The tendency for hallucination in current large language models (LLMs) negatively impacts dialogue systems. Such hallucinations produce factually incorrect responses that may mislead users and undermine system trust. Existing refinement methods for dialogue systems typically operate at the response level, overlooking the fact that a single response may contain multiple verifiable or unverifiable facts. To address this gap, we propose Fine-Refine, a fine-grained refinement framework that decomposes responses into atomic units, verifies each unit using external knowledge, assesses fluency via perplexity, and iteratively corrects granular errors. We evaluate factuality across the HybriDialogue and OpendialKG datasets in terms of factual accuracy (fact score) and coverage (Not Enough Information Proportion), and experiments show that Fine-Refine substantially improves factuality, achieving up to a 7.63-point gain in dialogue fact score, with a small trade-off in dialogue quality.

We evaluate cognitive impairment (CI) classification from transcripts of speech in English, Slovene, and Korean. We compare zero-shot large language models (LLMs) used as direct classifiers under three input settings – transcript-only, linguistic-features-only, and combined – with supervised tabular approaches trained under a leave-one-out protocol. The tabular models operate on engineered linguistic features, transcript embeddings, and early or late fusion of both modalities. Across languages, zero-shot LLMs provide competitive no-training baselines, but supervised tabular models generally perform better, particularly when engineered linguistic features are included and combined with embeddings. Few-shot experiments focusing on embeddings indicate that the value of limited supervision is language-dependent, with some languages benefiting substantially from additional labelled examples while others remain constrained without richer feature representations. Overall, the results suggest that, in small-data CI detection, structured linguistic signals and simple fusion-based classifiers remain strong and reliable signals.

Environmental, Social, and Governance (ESG) considerations are increasingly integral to assessing corporate performance, reputation, and long-term sustainability. Yet, reliable ESG ratings remain limited for smaller companies and emerging markets. We introduce the first publicly available Slovene ESG sentiment dataset and a suite of models for automatic ESG sentiment detection. The dataset, derived from the MaCoCu Slovene news collection, combines large language model (LLM)-assisted filtering with human annotation of company-related ESG content. We evaluate the performance of monolingual (SloBERTa) and multilingual (XLM-R) models, embedding-based classifiers (TabPFN), hierarchical ensemble architectures, and large language models. Results show that LLMs achieve the strongest performance on Environmental (Gemma3-27B, F1-macro: 0.61) and Social aspects (gpt-oss 20B, F1-macro: 0.45), while fine-tuned SloBERTa is the best model on Governance classification (F1-macro: 0.54). We then show in a small case study how the best-preforming classifier (gpt-oss) can be applied to investigate ESG aspects for selected companies across a long time frame.

The current dominance of large language models in natural language processing is based on their contextual awareness. For text classification, text representation models, such as ELMo, BERT, and BERT derivatives, are typically fine-tuned for a specific problem. Most existing work focuses on English; in contrast, we present a large-scale multilingual empirical comparison of several monolingual and multilingual ELMo and BERT models using 14 classification tasks in nine languages. The results show, that the choice of best model largely depends on the task and language used, especially in a cross-lingual setting. In monolingual settings, monolingual BERT models tend to perform the best among BERT models. Among ELMo models, the ones trained on large corpora dominate. Cross-lingual knowledge transfer is feasible on most tasks already in a zero-shot setting without losing much performance.

We introduce ClarQ4LLM, an evaluation framework consisting of bilingual English-Chinese conversation tasks, conversational agents and evaluation metrics, designed to serve as a strong benchmark for assessing agents' ability to ask clarification questions in task-oriented dialogues. The benchmark includes 31 different task types, each with 10 unique dialogue scenarios between information seeker and provider agents. The scenarios require the seeker to ask questions to resolve uncertainty and gather necessary information to complete tasks. Unlike traditional benchmarks that evaluate agents based on fixed dialogue content, ClarQ4LLM includes a provider conversational agent to replicate the original human provider in the benchmark. This allows both current and future seeker agents to test their ability to complete information gathering tasks through dialogue by directly interacting with our provider agent. In tests, LLAMA3.1 405B seeker agent managed a maximum success rate of only 60.05%, showing that ClarQ4LLM presents a strong challenge for future research(1).

