Andreas Vlachos

Xiaochen Zhu, Georgi Karadzhov, Tom Stafford et al.

Multi-party dialogue is a critical setting for studying collaborative reasoning and decision-making, yet existing datasets rarely focus on structured, in-depth complex reasoning tasks. We introduce DeliChess, a novel dataset of group deliberation dialogues in which participants collaboratively solve multiple-choice chess puzzles. Each group first completes the puzzle individually, then engages in a multi-party discussion before submitting a revised collective answer. The dataset includes 107 dialogues with full transcripts, pre- and post-discussion choices, and metadata on puzzle difficulty and move quality. We evaluate performance using three metrics based on chess engine evaluations, and find that deliberation significantly improves group accuracy. We further analyse the role of probing utterances (i.e., messages that elicit proposals, justifications, or strategic reflection) using a classifier trained on prior deliberation data. While probing makes group performance more variable after discussion, it does not consistently lead to better performance. Our dataset offers a rich testbed for modelling group reasoning, dialogue dynamics, and the resolution of differing perspectives and opinions in a well-defined strategic domain.

Nedjma Ousidhoum, Zhangdie Yuan, Andreas Vlachos · cambridge

Fact-checking requires retrieving evidence related to a claim under investigation. The task can be formulated as question generation based on a claim, followed by question answering. However, recent question generation approaches assume that the answer is known and typically contained in a passage given as input, whereas such passages are what is being sought when verifying a claim. In this paper, we present {\it Varifocal}, a method that generates questions based on different focal points within a given claim, i.e.\ different spans of the claim and its metadata, such as its source and date. Our method outperforms previous work on a fact-checking question generation dataset on a wide range of automatic evaluation metrics. These results are corroborated by our manual evaluation, which indicates that our method generates more relevant and informative questions. We further demonstrate the potential of focal points in generating sets of clarification questions for product descriptions.

Ieva Raminta Staliūnaitė, James Bishop, Andreas Vlachos

The task of Error Prediction, namely predicting whether a model output is correct, is commonly tackled with Uncertainty Quantification (UQ). However, while uncertainty metrics capture when models lack knowledge or capacity to make a prediction, they also reflect aleatoric uncertainty, which is inherent in the model input and context. This paper presents a method for improving error prediction for Large Language Models (LLMs), by disentangling input ambiguity from UQ signal. We conduct experiments on the task of Question Answering (QA) with six UQ metrics and show that UQ metrics are more predictive of errors on unambiguous instances than on questions with multiple plausible answers. We use Gated Experts and Selective Prediction to incorporate gold and predicted ambiguity labels into the error prediction pipeline. We find that ambiguity information improves error prediction scores across model families, training and evaluation paradigms, datasets (including allegedly unambiguous ones), and sources of aleatoric uncertainty, yielding improvements of over 10 points of PRR for individual UQ metrics on standard datasets.

Michael Schlichtkrull, Nedjma Ousidhoum, Andreas Vlachos · cambridge

Automated fact-checking is often presented as an epistemic tool that fact-checkers, social media consumers, and other stakeholders can use to fight misinformation. Nevertheless, few papers thoroughly discuss how. We document this by analysing 100 highly-cited papers, and annotating epistemic elements related to intended use, i.e., means, ends, and stakeholders. We find that narratives leaving out some of these aspects are common, that many papers propose inconsistent means and ends, and that the feasibility of suggested strategies rarely has empirical backing. We argue that this vagueness actively hinders the technology from reaching its goals, as it encourages overclaiming, limits criticism, and prevents stakeholder feedback. Accordingly, we provide several recommendations for thinking and writing about the use of fact-checking artefacts.

Rami Aly, Andreas Vlachos · cambridge

A key component of fact verification is thevevidence retrieval, often from multiple documents. Recent approaches use dense representations and condition the retrieval of each document on the previously retrieved ones. The latter step is performed over all the documents in the collection, requiring storing their dense representations in an index, thus incurring a high memory footprint. An alternative paradigm is retrieve-and-rerank, where documents are retrieved using methods such as BM25, their sentences are reranked, and further documents are retrieved conditioned on these sentences, reducing the memory requirements. However, such approaches can be brittle as they rely on heuristics and assume hyperlinks between documents. We propose a novel retrieve-and-rerank method for multi-hop retrieval, that consists of a retriever that jointly scores documents in the knowledge source and sentences from previously retrieved documents using an autoregressive formulation and is guided by a proof system based on natural logic that dynamically terminates the retrieval process if the evidence is deemed sufficient. This method is competitive with current state-of-the-art methods on FEVER, HoVer and FEVEROUS-S, while using $5$ to $10$ times less memory than competing systems. Evaluation on an adversarial dataset indicates improved stability of our approach compared to commonly deployed threshold-based methods. Finally, the proof system helps humans predict model decisions correctly more often than using the evidence alone.

Neema Kotonya, Andreas Vlachos, Majid Yazdani et al. · cambridge, meta-ai

Policy Compliance Detection (PCD) is a task we encounter when reasoning over texts, e.g. legal frameworks. Previous work to address PCD relies heavily on modeling the task as a special case of Recognizing Textual Entailment. Entailment is applicable to the problem of PCD, however viewing the policy as a single proposition, as opposed to multiple interlinked propositions, yields poor performance and lacks explainability. To address this challenge, more recent proposals for PCD have argued for decomposing policies into expression trees consisting of questions connected with logic operators. Question answering is used to obtain answers to these questions with respect to a scenario. Finally, the expression tree is evaluated in order to arrive at an overall solution. However, this work assumes expression trees are provided by experts, thus limiting its applicability to new policies. In this work, we learn how to infer expression trees automatically from policy texts. We ensure the validity of the inferred trees by introducing constrained decoding using a finite state automaton to ensure the generation of valid trees. We determine through automatic evaluation that 63% of the expression trees generated by our constrained generation model are logically equivalent to gold trees. Human evaluation shows that 88% of trees generated by our model are correct.

