BACKGROUND: High-throughput sequencing technologies generate massive amounts of FASTQ data comprising nucleotide sequences, quality scores, and read identifiers, necessitating efficient compression to alleviate storage and transmission burdens. Compared to general-purpose compressors, specialized FASTQ compressors achieve higher compression performance by exploiting the inherent redundancy in FASTQ files. However, existing FASTQ-specialized compressors often suffer from limited data applicability and tend to over-optimize either compression ratio or compression speed at the expense of the other. RESULTS: We present zDUR, a reference-free FASTQ compressor designed for efficient and scalable handling of next-generation sequencing data across diverse platforms and sequencing data types. Benchmarking against six reference-free compressors on 15 representative datasets spanning four sequencing data types demonstrates that zDUR achieves a favorable overall balance between compression ratio and speed, with broad applicability across data types. In particular, on single-cell RNA-seq and spatial transcriptomics datasets, zDUR achieves over a tenfold increase in runtime performance while maintaining higher compression ratios than SPRING, one of the state-of-the-art reference-free FASTQ compressors. CONCLUSIONS: zDUR offers a scalable and efficient solution for reference-free FASTQ compression, balancing performance, speed, and usability across diverse datasets.

Artem Shelmanov, Ekaterina Fadeeva, Akim Tsvigun, Ivan Tsvigun, Zhuohan Xie, Igor Kiselev, Nico Daheim, Caiqi Zhang, Artem Vazhentsev, Mrinmaya Sachan, Preslav Nakov, Timothy Baldwin. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Artem Vazhentsev, Ekaterina Fadeeva, Rui Xing, Gleb Kuzmin, Ivan Lazichny, Alexander Panchenko, Preslav Nakov, Timothy Baldwin, Maxim Panov, Artem Shelmanov. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Abstract The rapid proliferation of large language models (LLMs) has stimulated researchers to seek effective and efficient approaches to deal with LLM hallucinations and low-quality outputs. Uncertainty quantification (UQ) is a key element of machine learning applications in dealing with such challenges. However, research to date on UQ for LLMs has been fragmented in terms of techniques and evaluation methodologies. In this work, we address this issue by introducing a novel benchmark that implements a collection of state-of-the-art UQ baselines and offers an environment for controllable and consistent evaluation of novel UQ techniques over various text generation tasks. Our benchmark also supports the assessment of confidence normalization methods in terms of their ability to provide interpretable scores. Using our benchmark, we conduct a large-scale empirical investigation of UQ and normalization techniques across eleven tasks, identifying the most effective approaches.