Large Language Models (LLMs) are known to produce hallucinations - factually incorrect or fabricated information - which poses significant challenges for many Natural Language Processing (NLP) applications, such as dialogue systems. As a result, detecting hallucinations has become a critical area of research. Current approaches to hallucination detection in dialogue systems primarily focus on verifying the factual consistency of generated responses. However, these responses often contain a mix of accurate, inaccurate or unverifiable facts, making one factual label overly simplistic and coarse-grained. In this paper, we introduce a benchmark, FineDialFact, for fine-grained dialogue fact verification, which involves verifying atomic facts extracted from dialogue responses. To support this, we construct a dataset based on publicly available dialogue datasets and evaluate it using various baseline methods. Experimental results demonstrate that methods incorporating Chain-of-Thought (CoT) reasoning can enhance performance in dialogue fact verification. Despite this, the best F1-score achieved on the HybriDialogue, an open-domain dialogue dataset, is only 0.75, indicating that the benchmark remains a challenging task for future research. Our dataset and code will be public on GitHub.

Large Language Model (LLM)-based agents can solve complex procedural tasks by interacting with environments over multiple turns, but this ability typically depends on large models, long contexts, and repeated inference calls. This makes advanced memory-augmented agents difficult to deploy on resource-constrained devices. We introduce DuoMem, a dual-space distillation framework that transfers procedural problem-solving ability from a large teacher model to compact student models. DuoMem distils in two complementary spaces: (1)context-space distillation, which replaces student-generated memories with higher-quality teacher-generated procedural memories prepended to the student's input, and (2)parameter-space distillation, which fine-tunes lightweight LoRA adapters on successful teacher trajectories. Evaluated on ALFWorld, a challenging embodied decision-making benchmark, DuoMem boosts a 4B-parameter model from 4.3

In the domain of text-to-image generative models, biases inherent in training datasets often propagate into generated content, posing significant ethical challenges, particularly in socially sensitive contexts. We introduce FairCoT, a novel framework that enhances fairness in text to image models through Chain of Thought (CoT) reasoning within multimodal generative large language models. FairCoT employs iterative CoT refinement to systematically mitigate biases, and dynamically adjusts textual prompts in real time, ensuring diverse and equitable representation in generated images. By integrating iterative reasoning processes, FairCoT addresses the limitations of zero shot CoT in sensitive scenarios, balancing creativity with ethical responsibility. Experimental evaluations across popular text-to-image systems including DALLE and various Stable Diffusion variants, demonstrate that FairCoT significantly enhances fairness and diversity without sacrificing image quality or semantic fidelity. By combining robust reasoning, lightweight deployment, and extensibility to multiple models, FairCoT represents a promising step toward more socially responsible and transparent AI driven content generation.

Large Language Models (LLMs) have demonstrated remarkable capabilities in comprehending and analyzing lengthy sequential inputs, owing to their extensive context windows that allow processing millions of tokens in a single forward pass. However, this paper uncovers a surprising limitation: LLMs fall short when handling long input sequences. We investigate this issue using three datasets and two tasks (sentiment analysis and news categorization) across various LLMs, including Claude 3, Gemini Pro, GPT 3.5 Turbo, Llama 3 Instruct, and Mistral Instruct models. To address this limitation, we propose and evaluate ad-hoc solutions that substantially enhance LLMs' performance on long input sequences by up to 50 and 50

Amid ongoing efforts to develop extremely large, multimodal models, there is increasing interest in efficient Small Language Models (SLMs) that can operate without reliance on large data-centre infrastructure. However, recent SLMs (e.g., LLaMA or Phi) with up to three billion parameters are predominantly trained in high-resource languages, such as English, which limits their applicability to industries that require robust NLP solutions for less-represented languages and low-resource settings, particularly those requiring low latency and adaptability to evolving label spaces. This paper examines a retrieval-based approach to multi-label text classification (MLC) for a media monitoring dataset, with a particular focus on less-represented languages, such as Slovene. This dataset presents an extreme MLC challenge, with instances labelled using up to twelve thousand categories. The proposed method, which combines retrieval with computationally efficient prediction, effectively addresses challenges related to multilinguality, resource constraints, and frequent label changes. We adopt a model-agnostic approach that does not rely on a specific model architecture or language selection. Our results demonstrate that techniques from the extreme multi-label text classification (XMC) domain outperform traditional Transformer-based encoder models, particularly in handling dynamic label spaces without requiring continuous fine-tuning. Additionally, we highlight the effectiveness of this approach in scenarios involving rare labels, where baseline models struggle with generalisation.

