Nick Ferguson

Nick Ferguson, Josh Pennington, Narek Beghian et al.

Unstructured documents like PDFs contain valuable structured information, but downstream systems require this data in reliable, standardized formats. LLMs are increasingly deployed to automate this extraction, making accuracy and reliability paramount. However, progress is bottlenecked by two gaps. First, no end-to-end benchmark evaluates PDF-to-JSON extraction under enterprise-scale schema breadth. Second, no principled methodology captures the semantics of nested extraction, where fields demand different notions of correctness (exact match for identifiers, tolerance for quantities, semantic equivalence for names), arrays require alignment, and omission must be distinguished from hallucination. We address both gaps with ExtractBench, an open-source benchmark and evaluation framework for PDF-to-JSON structured extraction. The benchmark pairs 35 PDF documents with JSON Schemas and human-annotated gold labels across economically valuable domains, yielding 12,867 evaluatable fields spanning schema complexities from tens to hundreds of fields. The evaluation framework treats the schema as an executable specification: each field declares its scoring metric. Baseline evaluations reveal that frontier models (GPT-5/5.2, Gemini-3 Flash/Pro, Claude 4.5 Opus/Sonnet) remain unreliable on realistic schemas. Performance degrades sharply with schema breadth, culminating in 0% valid output on a 369-field financial reporting schema across all tested models. We release ExtractBench at https://github.com/ContextualAI/extract-bench.

Balint Gyevnar, Nick Ferguson, Burkhard Schafer

The European Union has proposed the Artificial Intelligence Act which introduces detailed requirements of transparency for AI systems. Many of these requirements can be addressed by the field of explainable AI (XAI), however, there is a fundamental difference between XAI and the Act regarding what transparency is. The Act views transparency as a means that supports wider values, such as accountability, human rights, and sustainable innovation. In contrast, XAI views transparency narrowly as an end in itself, focusing on explaining complex algorithmic properties without considering the socio-technical context. We call this difference the ``transparency gap''. Failing to address the transparency gap, XAI risks leaving a range of transparency issues unaddressed. To begin to bridge this gap, we overview and clarify the terminology of how XAI and European regulation -- the Act and the related General Data Protection Regulation (GDPR) -- view basic definitions of transparency. By comparing the disparate views of XAI and regulation, we arrive at four axes where practical work could bridge the transparency gap: defining the scope of transparency, clarifying the legal status of XAI, addressing issues with conformity assessment, and building explainability for datasets.

Nick Ferguson, Liane Guillou, Kwabena Nuamah et al.

We present a study into the ability of paraphrase generation methods to increase the variety of natural language questions that the FRANK Question Answering system can answer. We first evaluate paraphrase generation methods on the LC-QuAD 2.0 dataset using both automatic metrics and human judgement, and discuss their correlation. Error analysis on the dataset is also performed using both automatic and manual approaches, and we discuss how paraphrase generation and evaluation is affected by data points which contain error. We then simulate an implementation of the best performing paraphrase generation method (an English-French backtranslation) into FRANK in order to test our original hypothesis, using a small challenge dataset. Our two main conclusions are that cleaning of LC-QuAD 2.0 is required as the errors present can affect evaluation; and that, due to limitations of FRANK's parser, paraphrase generation is not a method which we can rely on to improve the variety of natural language questions that FRANK can answer.

Nick Ferguson, Alan Bundy, Kwabena Nuamah

Recent advancements in Large Language Models (LLMs) are increasingly focused on "reasoning" ability, a concept with many overlapping definitions in the LLM discourse. We take a more structured approach, distinguishing meta-level reasoning (denoting the process of reasoning about intermediate steps required to solve a task) from object-level reasoning (which concerns the low-level execution of the aforementioned steps.) We design a novel question answering task, which is based around the values of geopolitical indicators for various countries over various years. Questions require breaking down into intermediate steps, retrieval of data, and mathematical operations over that data. The meta-level reasoning ability of LLMs is analysed by examining the selection of appropriate tools for answering questions. To bring greater depth to the analysis of LLMs beyond final answer accuracy, our task contains 'essential actions' against which we can compare the tool call output of LLMs to infer the strength of reasoning ability. We find that LLMs demonstrate good meta-level reasoning on our task, yet are flawed in some aspects of task understanding. We find that n-shot prompting has little effect on accuracy; error messages encountered do not often deteriorate performance; and provide additional evidence for the poor numeracy of LLMs. Finally, we discuss the generalisation and limitation of our findings to other task domains.

Nick Ferguson, Liane Guillou, Alan Bundy et al.

Large Language Models (LLMs) excel in natural language tasks but still face challenges in Question Answering (QA) tasks requiring complex, multi-step reasoning. We outline the types of reasoning required in some of these tasks, and reframe them in terms of meta-level reasoning (akin to high-level strategic reasoning or planning) and object-level reasoning (embodied in lower-level tasks such as mathematical reasoning). Franklin, a novel dataset with requirements of meta- and object-level reasoning, is introduced and used along with three other datasets to evaluate four LLMs at question answering tasks requiring multiple steps of reasoning. Results from human annotation studies suggest LLMs demonstrate meta-level reasoning with high frequency, but struggle with object-level reasoning tasks in some of the datasets used. Additionally, evidence suggests that LLMs find the object-level reasoning required for the questions in the Franklin dataset challenging, yet they do exhibit strong performance with respect to the meta-level reasoning requirements.