLM routing refers to a problem of dynamically selecting the most suitable LLM from a set of candidates for each query in a sequence of questions. Before the emergence of a universally dominant and affordable foundation model, routing is important because the choice of LLM must align with user traits, model expertise, and cost. To balance these three key factors, cascading algorithms such as FrugalGPT (Chen et al., 2024) and AutoMix (Aggarwal et al., 2024) were first proposed. The idea is to query a cheaper model first and advance the query to a more expensive one if the current response is unlikely to meet the user's expectation. A drawback of cascading is the accumulated cost and latency caused by calling multiple LLM candidates to generate the final response.

The emergence of large language models (LLMs) has demonstrated that systems trained solely on text can acquire extensive world knowledge, develop reasoning capabilities, and internalize abstract semantic concepts--showcasing properties that can be associated with general intelligence. This raises an intriguing question: Do such concepts emerge in models trained on other modalities, such as speech? Furthermore, when models are trained jointly on multiple modalities: Do they develop a richer, more structured semantic understanding? To explore this, we analyze the conceptual structures learned by speech and textual models both individually and jointly. We employ Latent Concept Analysis, an unsupervised method for uncovering and interpreting latent representations in neural networks, to examine how semantic abstractions form across modalities. For reproducibility we made scripts and other resources available to the community.

The increasing prevalence of synthetic data in training loops has raised concerns about model collapse, where generative models degrade when trained on their own outputs. While prior work focuses on this self-consuming process, we study an underexplored yet prevalent phenomenon: co-evolving generative models that shape each other's training through iterative feedback. This is common in multimodal AI ecosystems, such as social media platforms, where text models generate captions that guide image models, and the resulting images influence the future adaptation of the text model. We take a first step by analyzing such a system, modeling the text model as a multinomial distribution and the image model as a conditional multi-dimensional Gaussian distribution. Our analysis uncovers three key results. First, when one model remains fixed, the other collapses: a frozen image model causes the text model to lose diversity, while a frozen text model leads to an exponential contraction of image diversity, though fidelity remains bounded. Second, in fully interactive systems, mutual reinforcement accelerates collapse, with image contraction amplifying text homogenization and vice versa, leading to a Matthew effect where dominant texts sustain higher image diversity while rarer texts collapse faster. Third, we analyze stabilization strategies implicitly introduced by real-world external influences. Random corpus injections for text models and user-content injections for image models prevent collapse while preserving both diversity and fidelity. Our theoretical findings are further validated through experiments.

The paper was clear and well written. The data set and the evaluation that was conducted could be useful to the community. However, the paper unfairly characterizes or omits some previous work, and was not clear enough about the limitations and biases of their evaluation strategy. These points detract from a paper that otherwise makes an interesting contribution. First, there is an implied criticism of WordSim-353 and MEN at the bottom of page 2 that they only contain similarity judgments at the word level. However, there is a large amount of work on learning phrase and sentence-level embeddings in the recently literature that overcome these issues (see representative work by Mirella Lapata, Marco Baroni, Stephen Clarke, Richard Socher, among many others), which the paper does not mention.

The use of transformer-based models is growing rapidly throughout society. With this growth, it is important to understand how they work, and in particular, how the attention mechanisms represent concepts. Though there are many interpretability methods, many look at models through their neuronal activations, which are poorly understood. We describe different lenses through which to view neuron activations, and investigate the effectiveness in language models and vision transformers through various methods of neural ablation: zero ablation, mean ablation, activation resampling, and a novel approach we term 'peak ablation'. Through experimental analysis, we find that in different regimes and models, each method can offer the lowest degradation of model performance compared to other methods, with resampling usually causing the most significant performance deterioration. We make our code available at https://github.com/nickypro/investigating-ablation.

Cognitive decline is a natural process that occurs as individuals age. Early diagnosis of anomalous decline is crucial for initiating professional treatment that can enhance the quality of life of those affected. To address this issue, we propose a multimodal model capable of predicting Mild Cognitive Impairment and cognitive scores. The TAUKADIAL dataset is used to conduct the evaluation, which comprises audio recordings of clinical interviews. The proposed model demonstrates the ability to transcribe and differentiate between languages used in the interviews. Subsequently, the model extracts audio and text features, combining them into a multimodal architecture to achieve robust and generalized results. Our approach involves in-depth research to implement various features obtained from the proposed modalities.

Generating images from text has become easier because of the scaling of diffusion models and advancements in the field of vision and language. These models are trained using vast amounts of data from the Internet. Hence, they often contain undesirable content such as copyrighted material. As it is challenging to remove such data and retrain the models, methods for erasing specific concepts from pre-trained models have been investigated. We propose a novel concept-erasure method that updates the text encoder using few-shot unlearning in which a few real images are used. The discussion regarding the generated images after erasing a concept has been lacking. While there are methods for specifying the transition destination for concepts, the validity of the specified concepts is unclear. Our method implicitly achieves this by transitioning to the latent concepts inherent in the model or the images. Our method can erase a concept within 10 s, making concept erasure more accessible than ever before. Implicitly transitioning to related concepts leads to more natural concept erasure. We applied the proposed method to various concepts and confirmed that concept erasure can be achieved tens to hundreds of times faster than with current methods. By varying the parameters to be updated, we obtained results suggesting that, like previous research, knowledge is primarily accumulated in the feed-forward networks of the text encoder.