Revisiting Prompt Sensitivity in Large Language Models for Text Classification: The Role of Prompt Underspecification

Large language models (LLMs) are widely used as zero-shot and few-shot classifiers, where task behaviour is largely controlled through prompting. A growing number of works have observed that LLMs are sensitive to prompt variations, with small changes leading to large changes in performance. However, in many cases, the investigation of sensitivity is performed using underspecified prompts that provide minimal task instructions and weakly constrain the model's output space. In this work, we argue that a significant portion of the observed prompt sensitivity can be attributed to prompt underspecification. We systematically study and compare the sensitivity of underspecified prompts and prompts that provide specific instructions. Utilising performance analysis, logit analysis, and linear probing, we find that underspecified prompts exhibit higher performance variance and lower logit values for relevant tokens, while instruction-prompts suffer less from such problems. However, linear probing analysis suggests that the effects of prompt underspecification have only a marginal impact on the internal LLM representations, instead emerging in the final layers. Overall, our findings highlight the need for more rigour when investigating and mitigating prompt sensitivity.

Beyond the Checkbox: Strengthening DSA Compliance Through Social Media Algorithmic Auditing

Algorithms of online platforms are required under the Digital Services Act (DSA) to comply with specific obligations concerning algorithmic transparency, user protection and privacy. To verify compliance with these requirements, DSA mandates platforms to undergo independent audits. Little is known about current auditing practices and their effectiveness in ensuring such compliance. To this end, we bridge regulatory and technical perspectives by critically examining selected audit reports across three critical algorithmic-related provisions: restrictions on profiling minors, transparency in recommender systems, and limitations on targeted advertising using sensitive data. Our analysis shows significant inconsistencies in methodologies and lack of technical depth when evaluating AI-powered systems. To enhance the depth, scale, and independence of compliance assessments, we propose to employ algorithmic auditing -- a process of behavioural assessment of AI algorithms by means of simulating user behaviour, observing algorithm responses and analysing them for audited phenomena.

Better as Generators Than Classifiers: Leveraging LLMs and Synthetic Data for Low-Resource Multilingual Classification

Large Language Models (LLMs) have demonstrated remarkable multilingual capabilities, making them promising tools in both high- and low-resource languages. One particularly valuable use case is generating synthetic samples that can be used to train smaller models in low-resource scenarios where human-labelled data is scarce. In this work, we investigate whether these synthetic data generation capabilities can serve as a form of distillation, producing smaller models that perform on par with or even better than massive LLMs across languages and tasks. To this end, we use a state-of-the-art multilingual LLM to generate synthetic datasets covering 11 languages and 4 classification tasks. These datasets are then used to train smaller models via fine-tuning or instruction tuning, or as synthetic in-context examples for compact LLMs. Our experiments show that even small amounts of synthetic data enable smaller models to outperform the large generator itself, particularly in low-resource languages. Overall, the results suggest that LLMs are best utilised as generators (teachers) rather than classifiers, producing data that empowers smaller and more efficient multilingual models.

Revisiting Algorithmic Audits of TikTok: Poor Reproducibility and Short-term Validity of Findings

Social media platforms are constantly shifting towards algorithmically curated content based on implicit or explicit user feedback. Regulators, as well as researchers, are calling for systematic social media algorithmic audits as this shift leads to enclosing users in filter bubbles and leading them to more problematic content. An important aspect of such audits is the reproducibility and generalisability of their findings, as it allows to draw verifiable conclusions and audit potential changes in algorithms over time. In this work, we study the reproducibility of the existing sockpuppeting audits of TikTok recommender systems, and the generalizability of their findings. In our efforts to reproduce the previous works, we find multiple challenges stemming from social media platform changes and content evolution, but also the research works themselves. These drawbacks limit the audit reproducibility and require an extensive effort altogether with inevitable adjustments to the auditing methodology. Our experiments also reveal that these one-shot audit findings often hold only in the short term, implying that the reproducibility and generalizability of the audits heavily depend on the methodological choices and the state of algorithms and content on the platform. This highlights the importance of reproducible audits that allow us to determine how the situation changes in time.

Use Random Selection for Now: Investigation of Few-Shot Selection Strategies in LLM-based Text Augmentation for Classification

The generative large language models (LLMs) are increasingly used for data augmentation tasks, where text samples are paraphrased (or generated anew) and then used for classifier fine-tuning. Existing works on augmentation leverage the few-shot scenarios, where samples are given to LLMs as part of prompts, leading to better augmentations. Yet, the samples are mostly selected randomly and a comprehensive overview of the effects of other (more ``informed'') sample selection strategies is lacking. In this work, we compare sample selection strategies existing in few-shot learning literature and investigate their effects in LLM-based textual augmentation. We evaluate this on in-distribution and out-of-distribution classifier performance. Results indicate, that while some ``informed'' selection strategies increase the performance of models, especially for out-of-distribution data, it happens only seldom and with marginal performance increases. Unless further advances are made, a default of random sample selection remains a good option for augmentation practitioners.