Christine de Kock, Tom Stafford, Andreas Vlachos · cambridge

Disagreements are frequently studied from the perspective of either detecting toxicity or analysing argument structure. We propose a framework of dispute tactics that unifies these two perspectives, as well as other dialogue acts which play a role in resolving disputes, such as asking questions and providing clarification. This framework includes a preferential ordering among rebuttal-type tactics, ranging from ad hominem attacks to refuting the central argument. Using this framework, we annotate 213 disagreements (3,865 utterances) from Wikipedia Talk pages. This allows us to investigate research questions around the tactics used in disagreements; for instance, we provide empirical validation of the approach to disagreement recommended by Wikipedia. We develop models for multilabel prediction of dispute tactics in an utterance, achieving the best performance with a transformer-based label powerset model. Adding an auxiliary task to incorporate the ordering of rebuttal tactics further yields a statistically significant increase. Finally, we show that these annotations can be used to provide useful additional signals to improve performance on the task of predicting escalation.

Youmna Farag, Charlotte O. Brand, Jacopo Amidei et al. · cambridge

Recent research on argumentative dialogues has focused on persuading people to take some action, changing their stance on the topic of discussion, or winning debates. In this work, we focus on argumentative dialogues that aim to open up (rather than change) people's minds to help them become more understanding to views that are unfamiliar or in opposition to their own convictions. To this end, we present a dataset of 183 argumentative dialogues about 3 controversial topics: veganism, Brexit and COVID-19 vaccination. The dialogues were collected using the Wizard of Oz approach, where wizards leverage a knowledge-base of arguments to converse with participants. Open-mindedness is measured before and after engaging in the dialogue using a questionnaire from the psychology literature, and success of the dialogue is measured as the change in the participant's stance towards those who hold opinions different to theirs. We evaluate two dialogue models: a Wikipedia-based and an argument-based model. We show that while both models perform closely in terms of opening up minds, the argument-based model is significantly better on other dialogue properties such as engagement and clarity.

Georgi Karadzhov, Tom Stafford, Andreas Vlachos · cambridge

People leverage group discussions to collaborate in order to solve complex tasks, e.g. in project meetings or hiring panels. By doing so, they engage in a variety of conversational strategies where they try to convince each other of the best approach and ultimately reach a decision. In this work, we investigate methods for detecting what makes someone change their mind. To this end, we leverage a recently introduced dataset containing group discussions of people collaborating to solve a task. To find out what makes someone change their mind, we incorporate various techniques such as neural text classification and language-agnostic change point detection. Evaluation of these methods shows that while the task is not trivial, the best way to approach it is using a language-aware model with learning-to-rank training. Finally, we examine the cues that the models develop as indicative of the cause of a change of mind.

Julius Cheng, Andreas Vlachos

Minimum Bayes risk (MBR) decoding outputs the hypothesis with the highest expected utility over the model distribution for some utility function. It has been shown to improve accuracy over beam search in conditional language generation problems and especially neural machine translation, in both human and automatic evaluations. However, the standard sampling-based algorithm for MBR is substantially more computationally expensive than beam search, requiring a large number of samples as well as a quadratic number of calls to the utility function, limiting its applicability. We describe an algorithm for MBR which gradually grows the number of samples used to estimate the utility while pruning hypotheses that are unlikely to have the highest utility according to confidence estimates obtained with bootstrap sampling. Our method requires fewer samples and drastically reduces the number of calls to the utility function compared to standard MBR while being statistically indistinguishable in terms of accuracy. We demonstrate the effectiveness of our approach in experiments on three language pairs, using chrF++ and COMET as utility/evaluation metrics.

Xiaochen Zhu, Caiqi Zhang, Yizhou Chi et al. · cambridge

Multi-agent debate (MAD) is widely used to improve large language model (LLM) performance through test-time scaling, yet recent work shows that vanilla MAD often underperforms simple majority vote despite higher computational cost. Studies show that, under homogeneous agents and uniform belief updates, debate preserves expected correctness and therefore cannot reliably improve outcomes. Drawing on findings from human deliberation and collective decision-making, we identify two key mechanisms missing from vanilla MAD: (i) diversity of initial viewpoints and (ii) explicit, calibrated confidence communication. We propose two lightweight interventions. First, a diversity-aware initialisation that selects a more diverse pool of candidate answers, increasing the likelihood that a correct hypothesis is present at the start of debate. Second, a confidence-modulated debate protocol in which agents express calibrated confidence and condition their updates on others' confidence. We show theoretically that diversity-aware initialisation improves the prior probability of MAD success without changing the underlying update dynamics, while confidence-modulated updates enable debate to systematically drift to the correct hypothesis. Empirically, across six reasoning-oriented QA benchmarks, our methods consistently outperform vanilla MAD and majority vote. Our results connect human deliberation with LLM-based debate and demonstrate that simple, principled modifications can substantially enhance debate effectiveness.

Rami Aly, Marek Strong, Andreas Vlachos

Fact verification systems assess a claim's veracity based on evidence. An important consideration in designing them is faithfulness, i.e. generating explanations that accurately reflect the reasoning of the model. Recent works have focused on natural logic, which operates directly on natural language by capturing the semantic relation of spans between an aligned claim with its evidence via set-theoretic operators. However, these approaches rely on substantial resources for training, which are only available for high-resource languages. To this end, we propose to use question answering to predict natural logic operators, taking advantage of the generalization capabilities of instruction-tuned language models. Thus, we obviate the need for annotated training data while still relying on a deterministic inference system. In a few-shot setting on FEVER, our approach outperforms the best baseline by $4.3$ accuracy points, including a state-of-the-art pre-trained seq2seq natural logic system, as well as a state-of-the-art prompt-based classifier. Our system demonstrates its robustness and portability, achieving competitive performance on a counterfactual dataset and surpassing all approaches without further annotation on a Danish verification dataset. A human evaluation indicates that our approach produces more plausible proofs with fewer erroneous natural logic operators than previous natural logic-based systems.