Test-time Reinforcement Learning (TTRL) has shown promise in adapting foundation models for complex tasks at test-time, resulting in large performance improvements. TTRL leverages an elegant two-phase sampling strategy: first, multi-sampling derives a pseudo-label via majority voting, while subsequent downsampling and reward-based fine-tuning encourages the model to explore and learn diverse valid solutions, with the pseudo-label modulating the reward signal. Meanwhile, in-context learning has been widely explored at inference time and demonstrated the ability to enhance model performance without weight updates. However, TTRL's two-phase sampling strategy under-utilizes contextual guidance, which can potentially improve pseudo-label accuracy in the initial exploitation phase while regulating exploration in the second. To address this, we propose context-guided TTRL (CG-TTRL), integrating context dynamically into both sampling phases and propose a method for efficient context selection for on-device applications. Our evaluations on mathematical and scientific QA benchmarks show CG-TTRL outperforms TTRL (e.g. additional 7

Evaluating text-to-image (T2I) systems requires judging not only whether an image matches a prompt, but also whether socially salient attributes are represented faithfully and without unsupported inference. Existing automated evaluators typically rely on face-centric recognizers or contrastive image–text similarity, which provide limited diagnostic feedback and often force predictions even when visual evidence is ambiguous or absent. For fairness-sensitive attributes such as religion and disability, where cues may be contextual, indirect, or intentionally unspecified, these evaluators can therefore miss failure modes that careful human reviewers would notice. We introduce FairJudge, an abstention-aware evaluation protocol that uses instruction-following multimodal LLMs as structured judges for social-attribute prediction, profession grounding, and prompt–image alignment. The protocol constrains outputs to closed label sets, requires visible-evidence rationales, supports an explicit unspecified decision when cues are insufficient, and maps rubric-based alignment judgments to [-1,1]. These constraints turn MLLM judging from open-ended assessment into a parseable, auditable evaluation procedure. Across four attribute-prediction benchmarks and three profession/alignment benchmarks, FairJudge outperforms or complements CLIP, DeepFace, VIEScore, and VQAScore. Ablations show that closed labels, abstention, and evidence reporting are central to reliability. We further introduce DIVERSIFY and DIVERSIFY-Professions, two context-rich resources for evaluating social representation and profession grounding beyond face-visible or iconic cues. We release code, prompts, datasets, parser logs, and per-image judge outputs to support reproducible auditing.

This Ph.D. proposal introduces a plan to develop a computational framework to identify Self-aspects in text. The Self is a multifaceted construct and it is reflected in language. While it is described across disciplines like cognitive science and phenomenology, it remains underexplored in natural language processing (NLP). Many of the aspects of the Self align with psychological and other well-researched phenomena (e.g., those related to mental health), highlighting the need for systematic NLP-based analysis. In line with this, we plan to introduce an ontology of Self-aspects and a gold-standard annotated dataset. Using this foundation, we will develop and evaluate conventional discriminative models, generative large language models, and embedding-based retrieval approaches against four main criteria: interpretability, ground-truth adherence, accuracy, and computational efficiency. Top-performing models will be applied in case studies in mental health and empirical phenomenology.

In this work we examine LLMs' ability to ask clarification questions in task-oriented dialogues that follow the asynchronous instruction-giver/instruction-follower format. We present a new corpus that combines two existing annotations of the Minecraft Dialogue Corpus – one for reference and ambiguity in reference, and one for SDRT including clarifications – into a single common format providing the necessary information to experiment with clarifications and their relation to ambiguity. With this corpus we compare LLM actions with original human-generated clarification questions, examining how both humans and LLMs act in the case of ambiguity. We find that there is only a weak link between ambiguity and humans producing clarification questions in these dialogues, and low correlation between humans and LLMs. Humans hardly ever produce clarification questions for referential ambiguity, but often do so for task-based uncertainty. Conversely, LLMs produce more clarification questions for referential ambiguity, but less so for task uncertainty. We question if LLMs' ability to ask clarification questions is predicated on their recent ability to simulate reasoning, and test this with different reasoning approaches, finding that reasoning does appear to increase question frequency and relevancy.

