Anya Belz

Francesco Moramarco, Alex Papadopoulos Korfiatis, Mark Perera et al.

In recent years, machine learning models have rapidly become better at generating clinical consultation notes; yet, there is little work on how to properly evaluate the generated consultation notes to understand the impact they may have on both the clinician using them and the patient's clinical safety. To address this we present an extensive human evaluation study of consultation notes where 5 clinicians (i) listen to 57 mock consultations, (ii) write their own notes, (iii) post-edit a number of automatically generated notes, and (iv) extract all the errors, both quantitative and qualitative. We then carry out a correlation study with 18 automatic quality metrics and the human judgements. We find that a simple, character-based Levenshtein distance metric performs on par if not better than common model-based metrics like BertScore. All our findings and annotations are open-sourced.

Massimiliano Pronesti, Angelo Miculescu, Mohsin Kapdi et al.

Systematic reviews rely on forest plots to synthesise quantitative evidence across biomedical studies, but generating them remains a fragmented and labour-intensive process. Researchers must interpret complex clinical texts, manually extract outcome data from trials, define appropriate interventions and comparators, harmonise inconsistent study designs, and carry out meta-analytic computations-typically using specialised software that demands structured inputs and domain expertise. While recent work has demonstrated that large language models can extract study-level data from unstructured text, no existing system automates the complete pipeline from raw documents to synthesised forest plots. To address this gap, we introduce AutoForest, the first end-to-end system that generates publication-ready forest plots directly from biomedical papers. Given one or more study papers, AutoForest automatically suggests ICO (Intervention, Comparator, Outcome) elements, extracts outcome data, performs statistical synthesis, and renders the final forest plot. We describe the system architecture, user interface and demonstrate its effectiveness on real-world examples through a user study involving clinicians, showing how AutoForest can accelerate evidence synthesis and substantially lower the barrier to conducting meta-analyses.

Anya Belz, Maja Popović, Simon Mille

This paper describes and tests a method for carrying out quantified reproducibility assessment (QRA) that is based on concepts and definitions from metrology. QRA produces a single score estimating the degree of reproducibility of a given system and evaluation measure, on the basis of the scores from, and differences between, different reproductions. We test QRA on 18 system and evaluation measure combinations (involving diverse NLP tasks and types of evaluation), for each of which we have the original results and one to seven reproduction results. The proposed QRA method produces degree-of-reproducibility scores that are comparable across multiple reproductions not only of the same, but of different original studies. We find that the proposed method facilitates insights into causes of variation between reproductions, and allows conclusions to be drawn about what changes to system and/or evaluation design might lead to improved reproducibility.

Mohammed Sabry, Anya Belz

We study distillation for large language models under explicit compute constraints, with the goal of producing student models that are not only cheaper to train, but structurally efficient at inference time. While prior approaches to parameter-efficient distillation, such as LoRA, reduce adaptation cost, they leave the dense backbone unchanged and therefore fail to deliver meaningful inference savings. We propose Budgeted LoRA, a distillation framework that treats model compression as a structured compute allocation problem. Instead of using a fixed student architecture, we introduce a global compute budget that sets the final target fraction of dense computation retained. Under this constraint, the model redistributes capacity across dense and low-rank pathways via (i) module-level dense retention coefficients, (ii) adaptive low-rank allocation, and (iii) post-training compression that selectively removes, approximates, or preserves dense components. This formulation yields a family of students controlled by a single budget dial. Empirically, Budgeted LoRA matches standard LoRA perplexity at a moderate budget with a 1.74x compressed-module speedup; at an aggressive budget it achieves a 4.05x speedup with moderate perplexity degradation, and it preserves higher accuracy on function-style in-context learning probes. These results suggest that, under compute-constrained distillation, retaining behavior is less about matching perplexity or removing more parameters than it is about controlling how dense computation is transferred to low-rank pathways.

Tom Knoll, Francesco Moramarco, Alex Papadopoulos Korfiatis et al.