Liang Chen, Zekun Wang, Shuhuai Ren et al. · pku

Building on the foundations of language modeling in natural language processing, Next Token Prediction (NTP) has evolved into a versatile training objective for machine learning tasks across various modalities, achieving considerable success. As Large Language Models (LLMs) have advanced to unify understanding and generation tasks within the textual modality, recent research has shown that tasks from different modalities can also be effectively encapsulated within the NTP framework, transforming the multimodal information into tokens and predict the next one given the context. This survey introduces a comprehensive taxonomy that unifies both understanding and generation within multimodal learning through the lens of NTP. The proposed taxonomy covers five key aspects: Multimodal tokenization, MMNTP model architectures, unified task representation, datasets \& evaluation, and open challenges. This new taxonomy aims to aid researchers in their exploration of multimodal intelligence. An associated GitHub repository collecting the latest papers and repos is available at https://github.com/LMM101/Awesome-Multimodal-Next-Token-Prediction

Eric Chamoun, Marzieh Saeidi, Andreas Vlachos

Prior research has shown that typical fact-checking models for stand-alone claims struggle with claims made in dialogues. As a solution, fine-tuning these models on labelled dialogue data has been proposed. However, creating separate models for each use case is impractical, and we show that fine-tuning models for dialogue results in poor performance on typical fact-checking. To overcome this challenge, we present techniques that allow us to use the same models for both dialogue and typical fact-checking. These mainly focus on retrieval adaptation and transforming conversational inputs so that they can be accurately predicted by models trained on stand-alone claims. We demonstrate that a typical fact-checking model incorporating these techniques is competitive with state-of-the-art models fine-tuned for dialogue, while maintaining its accuracy on stand-alone claims.

Mubashara Akhtar, Michael Schlichtkrull, Andreas Vlachos

Current automated fact-checking (AFC) approaches typically evaluate evidence either implicitly via the predicted verdicts or through exact matches with predefined closed knowledge sources, such as Wikipedia. However, these methods are limited due to their reliance on evaluation metrics originally designed for other purposes and constraints from closed knowledge sources. In this work, we introduce \textbf{\textcolor{skyblue}{Ev\textsuperscript{2}}\textcolor{orangebrown}{R}} which combines the strengths of reference-based evaluation and verdict-level proxy scoring. Ev\textsuperscript{2}R jointly assesses how well the evidence aligns with the gold references and how reliably it supports the verdict, addressing the shortcomings of prior methods. We evaluate Ev\textsuperscript{2}R against three types of evidence evaluation approaches: reference-based, proxy-reference, and reference-less baselines. Assessments against human ratings and adversarial tests demonstrate that Ev\textsuperscript{2}R consistently outperforms existing scoring approaches in accuracy and robustness. It achieves stronger correlation with human judgments and greater robustness to adversarial perturbations, establishing it as a reliable metric for evidence evaluation in AFC.\footnote{Code is available at \href{https://github.com/mubasharaak/fc-evidence-evaluation}{https://github.com/mubasharaak/fc-evidence-evaluation}.}

Eric Chamoun, Yizhou Chi, Yulong Chen et al.

Scientific progress depends on sequences of enabling contributions, yet existing AI4Science benchmarks largely focus on citation prediction, literature retrieval, or idea generation rather than the dependencies that make progress possible. In this paper, we introduce discovery pathway forecasting: given a target scientific contribution and the prior literature available at a specified time, the task is to (1) identify the enabling contributions required to realize it and (2) ground each in prior work when such prior work exists. We present SciPaths, a benchmark of 262 expert-annotated gold pathways and 2,444 silver pathways constructed from machine learning and natural language processing papers, where each pathway records enabling contributions, roles, rationales, and prior-work groundings or unmapped decisions. Evaluating frontier and open-weight language models, we find that the best model reaches only 0.189 F1 under strict semantic matching, with core methodological dependencies hardest to recover. Prior-work grounding improves substantially when gold enabling contributions are provided, showing that decomposition quality is a major bottleneck for end-to-end pathway recovery. SciPaths therefore shifts evaluation toward a missing capability in scientific forecasting: reasoning backward from a target contribution to the enabling scientific building blocks and prior-work dependencies that make it feasible.

Rui Cao, Yuming Jiang, Michael Schlichtkrull et al.

Multimodal Large Language Models (MLLMs) can enhance trustworthiness by aligning with human preferences. As human preference labeling is laborious, recent works employ evaluation models for assessing MLLMs' responses, using the model-based assessments to automate preference dataset construction. This approach, however, faces challenges with MLLMs' lengthy and compositional responses, which often require diverse reasoning skills that a single evaluation model may not fully possess. Additionally, most existing methods rely on closed-source models as evaluators. To address limitations, we propose DecompGen, a decomposable framework that uses an ensemble of open-sourced expert models. DecompGen breaks down each response into atomic verification tasks, assigning each task to an appropriate expert model to generate fine-grained assessments. The DecompGen feedback is used to automatically construct our preference dataset, DGPref. MLLMs aligned with DGPref via preference learning show improvements in trustworthiness, demonstrating the effectiveness of DecompGen.

Georgi Karadzhov, Andreas Vlachos, Tom Stafford

We explore different aspects of cognitive diversity and its effect on the success of group deliberation. To evaluate this, we use 500 dialogues from small, online groups discussing the Wason Card Selection task - the DeliData corpus. Leveraging the corpus, we perform quantitative analysis evaluating three different measures of cognitive diversity. First, we analyse the effect of group size as a proxy measure for diversity. Second, we evaluate the effect of the size of the initial idea pool. Finally, we look into the content of the discussion by analysing discussed solutions, discussion patterns, and how conversational probing can improve those characteristics. Despite the reputation of groups for compounding bias, we show that small groups can, through dialogue, overcome intuitive biases and improve individual decision-making. Across a large sample and different operationalisations, we consistently find that greater cognitive diversity is associated with more successful group deliberation. Code and data used for the analysis are available in the repository: https://github.com/gkaradzhov/cognitive-diversity-groups-cogsci24.

