arXiv subjects

Ioannis Konstas

Publications and source records attributed to Ioannis Konstas.

At least 19 recordsLinked to original sources

arXiv2026

GPTNT: Benchmarking Real-Time Collaboration Between Multimodal Agents on Keep Talking And Nobody Explodes

Multimodal models are increasingly deployed to solve tasks collaboratively with humans or other artificial agents. While existing benchmarks show that they possess the fundamental capabilities, the various conditions that coincide when collaborating---time pressure, information asymmetry, and imperfect communication---have traditionally been studied in isolation. To address this gap, we introduce GPTNT, a benchmark built on the cooperative video game Keep Talking and Nobody Explodes, in which two agents must coordinate to defuse procedurally generated bomb puzzles against a live countdown. One agent has access to the bomb but not the instructions for defusing it; the other holds the instructions but cannot see or manipulate the bomb. Neither agent can succeed alone: the task requires contributions from both, and is solvable only through effective, efficient communication. We remove turn-taking proxies or simplifications, instead requiring agents to act asynchronously and communicate in real time. GPTNT is designed to expose how models collaborate versus how they perform alone: the instruction manual, the partner, or both, can optionally be withheld to surface what a model has memorised versus what it derives in the moment. We demonstrate that GPTNT poses a considerable challenge to the state-of-the-art: not one of the closed- and open-source models we test defuses a single bomb in real time, a bar that human players clear. In a range of controlled experiments, we explore where capabilities break down, identifying critical weaknesses in state tracking, efficient acting within the time budget, handling ambiguity, and error recovery. Since it runs on the real game, GPTNT benefits from procedural generation and inherits a living modding community: as models improve, the benchmark can be evolved to remain challenging, rather than being solved once and retired.

cs.AI↗

arXiv2026

When Chain-of-Thought Fails, the Solution Hides in the Hidden States

Whether intermediate reasoning is computationally useful or merely explanatory depends on whether chain-of-thought (CoT) tokens contain task-relevant information. We present a mechanistic causal analysis of CoT on GSM8K using activation patching: transferring token-level hidden states from a CoT generation to a direct-answer run for the same question, then measuring the effect on final-answer accuracy. Across models, generating after patching yields substantially higher accuracy than both direct-answer prompting and the original CoT trace, revealing that individual CoT tokens can encode sufficient information to recover the correct answer, even when the original trace is incorrect. This task-relevant information is more prevalent in correct than incorrect CoT runs and is unevenly distributed across tokens, concentrating in mid-to-late layers and appearing earlier in the reasoning trace. Moreover, patching language tokens such as verbs and entities carry task-solving information that steers generation toward correct reasoning, whereas mathematical tokens encode answer-proximal content that rarely succeeds. Patched outputs are often shorter and yet exceed the accuracy of a full CoT trace, suggesting complete reasoning chains are not always necessary. Together, these findings demonstrate that CoT encodes recoverable, token-level problem-solving information, offering new insight into how reasoning is represented and where it breaks down.

cs.CL↗

arXiv2026

MoRFI: Monotonic Sparse Autoencoder Feature Identification

Large language models (LLMs) acquire most of their factual knowledge during the pre-training stage, through next token prediction. Subsequent stages of post-training often introduce new facts outwith the parametric knowledge, giving rise to hallucinations. While it has been demonstrated that supervised fine-tuning (SFT) on new knowledge may exacerbate the problem, the underlying mechanisms are still poorly understood. We conduct a controlled fine-tuning experiment, focusing on closed-book QA, and identify latent directions causally implicated in this degradation. Specifically, we fine-tune Llama 3.1 8B, Gemma 2 9B and Mistral 7B v03 on seven controlled mixtures of a single QA dataset, controlling for the percentage of new knowledge and number of training epochs. By measuring performance on the test set, we validate that incrementally introducing new knowledge increases hallucinations, with the effect being more pronounced with prolonged training. We leverage pre-trained sparse autoencoders (SAEs) to analyze residual stream activations across various checkpoints for each model and propose Monotonic Relationship Feature Identification (MoRFI) for capturing causally relevant latents. MoRFI filters SAE features that respond monotonically to controlled fine-tuning data mixtures of a target property. Our findings are consistent with exposure to unknown facts disrupting the model's ability to retrieve stored knowledge along a set of directions in the residual stream. Our pipeline reliably discovers them across distinct models, partially recovering lost knowledge through single-latent interventions.