A growing body of work uses Natural Language Processing (NLP) methods to automatically generate medical notes from audio recordings of doctor-patient consultations. However, there are very few studies on how such systems could be used in clinical practice, how clinicians would adjust to using them, or how system design should be influenced by such considerations. In this paper, we present three rounds of user studies, carried out in the context of developing a medical note generation system. We present, analyse and discuss the participating clinicians' impressions and views of how the system ought to be adapted to be of value to them. Next, we describe a three-week test run of the system in a live telehealth clinical practice. Major findings include (i) the emergence of five different note-taking behaviours; (ii) the importance of the system generating notes in real time during the consultation; and (iii) the identification of a number of clinical use cases that could prove challenging for automatic note generation systems.

Mohammed Sabry, Anya Belz

Recent parameter-efficient finetuning (PEFT) techniques aim to improve over the considerable cost of fully finetuning large pretrained language models (PLM). As different PEFT techniques proliferate, it is becoming difficult to compare them, in particular in terms of (i) the structure and functionality they add to the PLM, (ii) the different types and degrees of efficiency improvements achieved, (iii) performance at different downstream tasks, and (iv) how differences in structure and functionality relate to efficiency and task performance. To facilitate such comparisons, this paper presents a reference architecture which standardises aspects shared by different PEFT techniques, while isolating differences to specific locations and interactions with the standard components. Through this process of standardising and isolating differences, a modular view of PEFT techniques emerges, supporting not only direct comparison of different techniques and their efficiency and task performance, but also systematic exploration of reusability and composability of the different types of finetuned modules. We demonstrate how the reference architecture can be applied to understand properties and relative advantages of PEFT techniques, hence to inform selection of techniques for specific tasks, and design choices for new PEFT techniques.

Aleksandar Savkov, Francesco Moramarco, Alex Papadopoulos Korfiatis et al.

Evaluating automatically generated text is generally hard due to the inherently subjective nature of many aspects of the output quality. This difficulty is compounded in automatic consultation note generation by differing opinions between medical experts both about which patient statements should be included in generated notes and about their respective importance in arriving at a diagnosis. Previous real-world evaluations of note-generation systems saw substantial disagreement between expert evaluators. In this paper we propose a protocol that aims to increase objectivity by grounding evaluations in Consultation Checklists, which are created in a preliminary step and then used as a common point of reference during quality assessment. We observed good levels of inter-annotator agreement in a first evaluation study using the protocol; further, using Consultation Checklists produced in the study as reference for automatic metrics such as ROUGE or BERTScore improves their correlation with human judgements compared to using the original human note.

Michela Lorandi, Anya Belz

LLMs like GPT are great at tasks involving English which dominates in their training data. In this paper, we look at how they cope with tasks involving languages that are severely under-represented in their training data, in the context of data-to-text generation for Irish, Maltese, Welsh and Breton. During the prompt-engineering phase we tested a range of prompt types and formats on GPT-3.5 and~4 with a small sample of example input/output pairs. We then fully evaluated the two most promising prompts in two scenarios: (i) direct generation into the under-resourced language, and (ii) generation into English followed by translation into the under-resourced language. We find that few-shot prompting works better for direct generation into under-resourced languages, but that the difference disappears when pivoting via English. The few-shot + translation system variants were submitted to the WebNLG 2023 shared task where they outperformed competitor systems by substantial margins in all languages on all metrics. We conclude that good performance on under-resourced languages can be achieved out-of-the box with state-of-the-art LLMs. However, our best results (for Welsh) remain well below the lowest ranked English system at WebNLG'20.

Massimiliano Pronesti, Anya Belz, Yufang Hou

Recent work on reinforcement learning with verifiable rewards (RLVR) has shown that large language models (LLMs) can be substantially improved using outcome-level verification signals, such as unit tests for code or exact-match checks for mathematics. In parallel, process supervision has long been explored as a way to shape the intermediate reasoning behaviour of LLMs, but existing approaches rely on neural judges to score chain-of-thought steps, leaving them vulnerable to opacity, bias, and reward hacking. To address this gap, we introduce Verifiable Process Reward Models (VPRMs), a reinforcement-learning framework in which intermediate reasoning steps are checked by deterministic, rule-based verifiers. We apply VPRMs to risk-of-bias assessment for medical evidence synthesis, a domain where guideline-defined criteria and rule-based decision paths enable programmatic verification of reasoning traces. Across multiple datasets, we find that VPRMs generate reasoning that adheres closely to domain rules and achieve substantially higher coherence between step-level decisions and final labels. Results show that VPRMs achieve up to 20% higher F1 than state-of-the-art models and 6.5% higher than verifiable outcome rewards, with substantial gains in evidence grounding and logical coherence.