Hardy, Andreas Vlachos

Recent work on abstractive summarization has made progress with neural encoder-decoder architectures. However, such models are often challenged due to their lack of explicit semantic modeling of the source document and its summary. In this paper, we extend previous work on abstractive summarization using Abstract Meaning Representation (AMR) with a neural language generation stage which we guide using the source document. We demonstrate that this guidance improves summarization results by 7.4 and 10.5 points in ROUGE-2 using gold standard AMR parses and parses obtained from an off-the-shelf parser respectively. We also find that the summarization performance using the latter is 2 ROUGE-2 points higher than that of a well-established neural encoder-decoder approach trained on a larger dataset. Code is available at \url{https://github.com/sheffieldnlp/AMR2Text-summ}

Joycelyn Teo, Rui Cao, Zhenyun Deng et al.

Automated Fact-Checking (AFC) relies on claim extraction as a first step, yet existing methods largely overlook the multimodal nature of today's misinformation. Social media posts often combine short, informal text with images such as memes, screenshots, and photos, creating challenges that differ from both text-only claim extraction and well-studied multimodal tasks like image captioning or visual question answering. In this work, we present the first benchmark for multimodal claim extraction from social media, consisting of posts containing text and one or more images, annotated with gold-standard claims derived from real-world fact-checkers. We evaluate state-of-the-art multimodal LLMs (MLLMs) under a three-part evaluation framework (semantic alignment, faithfulness, and decontextualization) and find that baseline MLLMs struggle to model rhetorical intent and contextual cues. To address this, we introduce MICE, an intent-aware framework which shows improvements in intent-critical cases.

Michael Schlichtkrull, Yulong Chen, Chenxi Whitehouse et al. · amazon-science

The Automated Verification of Textual Claims (AVeriTeC) shared task asks participants to retrieve evidence and predict veracity for real-world claims checked by fact-checkers. Evidence can be found either via a search engine, or via a knowledge store provided by the organisers. Submissions are evaluated using AVeriTeC score, which considers a claim to be accurately verified if and only if both the verdict is correct and retrieved evidence is considered to meet a certain quality threshold. The shared task received 21 submissions, 18 of which surpassed our baseline. The winning team was TUDA_MAI with an AVeriTeC score of 63%. In this paper we describe the shared task, present the full results, and highlight key takeaways from the shared task.

Caiqi Zhang, Zhijiang Guo, Andreas Vlachos · cambridge

This paper investigates the potential benefits of language-specific fact-checking models, focusing on the case of Chinese. We first demonstrate the limitations of translation-based methods and multilingual large language models (e.g., GPT-4), highlighting the need for language-specific systems. We further propose a Chinese fact-checking system that can better retrieve evidence from a document by incorporating context information. To better analyze token-level biases in different systems, we construct an adversarial dataset based on the CHEF dataset, where each instance has large word overlap with the original one but holds the opposite veracity label. Experimental results on the CHEF dataset and our adversarial dataset show that our proposed method outperforms translation-based methods and multilingual LLMs and is more robust toward biases, while there is still large room for improvement, emphasizing the importance of language-specific fact-checking systems.

Marek Strong, Andreas Vlachos

Reasoning over temporal and numerical data, such as time series, is a crucial aspect of fact-checking. While many systems have recently been developed to handle this form of evidence, their evaluation remains limited by existing datasets, which often lack structured evidence, provide insufficient justifications for verdicts, or rely on synthetic claims. In this paper, we introduce TSVer, a new benchmark dataset for fact verification focusing on temporal and numerical reasoning with time-series evidence. TSVer contains 287 real-world claims sourced from 38 fact-checking organizations and a curated database of 400 time series covering diverse domains. Each claim is annotated with time frames across all pertinent time series, along with a verdict and justifications reflecting how the evidence is used to reach the verdict. Using an LLM-assisted multi-step annotation process, we improve the quality of our annotations and achieve an inter-annotator agreement of kappa=0.745 on verdicts. We also develop a baseline for verifying claims against time-series evidence and show that even the state-of-the-art reasoning models like Gemini-2.5-Pro are challenged by time series, achieving a 63.37 accuracy score on verdicts and an Ev2R score of 48.63 on verdict justifications.

Zhangdie Yuan, Andreas Vlachos · cambridge

Despite progress in automated fact-checking, most systems require a significant amount of labeled training data, which is expensive. In this paper, we propose a novel zero-shot method, which instead of operating directly on the claim and evidence sentences, decomposes them into semantic triples augmented using external knowledge graphs, and uses large language models trained for natural language inference. This allows it to generalize to adversarial datasets and domains that supervised models require specific training data for. Our empirical results show that our approach outperforms previous zero-shot approaches on FEVER, FEVER-Symmetric, FEVER 2.0, and Climate-FEVER, while being comparable or better than supervised models on the adversarial and the out-of-domain datasets.

Caiqi Zhang, Xiaochen Zhu, Chengzu Li et al. · cambridge

Hallucination remains a major challenge for the safe and trustworthy deployment of large language models (LLMs) in factual content generation. Prior work has explored confidence estimation as an effective approach to hallucination detection, but often relies on post-hoc self-consistency methods that require computationally expensive sampling. Verbalized confidence offers a more efficient alternative, but existing approaches are largely limited to short-form question answering (QA) tasks and do not generalize well to open-ended generation. In this paper, we propose LoVeC (Long-form Verbalized Confidence), an on-the-fly verbalized confidence estimation method for long-form generation. Specifically, we use reinforcement learning (RL) to train LLMs to append numerical confidence scores to each generated statement, serving as a direct and interpretable signal of the factuality of generation. Our experiments consider both on-policy and off-policy RL methods, including DPO, ORPO, and GRPO, to enhance the model calibration. We introduce two novel evaluation settings, free-form tagging and iterative tagging, to assess different verbalized confidence estimation methods. Experiments on three long-form QA datasets show that our RL-trained models achieve better calibration and generalize robustly across domains. Also, our method is highly efficient, as it only requires adding a few tokens to the output being decoded.

