Deploying large language models (LLMs) in real-world applications is often hindered by strict computational and latency constraints. While dynamic inference offers the flexibility to adjust model behavior based on varying resource budgets, existing methods are frequently limited by hardware inefficiencies or performance degradation. In this paper, we introduce Balcony, a simple yet highly effective framework for depth-based dynamic inference. By freezing the pretrained LLM and inserting additional transformer layers at selected exit points, Balcony maintains the full model’s performance while enabling real-time adaptation to different computational budgets. These additional layers are trained using a straightforward self-distillation loss, aligning the sub-model outputs with those of the full model. This approach requires significantly fewer training tokens and tunable parameters, drastically reducing computational costs compared to prior methods. When applied to the LLaMA3-8B model, using only 0.2% of the original pretraining data, Balcony achieves minimal performance degradation while enabling significant speedups. Remarkably, we show that Balcony outperforms state-of-the-art methods such as Flextron and Layerskip on multiple models at various scales, as well as other leading compression techniques across a variety of benchmarks.

In this paper we present the system for Track A in the SemEval-2024 Task 1: Semantic Textual Relatedness for African and Asian Languages (STR). The proposed system integrates a Siamese Network architecture with pre-trained language models, including BERT, RoBERTa, and the Universal Sentence Encoder (USE). Through rigorous experimentation and analysis, we evaluate the performance of these models across multiple languages. Our findings reveal that the Universal Sentence Encoder excels in capturing semantic similarities, outperforming BERT and RoBERTa in most scenarios. Particularly notable is the USE’s exceptional performance in English and Marathi. These results emphasize the importance of selecting appropriate pre-trained models based on linguistic considerations and task requirements.

Text detoxification is a conditional text generation task aiming to remove offensive content from toxic text. It is highly useful for online forums and social media, where offensive content is frequently encountered. Intuitively, there are diverse ways to detoxify sentences while preserving their meanings, and we can select from detoxified sentences before displaying text to users. Conditional diffusion models are particularly suitable for this task given their demonstrated higher generative diversity than existing conditional text generation models based on language models. Nonetheless, text fluency declines when they are trained with insufficient data, which is the case for this task. In this work, we propose DiffuDetox, a mixed conditional and unconditional diffusion model for text detoxification. The conditional model takes toxic text as the condition and reduces its toxicity, yielding a diverse set of detoxified sentences. The unconditional model is trained to recover the input text, which allows the introduction of additional fluent text for training and thus ensures text fluency. Extensive experimental results and in-depth analysis demonstrate the effectiveness of our proposed DiffuDetox.

Offensive and toxic text on social media platforms can lead to polarization and divisiveness within online communities and hinders constructive dialogue. Text detoxification is a crucial task in natural language processing to ensure the generation of non-toxic and safe text. Text detoxification is a special case of the Text Style Transfer (TST) problem, where an input text is rephrased to an output text that preserves its content while modifying the style (in this case to a more neutral, non-toxic style). State-of-the-art methods for detoxification use supervised training of encoder-decoder models to produce gold-standard outputs with a standard likelihood-based objective. However, it can be hard for these models to deviate from their pretrained auto-encoder identity mapping. While previous methods have used unlikelihood-based losses to penalize input-to-output copying of toxic content, these methods also unfortunately penalize non-toxic content in the input that would be fine to preserve in the output. To address these issues, we introduce a novel contrastive unlikelihood objective (COUNT) that directly contrasts the gold standard rephrasing with the identity input-to-output mapping to effectively isolate and focus learning on non-toxic style transfer. We benchmark COUNT on two parallel datasets, ParaDetox and APPDIA, showing that it achieves significant improvements in jointly combined fluency, content preservation, and detoxification (i.e., the highest “J” score).

Social media has emerged as a key channel for seeking information. Online users spend several hours reading, posting, and searching for news on microblogging platforms daily. However, this could act as a double-edged sword especially when not all information online is reliable. Moreover, the inherently unmoderated nature of social media renders identifying unverified information ever more challenging. Most of the existing approaches for rumor tracking are not scalable because of their dependency on a significant amount of labeled data. In this work, we investigate this problem from different angles. We design an Active-Transfer Learning (ATL) strategy to identify rumors with a limited amount of annotated data. We go beyond that and investigate the impact of leveraging various machine learning approaches in addition to different contextual representations. We discuss the impact of multiple classifiers on a limited amount of annotated data followed by an interactive approach to gradually update the models by adding the least certain samples (LCS) from the pool of unlabeled data. Our proposed Active Learning (AL) strategy achieves faster convergence in terms of the F-score while requiring fewer annotated samples (42% of the whole dataset for the best model).