Roi Blanco

Natalia Tarasova, Enrique Balp-Straffon, Aleksei Iancheruk et al.

Building infrastructure-as-code (IaC) in cloud computing is a critical task, underpinning the reliability, scalability, and security of modern software systems. Despite the remarkable progress of large language models (LLMs) in software engineering -- demonstrated across many dedicated benchmarks -- their capabilities in developing IaC remain underexplored. Unlike existing IaC benchmarks that predominantly center on declarative paradigms such as Terraform and involve generating entire codebases from scratch, our benchmark reflects the incremental code edits common in enterprise development with imperative tools like the AWS CDK. We present SWE-InfraBench, a diverse evaluation dataset sourced from dozens of real-world IaC codebases that challenge LLMs to perform realistic code modifications in AWS CDK repositories. Each example requires models to implement changes to existing codebases based on natural language instructions, with success determined by passing provided test cases. These tasks demand sophisticated reasoning about cloud resource dependencies and implementation patterns beyond conventional code generation challenges. Our evaluation results reveal significant limitations in current LLMs showing that even state-of-the-art systems struggle with many tasks -- the best model, Sonnet 3.7, succeeds in only 34\% of cases, while specialized reasoning models like DeepSeek R1 achieve just 24% success. The SWE-InfraBench dataset is available at: https://www.kaggle.com/datasets/64e59070fd51c0278560b01eb5dc4f3c447d5268cdabe5a350d2969e4413fea5

Laura Aina, Nikos Voskarides, Roi Blanco

Pre-trained language models (LMs) obtain state-of-the-art performance when adapted to text classification tasks. However, when using such models in real-world applications, efficiency considerations are paramount. In this paper, we study how different training procedures that adapt LMs to text classification perform, as we vary model and train set size. More specifically, we compare standard fine-tuning, prompting, and knowledge distillation (KD) when the teacher was trained with either fine-tuning or prompting. Our findings suggest that even though fine-tuning and prompting work well to train large LMs on large train sets, there are more efficient alternatives that can reduce compute or data cost. Interestingly, we find that prompting combined with KD can reduce compute and data cost at the same time.

Ege Beyazit, KL Navaneet, Prashant Mathur et al.

Black-box Large Language Models (LLMs) provide practical and accessible alternatives to other machine learning methods, as they require minimal labeled data and machine learning expertise to develop solutions for various decision making problems. However, for applications that need operating with constraints on specific metrics (e.g., precision $\geq$ 95%), decision making with black-box LLMs remains unfavorable, due to their low numerical output cardinalities. This results in limited control over their operating points, preventing fine-grained adjustment of their decision making behavior. In this paper, we study using black-box LLMs as classifiers, focusing on efficiently improving their operational granularity without performance loss. Specifically, we first investigate the reasons behind their low-cardinality numerical outputs and show that they are biased towards generating rounded but informative verbalized probabilities. Then, we experiment with standard prompt engineering, uncertainty estimation and confidence elicitation techniques, and observe that they do not effectively improve operational granularity without sacrificing performance or increasing inference cost. Finally, we propose efficient approaches to significantly increase the number and diversity of available operating points. Our proposed approaches provide finer-grained operating points and achieve comparable to or better performance than the benchmark methods across 11 datasets and 3 LLMs.

Seraphina Goldfarb-Tarrant, Adam Lopez, Roi Blanco et al.

Sentiment analysis (SA) systems are used in many products and hundreds of languages. Gender and racial biases are well-studied in English SA systems, but understudied in other languages, with few resources for such studies. To remedy this, we build a counterfactual evaluation corpus for gender and racial/migrant bias in four languages. We demonstrate its usefulness by answering a simple but important question that an engineer might need to answer when deploying a system: What biases do systems import from pre-trained models when compared to a baseline with no pre-training? Our evaluation corpus, by virtue of being counterfactual, not only reveals which models have less bias, but also pinpoints changes in model bias behaviour, which enables more targeted mitigation strategies. We release our code and evaluation corpora to facilitate future research.

Stefanos Angelidis, Lea Frermann, Diego Marcheggiani et al.

We present a system for answering questions based on the full text of books (BookQA), which first selects book passages given a question at hand, and then uses a memory network to reason and predict an answer. To improve generalization, we pretrain our memory network using artificial questions generated from book sentences. We experiment with the recently published NarrativeQA corpus, on the subset of Who questions, which expect book characters as answers. We experimentally show that BERT-based retrieval and pretraining improve over baseline results significantly. At the same time, we confirm that NarrativeQA is a highly challenging data set, and that there is need for novel research in order to achieve high-precision BookQA results. We analyze some of the bottlenecks of the current approach, and we argue that more research is needed on text representation, retrieval of relevant passages, and reasoning, including commonsense knowledge.

Joel Mackenzie, J. Shane Culpepper, Roi Blanco et al.

Scalable web search systems typically employ multi-stage retrieval architectures, where an initial stage generates a set of candidate documents that are then pruned and re-ranked. Since subsequent stages typically exploit a multitude of features of varying costs using machine-learned models, reducing the number of documents that are considered at each stage improves latency. In this work, we propose and validate a unified framework that can be used to predict a wide range of performance-sensitive parameters which minimize effectiveness loss, while simultaneously minimizing query latency, across all stages of a multi-stage search architecture. Furthermore, our framework can be easily applied in large-scale IR systems, can be trained without explicitly requiring relevance judgments, and can target a variety of different efficiency-effectiveness trade-offs, making it well suited to a wide range of search scenarios. Our results show that we can reliably predict a number of different parameters on a per-query basis, while simultaneously detecting and minimizing the likelihood of tail-latency queries that exceed a pre-specified performance budget. As a proof of concept, we use the prediction framework to help alleviate the problem of tail-latency queries in early stage retrieval. On the standard ClueWeb09B collection and 31k queries, we show that our new hybrid system can reliably achieve a maximum query time of 200 ms with a 99.99% response time guarantee without a significant loss in overall effectiveness. The solutions presented are practical, and can easily be used in large-scale distributed search engine deployments with a small amount of additional overhead.

Christina Lioma, Roi Blanco

Automatic language processing tools typically assign to terms so-called weights corresponding to the contribution of terms to information content. Traditionally, term weights are computed from lexical statistics, e.g., term frequencies. We propose a new type of term weight that is computed from part of speech (POS) n-gram statistics. The proposed POS-based term weight represents how informative a term is in general, based on the POS contexts in which it generally occurs in language. We suggest five different computations of POS-based term weights by extending existing statistical approximations of term information measures. We apply these POS-based term weights to information retrieval, by integrating them into the model that matches documents to queries. Experiments with two TREC collections and 300 queries, using TF-IDF & BM25 as baselines, show that integrating our POS-based term weights to retrieval always leads to gains (up to +33.7% from the baseline). Additional experiments with a different retrieval model as baseline (Language Model with Dirichlet priors smoothing) and our best performing POS-based term weight, show retrieval gains always and consistently across the whole smoothing range of the baseline.