Zhangdie Yuan, Eric Chamoun, Rami Aly et al. · cambridge

This paper introduces PRobELM (Plausibility Ranking Evaluation for Language Models), a benchmark designed to assess language models' ability to discern more plausible from less plausible scenarios through their parametric knowledge. While benchmarks such as TruthfulQA emphasise factual accuracy or truthfulness, and others such as COPA explore plausible scenarios without explicitly incorporating world knowledge, PRobELM seeks to bridge this gap by evaluating models' capabilities to prioritise plausible scenarios that leverage world knowledge over less plausible alternatives. This design allows us to assess the potential of language models for downstream use cases such as literature-based discovery where the focus is on identifying information that is likely but not yet known. Our benchmark is constructed from a dataset curated from Wikidata edit histories, tailored to align the temporal bounds of the training data for the evaluated models. PRobELM facilitates the evaluation of language models across multiple prompting types, including statement, text completion, and question-answering. Experiments with 10 models of various sizes and architectures on the relationship between model scales, training recency, and plausibility performance, reveal that factual accuracy does not directly correlate with plausibility performance and that up-to-date training data enhances plausibility assessment across different model architectures.

Pietro Lesci, Andreas Vlachos

Active learning for imbalanced classification tasks is challenging as the minority classes naturally occur rarely. Gathering a large pool of unlabelled data is thus essential to capture minority instances. Standard pool-based active learning is computationally expensive on large pools and often reaches low accuracy by overfitting the initial decision boundary, thus failing to explore the input space and find minority instances. To address these issues we propose AnchorAL. At each iteration, AnchorAL chooses class-specific instances from the labelled set, or anchors, and retrieves the most similar unlabelled instances from the pool. This resulting subpool is then used for active learning. Using a small, fixed-sized subpool AnchorAL allows scaling any active learning strategy to large pools. By dynamically selecting different anchors at each iteration it promotes class balance and prevents overfitting the initial decision boundary, thus promoting the discovery of new clusters of minority instances. In experiments across different classification tasks, active learning strategies, and model architectures AnchorAL is (i) faster, often reducing runtime from hours to minutes, (ii) trains more performant models, (iii) and returns more balanced datasets than competing methods.

Pietro Lesci, Clara Meister, Thomas Hofmann et al.

Modern language models are typically trained over subword sequences, but ultimately define probabilities over character-strings. Ideally, the choice of the tokeniser -- which maps character-strings to subwords -- should not affect the probability assigned to the underlying character-string; in practice, it does. We define this mismatch as tokenisation bias. In this work, we quantify one particular type of tokenisation bias: the effect of including or not a subword (e.g., $\langle hello \rangle$) in a tokeniser's vocabulary on the probability a trained model assigns to the corresponding characters (i.e., \textit{``hello''}). Estimating this effect is challenging because each model is trained with only one tokeniser. We address this by framing tokenisation bias as a causal effect and estimating it using the regression discontinuity design. Specifically, we exploit the fact that tokenisation algorithms rank subwords and add the first $K$ to a tokeniser's vocabulary, where $K$ is an arbitrary cutoff point. As such, we can estimate a causal effect by comparing similar subwords around this cutoff. Experimentally, we find that tokenisation consistently affects models' outputs across scales, vocabularies, and tokenisers. Notably, a subword's presence in a small model's vocabulary may increase its characters' probability by up to 17 times, highlighting tokenisation as a key design choice in language modelling.

Xiaochen Zhu, Caiqi Zhang, Tom Stafford et al. · cambridge

The conformity effect describes the tendency of individuals to align their responses with the majority. Studying this bias in large language models (LLMs) is crucial, as LLMs are increasingly used in various information-seeking and decision-making tasks as conversation partners to improve productivity. Thus, conformity to incorrect responses can compromise their effectiveness. In this paper, we adapt psychological experiments to examine the extent of conformity in popular LLMs. Our findings reveal that all tested models exhibit varying levels of conformity toward the majority, regardless of their initial choice or correctness, across different knowledge domains. Notably, we are the first to show that LLMs are more likely to conform when they are more uncertain in their own prediction. We further explore factors that influence conformity, such as training paradigms and input characteristics, finding that instruction-tuned models are less susceptible to conformity, while increasing the naturalness of majority tones amplifies conformity. Finally, we propose two interventions, Devil's Advocate and Question Distillation, to mitigate conformity, providing insights into building more robust language models.

Rui Cao, Zifeng Ding, Zhijiang Guo et al.

Textual claims are often accompanied by images to enhance their credibility and spread on social media, but this also raises concerns about the spread of misinformation. Existing datasets for automated verification of image-text claims remain limited, as they often consist of synthetic claims and lack evidence annotations to capture the reasoning behind the verdict. In this work, we introduce AVerImaTeC, a dataset consisting of 1,297 real-world image-text claims. Each claim is annotated with question-answer (QA) pairs containing evidence from the web, reflecting a decomposed reasoning regarding the verdict. We mitigate common challenges in fact-checking datasets such as contextual dependence, temporal leakage, and evidence insufficiency, via claim normalization, temporally constrained evidence annotation, and a two-stage sufficiency check. We assess the consistency of the annotation in AVerImaTeC via inter-annotator studies, achieving a $κ=0.742$ on verdicts and $74.7\%$ consistency on QA pairs. We also propose a novel evaluation method for evidence retrieval and conduct extensive experiments to establish baselines for verifying image-text claims using open-web evidence.