cs.CL↗

arXiv2026

Retrievit: In-context Retrieval Capabilities of Transformers, State Space Models, and Hybrid Architectures

Transformers excel at in-context retrieval but suffer from quadratic complexity with sequence length, while State Space Models (SSMs) offer efficient linear-time processing but have limited retrieval capabilities. We investigate whether hybrid architectures combining Transformers and SSMs can achieve the best of both worlds on two synthetic in-context retrieval tasks. The first task, n-gram retrieval, requires the model to reproduce an n-gram that succeeds the query within the input sequence. The second task, position retrieval, presents the model with a query token and requires it to perform a two-hop lookup: first locating the corresponding element in the sequence, and then outputting its positional index. Under controlled conditions, we assess data efficiency, length generalization, robustness to out of domain training examples, and learned representations across Transformers, SSMs, and hybrid architectures. We find that hybrid models outperform SSMs and match or exceed Transformers in terms of data efficiency and extrapolation for tasks that require precise information retrieval from the input context. However, Transformers maintain superiority in position retrieval tasks. Through representation analysis, we discover that SSM-based models develop locality-aware embeddings where tokens representing adjacent positions become neighbors in embedding space, forming interpretable structures. This property is absent in Transformers as causal attention is sufficient for acquiring positional associations, and the introduction of positional encoding amplifies this behavior, leading to improvement in data efficiency. SSMs on the other hand update their internal representations incrementally and without positional encodings, are required to learn these associations. Our findings reveal fundamental differences in how Transformers and SSMs, and hybrid models learn positional associations.

cs.AI↗

arXiv2025

Erasing 'Ugly' from the Internet: Propagation of the Beauty Myth in Text-Image Models

Social media has exacerbated the promotion of Western beauty norms, leading to negative self-image, particularly in women and girls, and causing harm such as body dysmorphia. Increasingly content on the internet has been artificially generated, leading to concerns that these norms are being exaggerated. The aim of this work is to study how generative AI models may encode 'beauty' and erase 'ugliness', and discuss the implications of this for society. To investigate these aims, we create two image generation pipelines: a text-to-image model and a text-to-language model-to image model. We develop a structured beauty taxonomy which we use to prompt three language models (LMs) and two text-to-image models to cumulatively generate 5984 images using our two pipelines. We then recruit women and non-binary social media users to evaluate 1200 of the images through a Likert-scale within-subjects study. Participants show high agreement in their ratings. Our results show that 86.5% of generated images depicted people with lighter skin tones, 22% contained explicit content despite Safe for Work (SFW) training, and 74% were rated as being in a younger age demographic. In particular, the images of non-binary individuals were rated as both younger and more hypersexualised, indicating troubling intersectional effects. Notably, prompts encoded with 'negative' or 'ugly' beauty traits (such as "a wide nose") consistently produced higher Not SFW (NSFW) ratings regardless of gender. This work sheds light on the pervasive demographic biases related to beauty standards present in generative AI models -- biases that are actively perpetuated by model developers, such as via negative prompting. We conclude by discussing the implications of this on society, which include pollution of the data streams and active erasure of features that do not fall inside the stereotype of what is considered beautiful by developers.

cs.CV↗

arXiv2025

Can NLP Tackle Hate Speech in the Real World? Stakeholder-Informed Feedback and Survey on Counterspeech

Counterspeech, i.e. the practice of responding to online hate speech, has gained traction in NLP as a promising intervention. While early work emphasised collaboration with non-governmental organisation stakeholders, recent research trends have shifted toward automated pipelines that reuse a small set of legacy datasets, often without input from affected communities. This paper presents a systematic review of 74 NLP studies on counterspeech, analysing the extent to which stakeholder participation influences dataset creation, model development, and evaluation. To complement this analysis, we conducted a participatory case study with five NGOs specialising in online Gender-Based Violence (oGBV), identifying stakeholder-informed practices for counterspeech generation. Our findings reveal a growing disconnect between current NLP research and the needs of communities most impacted by toxic online content. We conclude with concrete recommendations for re-centring stakeholder expertise in counterspeech research.