Michela Lorandi, Anya Belz

Parameter-efficient fine-tuning (PEFT) techniques offer task-specific fine-tuning at a fraction of the cost of full fine-tuning, but require separate fine-tuning for every new task (combination). In this paper, we explore three ways of generalising beyond single-task training/inference: (i) training on combinations of multiple, related datasets; (ii) at inference, composing the weight matrices of separately trained PEFT modules; and (iii) at inference, composing the outputs of separately trained PEFT modules. We test these approaches on three different LLMs, QLoRA as the PEFT technique, and three sets of controlled text generation datasets for sentiment control, topic control, and multi-attribute control. We find that summing PEFT module outputs is a particularly strong composition method, which consistently either outperforms or matches the performance of alternative approaches. This is the case even when comparing against single-task specialised modules on the single-task test set, where three-module output composition achieves an average 2% point performance increase across all models for sentiment control.

Michela Lorandi, Anya Belz

Background: Many different approaches to controlled text generation (CTG) have been proposed over recent years, but it is difficult to get a clear picture of which approach performs best, because different datasets and evaluation methods are used in each case to assess the control achieved. Objectives: Our aim in the work reported in this paper is to develop an approach to evaluation that enables us to comparatively evaluate different CTG systems in a manner that is both informative and fair to the individual systems. Methods: We use a level-playing-field (LPF) approach to comparative evaluation where we (i) generate and process all system outputs in a standardised way, and (ii) apply a shared set of evaluation methods and datasets, selected based on those currently in use, in order to ensure fair evaluation. Results: When re-evaluated in this way, performance results for a representative set of current CTG systems differ substantially from originally reported results, in most cases for the worse. This highlights the importance of a shared standardised way of assessing controlled generation. Conclusions: The discrepancies revealed by LPF evaluation demonstrate the urgent need for standardised, reproducible evaluation practices in CTG. Our results suggest that without such practices, published performance claims may substantially misrepresent true system capabilities.

Michela Lorandi, Anya Belz

The performance of NLP methods for severely under-resourced languages cannot currently hope to match the state of the art in NLP methods for well resourced languages. We explore the extent to which pretrained large language models (LLMs) can bridge this gap, via the example of data-to-text generation for Irish, Welsh, Breton and Maltese. We test LLMs on these under-resourced languages and English, in a range of scenarios. We find that LLMs easily set the state of the art for the under-resourced languages by substantial margins, as measured by both automatic and human evaluations. For all our languages, human evaluation shows on-a-par performance with humans for our best systems, but BLEU scores collapse compared to English, casting doubt on the metric's suitability for evaluating non-task-specific systems. Overall, our results demonstrate the great potential of LLMs to bridge the performance gap for under-resourced languages.

Anya Belz

Reproduction studies reported in NLP provide individual data points which in combination indicate worryingly low levels of reproducibility in the field. Because each reproduction study reports quantitative conclusions based on its own, often not explicitly stated, criteria for reproduction success/failure, the conclusions drawn are hard to interpret, compare, and learn from. In this paper, we present QRA++, a quantitative approach to reproducibility assessment that (i) produces continuous-valued degree of reproducibility assessments at three levels of granularity; (ii) utilises reproducibility measures that are directly comparable across different studies; and (iii) grounds expectations about degree of reproducibility in degree of similarity between experiments. QRA++ enables more informative reproducibility assessments to be conducted, and conclusions to be drawn about what causes reproducibility to be better/poorer. We illustrate this by applying QRA++ to three example sets of comparable experiments, revealing clear evidence that degree of reproducibility depends on similarity of experiment properties, but also system type and evaluation method.

Massimiliano Pronesti, Joao Bettencourt-Silva, Paul Flanagan et al.