Zhangdie Yuan, Zifeng Ding, Andreas Vlachos · cambridge

Forecasting is an important task in many domains, such as technology and economics. However existing forecasting benchmarks largely lack comprehensive confidence assessment, focus on limited question types, and often consist of artificial questions that do not align with real-world human forecasting needs. To address these gaps, we introduce FOReCAst (Future Outcome Reasoning and Confidence Assessment), a benchmark that evaluates models' ability to make predictions and their confidence in them. FOReCAst spans diverse forecasting scenarios involving Boolean questions, timeframe prediction, and quantity estimation, enabling a comprehensive evaluation of both prediction accuracy and confidence calibration for real-world applications.

Xiaochen Zhu, Georgi Karadzhov, Chenxi Whitehouse et al.

Diffusion models have shown promise in text generation, but often struggle with generating long, coherent, and contextually accurate text. Token-level diffusion doesn't model word-order dependencies explicitly and operates on short, fixed output windows, while passage-level diffusion struggles with learning robust representations for long-form text. To address these challenges, we propose Segment-Level Diffusion (SLD), a framework that enhances diffusion-based text generation through text segmentation, robust representation training with adversarial and contrastive learning, and improved latent-space guidance. By segmenting long-form outputs into multiple latent representations and decoding them with an autoregressive decoder, SLD simplifies diffusion predictions and improves scalability. Experiments on four datasets demonstrate that, when compared to other diffusion and autoregressive baselines SLD achieves competitive or superior fluency, coherence, and contextual compatibility in automatic and human evaluations.

Julius Cheng, Maike Züfle, Vilém Zouhar et al. · eth-zurich

Reranking a list of candidates from a machine translation system with an external scoring model and returning the highest-scoring candidate remains a simple and effective method for improving the overall output quality. Translation scoring models continue to grow in size, with the best models being comparable to generation models. Thus, reranking can add substantial computational cost to the translation pipeline. In this work, we pose reranking as a Bayesian optimization (BayesOpt) problem. By strategically selecting candidates to score based on a balance of exploration and exploitation, we show that it is possible to find top-scoring candidates when scoring only a fraction of the candidate list. For instance, our method achieves the same CometKiwi score using only 70 scoring evaluations compared a baseline system using 180. We present a multi-fidelity setting for BayesOpt, where the candidates are first scored with a cheaper but noisier proxy scoring model, which further improves the cost-performance tradeoff when using smaller but well-trained distilled proxy scorers.

Rami Aly, Andreas Vlachos

Fact verification on tabular evidence incentivises the use of symbolic reasoning models where a logical form is constructed (e.g. a LISP-style program), providing greater verifiability than fully neural approaches. However, these systems typically rely on well-formed tables, restricting their use in many scenarios. An emerging symbolic reasoning paradigm for textual evidence focuses on natural logic inference, which constructs proofs by modelling set-theoretic relations between a claim and its evidence in natural language. This approach provides flexibility and transparency but is less compatible with tabular evidence since the relations do not extend to arithmetic functions. We propose a set-theoretic interpretation of numerals and arithmetic functions in the context of natural logic, enabling the integration of arithmetic expressions in deterministic proofs. We leverage large language models to generate arithmetic expressions by generating questions about salient parts of a claim which are answered by executing appropriate functions on tables. In a few-shot setting on FEVEROUS, we achieve an accuracy of 71.4, outperforming both fully neural and symbolic reasoning models by 3.4 points. When evaluated on TabFact without any further training, our method remains competitive with an accuracy lead of 0.5 points.

Rui Cao, Zhenyun Deng, Yulong Chen et al.

The Automatic Verification of Image-Text Claims (AVerImaTeC) shared task aims to advance system development for retrieving evidence and verifying real-world image-text claims. Participants were allowed to either employ external knowledge sources, such as web search engines, or leverage the curated knowledge store provided by the organizers. System performance was evaluated using the AVerImaTeC score, defined as a conditional verdict accuracy in which a verdict is considered correct only when the associated evidence score exceeds a predefined threshold. The shared task attracted 14 submissions during the development phase and 6 submissions during the testing phase. All participating systems in the testing phase outperformed the baseline provided. The winning team, HUMANE, achieved an AVerImaTeC score of 0.5455. This paper provides a detailed description of the shared task, presents the complete evaluation results, and discusses key insights and lessons learned.

Zhangdie Yuan, Zifeng Ding, Andreas Vlachos · cambridge

Prior work has largely treated future event prediction as a static task, failing to consider how forecasts and the confidence in them should evolve as new evidence emerges. To address this gap, we introduce EVOLVECAST, a framework for evaluating whether large language models appropriately revise their predictions in response to new information. In particular, EVOLVECAST assesses whether LLMs adjust their forecasts when presented with information released after their training cutoff. We use human forecasters as a comparative reference to analyze prediction shifts and confidence calibration under updated contexts. While LLMs demonstrate some responsiveness to new information, their updates are often inconsistent or overly conservative. We further find that neither verbalized nor logits-based confidence estimates consistently outperform the other, and both remain far from the human reference standard. Across settings, models tend to express conservative bias, underscoring the need for more robust approaches to belief updating.

Zhenyun Deng, Yulong Chen, Andreas Vlachos

Extracting individual sentences from a document as evidence or reasoning steps is commonly done in many NLP tasks. However, extracted sentences often lack context necessary to make them understood, e.g., coreference and background information. To this end, we propose a content selection and planning framework for zero-shot decontextualisation, which determines what content should be mentioned and in what order for a sentence to be understood out of context. Specifically, given a potentially ambiguous sentence and its context, we first segment it into basic semantically-independent units. We then identify potentially ambiguous units from the given sentence, and extract relevant units from the context based on their discourse relations. Finally, we generate a content plan to rewrite the sentence by enriching each ambiguous unit with its relevant units. Experimental results demonstrate that our approach is competitive for sentence decontextualisation, producing sentences that exhibit better semantic integrity and discourse coherence, outperforming existing methods.

Yulong Chen, Michael Sejr Schlichtkrull, Zhenyun Deng et al.