cs.CL↗

arXiv2025

CROPE: Evaluating In-Context Adaptation of Vision and Language Models to Culture-Specific Concepts

As Vision and Language models (VLMs) are reaching users across the globe, assessing their cultural understanding has become a critical challenge. In this paper, we introduce CROPE, a visual question answering benchmark designed to probe the knowledge of culture-specific concepts and evaluate the capacity for cultural adaptation through contextual information. This allows us to distinguish between parametric knowledge acquired during training and contextual knowledge provided during inference via visual and textual descriptions. Our evaluation of several state-of-the-art open VLMs shows large performance disparities between culture-specific and common concepts in the parametric setting. Moreover, experiments with contextual knowledge indicate that models struggle to effectively utilize multimodal information and bind culture-specific concepts to their depictions. Our findings reveal limitations in the cultural understanding and adaptability of current VLMs that need to be addressed toward more culturally inclusive models.

cs.CL↗

arXiv2024

Investigating the Role of Instruction Variety and Task Difficulty in Robotic Manipulation Tasks

Evaluating the generalisation capabilities of multimodal models based solely on their performance on out-of-distribution data fails to capture their true robustness. This work introduces a comprehensive evaluation framework that systematically examines the role of instructions and inputs in the generalisation abilities of such models, considering architectural design, input perturbations across language and vision modalities, and increased task complexity. The proposed framework uncovers the resilience of multimodal models to extreme instruction perturbations and their vulnerability to observational changes, raising concerns about overfitting to spurious correlations. By employing this evaluation framework on current Transformer-based multimodal models for robotic manipulation tasks, we uncover limitations and suggest future advancements should focus on architectural and training innovations that better integrate multimodal inputs, enhancing a model's generalisation prowess by prioritising sensitivity to input content over incidental correlations.

cs.CL↗

arXiv2024

Reasoning or a Semblance of it? A Diagnostic Study of Transitive Reasoning in LLMs

Evaluating Large Language Models (LLMs) on reasoning benchmarks demonstrates their ability to solve compositional questions. However, little is known of whether these models engage in genuine logical reasoning or simply rely on implicit cues to generate answers. In this paper, we investigate the transitive reasoning capabilities of two distinct LLM architectures, LLaMA 2 and Flan-T5, by manipulating facts within two compositional datasets: QASC and Bamboogle. We controlled for potential cues that might influence the models' performance, including (a) word/phrase overlaps across sections of test input; (b) models' inherent knowledge during pre-training or fine-tuning; and (c) Named Entities. Our findings reveal that while both models leverage (a), Flan-T5 shows more resilience to experiments (b and c), having less variance than LLaMA 2. This suggests that models may develop an understanding of transitivity through fine-tuning on knowingly relevant datasets, a hypothesis we leave to future work.

cs.CL↗

arXiv2024

Re-examining Sexism and Misogyny Classification with Annotator Attitudes

Gender-Based Violence (GBV) is an increasing problem online, but existing datasets fail to capture the plurality of possible annotator perspectives or ensure the representation of affected groups. We revisit two important stages in the moderation pipeline for GBV: (1) manual data labelling; and (2) automated classification. For (1), we examine two datasets to investigate the relationship between annotator identities and attitudes and the responses they give to two GBV labelling tasks. To this end, we collect demographic and attitudinal information from crowd-sourced annotators using three validated surveys from Social Psychology. We find that higher Right Wing Authoritarianism scores are associated with a higher propensity to label text as sexist, while for Social Dominance Orientation and Neosexist Attitudes, higher scores are associated with a negative tendency to do so. For (2), we conduct classification experiments using Large Language Models and five prompting strategies, including infusing prompts with annotator information. We find: (i) annotator attitudes affect the ability of classifiers to predict their labels; (ii) including attitudinal information can boost performance when we use well-structured brief annotator descriptions; and (iii) models struggle to reflect the increased complexity and imbalanced classes of the new label sets.