Extracting scientific evidence from biomedical studies for clinical research questions (e.g., Does stem cell transplantation improve quality of life in patients with medically refractory Crohn's disease compared to placebo?) is a crucial step in synthesising biomedical evidence. In this paper, we focus on the task of document-level scientific evidence extraction for clinical questions with conflicting evidence. To support this task, we create a dataset called CochraneForest, leveraging forest plots from Cochrane systematic reviews. It comprises 202 annotated forest plots, associated clinical research questions, full texts of studies, and study-specific conclusions. Building on CochraneForest, we propose URCA (Uniform Retrieval Clustered Augmentation), a retrieval-augmented generation framework designed to tackle the unique challenges of evidence extraction. Our experiments show that URCA outperforms the best existing methods by up to 10.3% in F1 score on this task. However, the results also underscore the complexity of CochraneForest, establishing it as a challenging testbed for advancing automated evidence synthesis systems.

Michela Lorandi, Anya Belz

Rerunning a metric-based evaluation should be more straightforward, and results should be closer, than in a human-based evaluation, especially where code and model checkpoints are made available by the original authors. As this report of our efforts to rerun a metric-based evaluation of a set of single-attribute and multiple-attribute controllable text generation (CTG) techniques shows however, such reruns of evaluations do not always produce results that are the same as the original results, and can reveal errors in the reporting of the original work.

Mohammed Sabry, Anya Belz

Does explicitly exercising the induction circuit during pretraining improve in-context learning (ICL), or is natural text sufficient when compute is held constant (iso-FLOPs)? To test whether targeted synthetic data can accelerate induction-head emergence and enhance ICL, we introduce Bi-Induct, a lightweight curriculum that injects forward-copy (Induction), backward-copy (Anti), or a balanced mix into the pretraining stream. We train models from 0.13B to 1B parameters under iso-FLOPs, evaluating (i) few-shot ICL benchmarks, (ii) head-level telemetry, and (iii) held-out language modeling perplexity. Our findings challenge the assumption that early induction circuit activation directly improves ICL. While Bi-Induct accelerates induction-head emergence at small scales, this does not consistently yield stronger generalization. On standard LM benchmarks, Bi-Induct matches natural-only training; on function-style ICL probes, the 1B natural-only performs best. Stress tests (e.g., label permutation, HITS@1 vs. HITS@3, 1 vs. 10 shots) preserve these trends. Telemetry shows larger natural-only models develop broader, earlier induction heads without explicit induction patterns. Anti-induction data fails to elicit meaningful activation. Perplexity penalties from synthetic data shrink with scale, suggesting larger models can absorb non-natural patterns with minimal cost. Crucially, ablating the top 2% of induction heads degrades ICL more than random ablations, especially for natural-only models, indicating more centralized, load-bearing circuits. Bi-Induct variants exhibit more redundant induction activity, implying different circuit utilization. Overall, inducing activation is not sufficient: ICL gains depend on these circuits becoming functionally necessary. These results underscore mechanism-aware pretraining diagnostics and data mixtures that foster load-bearing, not merely present, structure.

Anya Belz, Simon Mille, Craig Thomson

Prior work has shown that two NLP evaluation experiments that report results for the same quality criterion name (e.g. Fluency) do not necessarily evaluate the same aspect of quality, and the comparability implied by the name can be misleading. Not knowing when two evaluations are comparable in this sense means we currently lack the ability to draw reliable conclusions about system quality on the basis of multiple, independently conducted evaluations. This in turn hampers the ability of the field to progress scientifically as a whole, a pervasive issue in NLP since its beginning (Sparck Jones, 1981). It is hard to see how the issue of unclear comparability can be fully addressed other than by the creation of a standard set of quality criterion names and definitions that the several hundred quality criterion names actually in use in the field can be mapped to, and grounded in. Taking a strictly descriptive approach, the QCET Quality Criteria for Evaluation Taxonomy derives a standard set of quality criterion names and definitions from three surveys of evaluations reported in NLP, and structures them into a hierarchy where each parent node captures common aspects of its child nodes. We present QCET and the resources it consists of, and discuss its three main uses in (i) establishing comparability of existing evaluations, (ii) guiding the design of new evaluations, and (iii) assessing regulatory compliance.

Massimiliano Pronesti, Michela Lorandi, Paul Flanagan et al.