Political pledges reflect candidates' policy commitments, but tracking their fulfilment requires reasoning over incremental evidence distributed across multiple, dynamically updated sources. Existing methods simplify this task into a document classification task, overlooking its dynamic, temporal and multi-document nature. To address this issue, we introduce \textsc{PledgeTracker}, a system that reformulates pledge verification into structured event timeline construction. PledgeTracker consists of three core components: (1) a multi-step evidence retrieval module; (2) a timeline construction module and; (3) a fulfilment filtering module, allowing the capture of the evolving nature of pledge fulfilment and producing interpretable and structured timelines. We evaluate PledgeTracker in collaboration with professional fact-checkers in real-world workflows, demonstrating its effectiveness in retrieving relevant evidence and reducing human verification effort.

Zifeng Ding, Shenyang Huang, Zeyu Cao et al. · cambridge

Forecasting future links is a central task in temporal graph (TG) reasoning, requiring models to leverage historical interactions to predict upcoming ones. Traditional neural approaches, such as temporal graph neural networks, achieve strong performance but lack explainability and cannot be applied to unseen graphs without retraining. Recent studies have begun to explore using large language models (LLMs) for graph reasoning, but most of them are constrained to static graphs or small synthetic TGs and lack the evaluation of the quality of reasoning traces generated by LLMs. In this work, we present Reasoning-Enhanced Learning for Temporal Graphs (ReaL-TG), a reinforcement learning framework that fine-tunes LLMs to perform explainable link forecasting on real-world TGs. ReaL-TG uses outcome-based reward to encourage models to self-explore reasoning strategies from graph structure and to produce explanations that directly justify their predictions. To enable evaluation on LLM-generated reasoning traces, we propose a new evaluation protocol combining ranking metrics with an LLM-as-a-Judge system that assesses both the quality of reasoning and the impact of hallucinations. Experiments with ReaL-TG-4B, obtained by fine-tuning Qwen3-4B under our framework, show that it outperforms much larger frontier LLMs, including GPT-5 mini, on ranking metrics, while producing high-quality explanations confirmed by both the LLM judge and human evaluation.

Ieva Raminta Staliūnaitė, Julius Cheng, Andreas Vlachos

The predictive uncertainty of machine translation (MT) models is typically used as a quality estimation proxy. In this work, we posit that apart from confidently translating when a single correct translation exists, models should also maintain uncertainty when the input is ambiguous. We use uncertainty to measure gender bias in MT systems. When the source sentence includes a lexeme whose gender is not overtly marked, but whose target-language equivalent requires gender specification, the model must infer the appropriate gender from the context and can be susceptible to biases. Prior work measured bias via gender accuracy, however it cannot be applied to ambiguous cases. Using semantic uncertainty, we are able to assess bias when translating both ambiguous and unambiguous source sentences, and find that high translation accuracy does not correlate with exhibiting uncertainty appropriately, and that debiasing affects the two cases differently.

Michalis Korakakis, Andreas Vlachos, Adrian Weller

Empirical risk minimization (ERM) incentivizes models to exploit shortcuts, i.e., spurious correlations between input attributes and labels that are prevalent in the majority of the training data but unrelated to the task at hand. This reliance hinders generalization on minority examples, where such correlations do not hold. Existing shortcut mitigation approaches are model-specific, difficult to tune, computationally expensive, and fail to improve learned representations. To address these issues, we propose InterpoLated Learning (InterpoLL) which interpolates the representations of majority examples to include features from intra-class minority examples with shortcut-mitigating patterns. This weakens shortcut influence, enabling models to acquire features predictive across both minority and majority examples. Experimental results on multiple natural language understanding tasks demonstrate that InterpoLL improves minority generalization over both ERM and state-of-the-art shortcut mitigation methods, without compromising accuracy on majority examples. Notably, these gains persist across encoder, encoder-decoder, and decoder-only architectures, demonstrating the method's broad applicability.

Eric Chamoun, Nedjma Ousidhoum, Michael Schlichtkrull et al.

Clarifying the research framing of NLP artefacts (e.g., models, datasets, etc.) is crucial to aligning research with practical applications. Recent studies manually analyzed NLP research across domains, showing that few papers explicitly identify key stakeholders, intended uses, or appropriate contexts. In this work, we propose to automate this analysis, developing a three-component system that infers research framings by first extracting key elements (means, ends, stakeholders), then linking them through interpretable rules and contextual reasoning. We evaluate our approach on two domains: automated fact-checking using an existing dataset, and hate speech detection for which we annotate a new dataset-achieving consistent improvements over strong LLM baselines. Finally, we apply our system to recent automated fact-checking papers and uncover three notable trends: a rise in vague or underspecified research goals, increased emphasis on scientific exploration over application, and a shift toward supporting human fact-checkers rather than pursuing full automation.

Zhangdie Yuan, Andreas Vlachos · cambridge

Capturing symmetric (e.g., country borders another country) and antisymmetric (e.g., parent_of) relations is crucial for a variety of applications. This paper tackles this challenge by introducing a novel Wikidata-derived natural language inference dataset designed to evaluate large language models (LLMs). Our findings reveal that LLMs perform comparably to random chance on this benchmark, highlighting a gap in relational understanding. To address this, we explore encoder retraining via contrastive learning with k-nearest neighbors. The retrained encoder matches the performance of fine-tuned classification heads while offering additional benefits, including greater efficiency in few-shot learning and improved mitigation of catastrophic forgetting.

Jooyoung Lee, Xiaochen Zhu, Georgi Karadzhov et al.