Fighting Randomness with Randomness: Mitigating Optimisation Instability of Fine-Tuning using Delayed Ensemble and Noisy Interpolation

While fine-tuning of pre-trained language models generally helps to overcome the lack of labelled training samples, it also displays model performance instability. This instability mainly originates from randomness in initialisation or data shuffling. To address this, researchers either modify the training process or augment the available samples, which typically results in increased computational costs. We propose a new mitigation strategy, called Delayed Ensemble with Noisy Interpolation (DENI), that leverages the strengths of ensembling, noise regularisation and model interpolation, while retaining computational efficiency. We compare DENI with 9 representative mitigation strategies across 3 models, 4 tuning strategies and 7 text classification datasets. We show that: 1) DENI outperforms the best performing mitigation strategy (Ensemble), while using only a fraction of its cost; 2) the mitigation strategies are beneficial for parameter-efficient fine-tuning (PEFT) methods, outperforming full fine-tuning in specific cases; and 3) combining DENI with data augmentation often leads to even more effective instability mitigation.

Fine-Tuning, Prompting, In-Context Learning and Instruction-Tuning: How Many Labelled Samples Do We Need?

When solving a task with limited labelled data, researchers can either use a general large language model without further update, or use the few examples to tune a specialised smaller model. When enough labels are available, the specialised models outperform the general ones on many NLP tasks. In this work, we aim to investigate how many labelled samples are required for the specialised models to achieve this superior performance, while taking the results variance into consideration. Observing the behaviour of prompting, in-context learning, fine-tuning and instruction-tuning, identifying their break-even points when increasing number of labelled training samples across three tasks of varying complexity, we find that the specialised models often need only few samples ($100-1000$) to be on par or better than the general ones. At the same time, the amount of required labelled data strongly depends on the task complexity and results variance.

On Sensitivity of Learning with Limited Labelled Data to the Effects of Randomness: Impact of Interactions and Systematic Choices

While learning with limited labelled data can improve performance when the labels are lacking, it is also sensitive to the effects of uncontrolled randomness introduced by so-called randomness factors (e.g., varying order of data). We propose a method to systematically investigate the effects of randomness factors while taking the interactions between them into consideration. To measure the true effects of an individual randomness factor, our method mitigates the effects of other factors and observes how the performance varies across multiple runs. Applying our method to multiple randomness factors across in-context learning and fine-tuning approaches on 7 representative text classification tasks and meta-learning on 3 tasks, we show that: 1) disregarding interactions between randomness factors in existing works caused inconsistent findings due to incorrect attribution of the effects of randomness factors, such as disproving the consistent sensitivity of in-context learning to sample order even with random sample selection; and 2) besides mutual interactions, the effects of randomness factors, especially sample order, are also dependent on more systematic choices unexplored in existing works, such as number of classes, samples per class or choice of prompt format.

Automatic Combination of Sample Selection Strategies for Few-Shot Learning

In few-shot learning, such as meta-learning, few-shot fine-tuning or in-context learning, the limited number of samples used to train a model have a significant impact on the overall success. Although a large number of sample selection strategies exist, their impact on the performance of few-shot learning is not extensively known, as most of them have been so far evaluated in typical supervised settings only. In this paper, we thoroughly investigate the impact of 20 sample selection strategies on the performance of 5 few-shot learning approaches over 8 image and 6 text datasets. In addition, we propose a new method for automatic combination of sample selection strategies (ACSESS) that leverages the strengths and complementary information of the individual strategies. The experimental results show that our method consistently outperforms the individual selection strategies, as well as the recently proposed method for selecting support examples for in-context learning. We also show a strong modality, dataset and approach dependence for the majority of strategies as well as their dependence on the number of shots - demonstrating that the sample selection strategies play a significant role for lower number of shots, but regresses to random selection at higher number of shots.

Effects of diversity incentives on sample diversity and downstream model performance in LLM-based text augmentation

The latest generative large language models (LLMs) have found their application in data augmentation tasks, where small numbers of text samples are LLM-paraphrased and then used to fine-tune the model. However, more research is needed to assess how different prompts, seed data selection strategies, filtering methods, or model settings affect the quality of paraphrased data (and downstream models). In this study, we investigate three text diversity incentive methods well established in crowdsourcing: taboo words, hints by previous outlier solutions, and chaining on previous outlier solutions. Using these incentive methods as part of instructions to LLMs augmenting text datasets, we measure their effects on generated texts' lexical diversity and downstream model performance. We compare the effects over 5 different LLMs and 6 datasets. We show that diversity is most increased by taboo words, while downstream model performance is highest when previously created paraphrases are used as hints.