Systematic reviews in medicine play a critical role in evidence-based decision-making by aggregating findings from multiple studies. A central bottleneck in automating this process is extracting numeric evidence and determining study-level conclusions for specific outcomes and comparisons. Prior work has framed this problem as a textual inference task by retrieving relevant content fragments and inferring conclusions from them. However, such approaches often rely on shallow textual cues and fail to capture the underlying numeric reasoning behind expert assessments. In this work, we conceptualise the problem as one of quantitative reasoning. Rather than inferring conclusions from surface text, we extract structured numerical evidence (e.g., event counts or standard deviations) and apply domain knowledge informed logic to derive outcome-specific conclusions. We develop a numeric reasoning system composed of a numeric data extraction model and an effect estimate component, enabling more accurate and interpretable inference aligned with the domain expert principles. We train the numeric data extraction model using different strategies, including supervised fine-tuning (SFT) and reinforcement learning (RL) with a new value reward model. When evaluated on the CochraneForest benchmark, our best-performing approach -- using RL to train a small-scale number extraction model -- yields up to a 21% absolute improvement in F1 score over retrieval-based systems and outperforms general-purpose LLMs of over 400B parameters by up to 9%. Our results demonstrate the promise of reasoning-driven approaches for automating systematic evidence synthesis.

Anya Belz, Craig Thomson

This paper presents version 3.0 of the Human Evaluation Datasheet (HEDS). This update is the result of our experience using HEDS in the context of numerous recent human evaluation experiments, including reproduction studies, and of feedback received. Our main overall goal was to improve clarity, and to enable users to complete the datasheet more consistently and comparably. The HEDS 3.0 package consists of the digital data sheet, documentation, and code for exporting completed data sheets as latex files, all available from the HEDS GitHub.

Mohammed Sabry, Anya Belz

As the cost of training ever larger language models has grown, so has the interest in reusing previously learnt knowledge. Transfer learning methods have shown how reusing non-task-specific knowledge can help in subsequent task-specific learning. In this paper, we investigate the inverse: porting whole functional modules that encode task-specific knowledge from one model to another. We designed a study comprising 1,440 training/testing runs to test the portability of modules trained by parameter-efficient finetuning (PEFT) techniques, using sentiment analysis as an example task. We test portability in a wide range of scenarios, involving different PEFT techniques and different pretrained host models, among other dimensions. We compare the performance of ported modules with that of equivalent modules trained (i) from scratch, and (ii) from parameters sampled from the same distribution as the ported module. We find that the ported modules far outperform the two alternatives tested, but that there are interesting performance differences between the four PEFT techniques. We conclude that task-specific knowledge in the form of structurally modular sets of parameters as produced by PEFT techniques is highly portable, but that degree of success depends on type of PEFT and on differences between originating and receiving pretrained models.

Anya Belz, Craig Thomson, Ehud Reiter et al.

We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13\% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

Anya Belz

Reproducibility has become an intensely debated topic in NLP and ML over recent years, but no commonly accepted way of assessing reproducibility, let alone quantifying it, has so far emerged. The assumption has been that wider scientific reproducibility terminology and definitions are not applicable to NLP/ML, with the result that many different terms and definitions have been proposed, some diametrically opposed. In this paper, we test this assumption, by taking the standard terminology and definitions from metrology and applying them directly to NLP/ML. We find that we are able to straightforwardly derive a practical framework for assessing reproducibility which has the desirable property of yielding a quantified degree of reproducibility that is comparable across different reproduction studies.

Anastasia Shimorina, Anya Belz

This paper introduces the Human Evaluation Datasheet, a template for recording the details of individual human evaluation experiments in Natural Language Processing (NLP). Originally taking inspiration from seminal papers by Bender and Friedman (2018), Mitchell et al. (2019), and Gebru et al. (2020), the Human Evaluation Datasheet is intended to facilitate the recording of properties of human evaluations in sufficient detail, and with sufficient standardisation, to support comparability, meta-evaluation, and reproducibility tests.

Anya Belz, Shubham Agarwal, Anastasia Shimorina et al.

Against the background of what has been termed a reproducibility crisis in science, the NLP field is becoming increasingly interested in, and conscientious about, the reproducibility of its results. The past few years have seen an impressive range of new initiatives, events and active research in the area. However, the field is far from reaching a consensus about how reproducibility should be defined, measured and addressed, with diversity of views currently increasing rather than converging. With this focused contribution, we aim to provide a wide-angle, and as near as possible complete, snapshot of current work on reproducibility in NLP, delineating differences and similarities, and providing pointers to common denominators.