The proliferation of generative models has presented significant challenges in distinguishing authentic human-authored content from deepfake content. Collaborative human efforts, augmented by AI tools, present a promising solution. In this study, we explore the potential of DeepFakeDeLiBot, a deliberation-enhancing chatbot, to support groups in detecting deepfake text. Our findings reveal that group-based problem-solving significantly improves the accuracy of identifying machine-generated paragraphs compared to individual efforts. While engagement with DeepFakeDeLiBot does not yield substantial performance gains overall, it enhances group dynamics by fostering greater participant engagement, consensus building, and the frequency and diversity of reasoning-based utterances. Additionally, participants with higher perceived effectiveness of group collaboration exhibited performance benefits from DeepFakeDeLiBot. These findings underscore the potential of deliberative chatbots in fostering interactive and productive group dynamics while ensuring accuracy in collaborative deepfake text detection. \textit{Dataset and source code used in this study will be made publicly available upon acceptance of the manuscript.

Michalis Korakakis, Andreas Vlachos, Adrian Weller

Active Learning aims to minimize annotation effort by selecting the most useful instances from a pool of unlabeled data. However, typical active learning methods overlook the presence of distinct example groups within a class, whose prevalence may vary, e.g., in occupation classification datasets certain demographics are disproportionately represented in specific classes. This oversight causes models to rely on shortcuts for predictions, i.e., spurious correlations between input attributes and labels occurring in well-represented groups. To address this issue, we propose Active Learning Via INterpolation (ALVIN), which conducts intra-class interpolations between examples from under-represented and well-represented groups to create anchors, i.e., artificial points situated between the example groups in the representation space. By selecting instances close to the anchors for annotation, ALVIN identifies informative examples exposing the model to regions of the representation space that counteract the influence of shortcuts. Crucially, since the model considers these examples to be of high certainty, they are likely to be ignored by typical active learning methods. Experimental results on six datasets encompassing sentiment analysis, natural language inference, and paraphrase detection demonstrate that ALVIN outperforms state-of-the-art active learning methods in both in-distribution and out-of-distribution generalization.

Lexin Zhou, Youmna Farag, Andreas Vlachos

Research on dialogue constructiveness assessment focuses on (i) analysing conversational factors that influence individuals to take specific actions, win debates, change their perspectives or broaden their open-mindedness and (ii) predicting constructiveness outcomes following dialogues for such use cases. These objectives can be achieved by training either interpretable feature-based models (which often involve costly human annotations) or neural models such as pre-trained language models (which have empirically shown higher task accuracy but lack interpretability). In this paper we propose an LLM feature-based framework for dialogue constructiveness assessment that combines the strengths of feature-based and neural approaches, while mitigating their downsides. The framework first defines a set of dataset-independent and interpretable linguistic features, which can be extracted by both prompting an LLM and simple heuristics. Such features are then used to train LLM feature-based models. We apply this framework to three datasets of dialogue constructiveness and find that our LLM feature-based models outperform or performs at least as well as standard feature-based models and neural models. We also find that the LLM feature-based model learns more robust prediction rules instead of relying on superficial shortcuts, which often trouble neural models.

Pietro Lesci, Clara Meister, Thomas Hofmann et al.

Understanding memorisation in language models has practical and societal implications, e.g., studying models' training dynamics or preventing copyright infringements. Prior work defines memorisation as the causal effect of training with an instance on the model's ability to predict that instance. This definition relies on a counterfactual: the ability to observe what would have happened had the model not seen that instance. Existing methods struggle to provide computationally efficient and accurate estimates of this counterfactual. Further, they often estimate memorisation for a model architecture rather than for a specific model instance. This paper fills an important gap in the literature, proposing a new, principled, and efficient method to estimate memorisation based on the difference-in-differences design from econometrics. Using this method, we characterise a model's memorisation profile--its memorisation trends across training--by only observing its behaviour on a small set of instances throughout training. In experiments with the Pythia model suite, we find that memorisation (i) is stronger and more persistent in larger models, (ii) is determined by data order and learning rate, and (iii) has stable trends across model sizes, thus making memorisation in larger models predictable from smaller ones.

Zhenyun Deng, Michael Schlichtkrull, Andreas Vlachos

Selecting which claims to check is a time-consuming task for human fact-checkers, especially from documents consisting of multiple sentences and containing multiple claims. However, existing claim extraction approaches focus more on identifying and extracting claims from individual sentences, e.g., identifying whether a sentence contains a claim or the exact boundaries of the claim within a sentence. In this paper, we propose a method for document-level claim extraction for fact-checking, which aims to extract check-worthy claims from documents and decontextualise them so that they can be understood out of context. Specifically, we first recast claim extraction as extractive summarization in order to identify central sentences from documents, then rewrite them to include necessary context from the originating document through sentence decontextualisation. Evaluation with both automatic metrics and a fact-checking professional shows that our method is able to extract check-worthy claims from documents more accurately than previous work, while also improving evidence retrieval.

Mubashara Akhtar, Michael Schlichtkrull, Zhijiang Guo et al.

Misinformation is often conveyed in multiple modalities, e.g. a miscaptioned image. Multimodal misinformation is perceived as more credible by humans, and spreads faster than its text-only counterparts. While an increasing body of research investigates automated fact-checking (AFC), previous surveys mostly focus on text. In this survey, we conceptualise a framework for AFC including subtasks unique to multimodal misinformation. Furthermore, we discuss related terms used in different communities and map them to our framework. We focus on four modalities prevalent in real-world fact-checking: text, image, audio, and video. We survey benchmarks and models, and discuss limitations and promising directions for future research

Michael Schlichtkrull, Zhijiang Guo, Andreas Vlachos

Existing datasets for automated fact-checking have substantial limitations, such as relying on artificial claims, lacking annotations for evidence and intermediate reasoning, or including evidence published after the claim. In this paper we introduce AVeriTeC, a new dataset of 4,568 real-world claims covering fact-checks by 50 different organizations. Each claim is annotated with question-answer pairs supported by evidence available online, as well as textual justifications explaining how the evidence combines to produce a verdict. Through a multi-round annotation process, we avoid common pitfalls including context dependence, evidence insufficiency, and temporal leakage, and reach a substantial inter-annotator agreement of $κ=0.619$ on verdicts. We develop a baseline as well as an evaluation scheme for verifying claims through several question-answering steps against the open web.