Multilingual vision-language models (VLMs) promise universal image-text retrieval, yet their social biases remain underexplored. We perform the first systematic audit of four public multilingual CLIP variants: M-CLIP, NLLB-CLIP, CAPIVARA-CLIP, and the debiased SigLIP-2, covering ten languages that differ in resource availability and morphological gender marking. Using balanced subsets of FairFace and the PATA stereotype suite in a zero-shot setting, we quantify race and gender bias and measure stereotype amplification. Contrary to the intuition that multilinguality mitigates bias, every model exhibits stronger gender skew than its English-only baseline. CAPIVARA-CLIP shows its largest biases precisely in the low-resource languages it targets, while the shared encoder of NLLB-CLIP and SigLIP-2 transfers English gender stereotypes into gender-neutral languages; loosely coupled encoders largely avoid this leakage. Although SigLIP-2 reduces agency and communion skews, it inherits – and in caption-sparse contexts (e.g., Xhosa) amplifies – the English anchor's crime associations. Highly gendered languages consistently magnify all bias types, yet gender-neutral languages remain vulnerable whenever cross-lingual weight sharing imports foreign stereotypes. Aggregated metrics thus mask language-specific hot spots, underscoring the need for fine-grained, language-aware bias evaluation in future multilingual VLM research.

Vision-language models (VLMs) deliver strong zero-shot recognition but frequently inherit social biases from their training data. We systematically disentangle three design factors – model size, training-data scale, and training-data source – by comparing CLIP and OpenCLIP, two models that share an identical contrastive objective yet differ in encoder width and in the image-text corpora on which they are pre-trained (400M proprietary pairs vs. 400M/2B LAION). Across balanced face-analysis benchmarks, enlarging the encoder reduces gender skew in CLIP but amplifies both gender and racial skew in OpenCLIP; increasing the LAION corpus from 400M to 2B further increases OpenCLIP bias. At matched model and data budgets, substituting proprietary data with LAION improves gender fairness while increasing racial skew, underscoring data source as the primary driver of bias patterns. We also evaluate three post-hoc, test-time debiasing strategies – Bias Prompts, Prompt Array, and SANER. Debiasing reduces but does not eliminate harm, and its effectiveness is source- and size-dependent: Bias Prompts most effectively reduce gender skew in CLIP at smaller model sizes, whereas Prompt Array and SANER more reliably reduce racial skew in OpenCLIP; scaling LAION reconfigures which method is most fair. Taken together, these findings challenge the assumption that bigger models or datasets are automatically fairer and foreground training data source as the key determinant of both bias and mitigation efficacy. We release code and evaluation scripts to enable transparent, reproducible auditing of future VLMs.

From natural language processing to vision, Scaled Dot Product Attention (SDPA) is the backbone of most modern deep learning applications. Unfortunately, its memory and computational requirements can be prohibitive in low-resource settings. In this paper, we improve its efficiency without sacrificing its versatility. We propose three attention variants where we remove consecutive linear transformations or add a novel one, and evaluate them on a range of standard NLP and vision tasks. Our proposed models are substantially lighter than standard SDPA (and have 25-50 show that the performance cost of these changes is negligible relative to size reduction and that in one case (Super Attention) we succeed in outperforming SDPA by up to 10

Large Language Models (LLMs) succeed in many natural language processing tasks. However, their tendency to hallucinate - generate plausible but inconsistent or factually incorrect text - can cause significant problems in certain tasks, including response generation in dialogue. To mitigate this issue, we propose two novel graph knowledge-augmented frameworks, Dialogue Response Generation via Textualised Graphs (TG-DRG) and Graph-Aware Dialogue Response Generation (GA-DRG), which combine reasoning-guided dialogue reformulation, dialogue sense knowledge selection, and graph-enhanced response generation to improve the factuality of dialogue responses. To evaluate the factuality of generated responses, we propose a dialogue fact score that addresses the limitations of existing fact-score methods in dialogue settings, providing a more reliable assessment of factual consistency. We evaluate our methods using different baselines on the OpendialKG and HybriDialogue datasets. Our methods noticeably improve factuality compared to other graph knowledge-augmentation baselines, including the state-of-the-art G-retriever, achieving improvements of 3.47