cs.CL↗

arXiv2024

Voices in a Crowd: Searching for Clusters of Unique Perspectives

Language models have been shown to reproduce underlying biases existing in their training data, which is the majority perspective by default. Proposed solutions aim to capture minority perspectives by either modelling annotator disagreements or grouping annotators based on shared metadata, both of which face significant challenges. We propose a framework that trains models without encoding annotator metadata, extracts latent embeddings informed by annotator behaviour, and creates clusters of similar opinions, that we refer to as voices. Resulting clusters are validated post-hoc via internal and external quantitative metrics, as well a qualitative analysis to identify the type of voice that each cluster represents. Our results demonstrate the strong generalisation capability of our framework, indicated by resulting clusters being adequately robust, while also capturing minority perspectives based on different demographic factors throughout two distinct datasets.

cs.CL↗

arXiv2024

Enhancing Continual Learning in Visual Question Answering with Modality-Aware Feature Distillation

Continual learning focuses on incrementally training a model on a sequence of tasks with the aim of learning new tasks while minimizing performance drop on previous tasks. Existing approaches at the intersection of Continual Learning and Visual Question Answering (VQA) do not study how the multimodal nature of the input affects the learning dynamics of a model. In this paper, we demonstrate that each modality evolves at different rates across a continuum of tasks and that this behavior occurs in established encoder-only models as well as modern recipes for developing Vision & Language (VL) models. Motivated by this observation, we propose a modality-aware feature distillation (MAFED) approach which outperforms existing baselines across models of varying scale in three multimodal continual learning settings. Furthermore, we provide ablations showcasing that modality-aware distillation complements experience replay. Overall, our results emphasize the importance of addressing modality-specific dynamics to prevent forgetting in multimodal continual learning.

cs.CV↗

arXiv2024

AlanaVLM: A Multimodal Embodied AI Foundation Model for Egocentric Video Understanding

AI personal assistants deployed via robots or wearables require embodied understanding to collaborate with humans effectively. However, current Vision-Language Models (VLMs) primarily focus on third-person view videos, neglecting the richness of egocentric perceptual experience. To address this gap, we propose three key contributions. First, we introduce the Egocentric Video Understanding Dataset (EVUD) for training VLMs on video captioning and question answering tasks specific to egocentric videos. Second, we present AlanaVLM, a 7B parameter VLM trained using parameter-efficient methods on EVUD. Finally, we evaluate AlanaVLM's capabilities on OpenEQA, a challenging benchmark for embodied video question answering. Our model achieves state-of-the-art performance, outperforming open-source models including strong Socratic models using GPT-4 as a planner by 3.6%. Additionally, we outperform Claude 3 and Gemini Pro Vision 1.0 and showcase competitive results compared to Gemini Pro 1.5 and GPT-4V, even surpassing the latter in spatial reasoning. This research paves the way for building efficient VLMs that can be deployed in robots or wearables, leveraging embodied video understanding to collaborate seamlessly with humans in everyday tasks, contributing to the next generation of Embodied AI.

cs.CV↗

arXiv2024

Task Formulation Matters When Learning Continually: A Case Study in Visual Question Answering

Continual learning aims to train a model incrementally on a sequence of tasks without forgetting previous knowledge. Although continual learning has been widely studied in computer vision, its application to Vision+Language tasks is not that straightforward, as settings can be parameterized in multiple ways according to their input modalities. In this paper, we present a detailed study of how different settings affect performance for Visual Question Answering. We first propose three plausible task formulations and demonstrate their impact on the performance of continual learning algorithms. We break down several factors of task similarity, showing that performance and sensitivity to task order highly depend on the shift of the output distribution. We also investigate the potential of pretrained models and compare the robustness of transformer models with different visual embeddings. Finally, we provide an analysis interpreting model representations and their impact on forgetting. Our results highlight the importance of stabilizing visual representations in deeper layers.

cs.LG↗

arXiv2023

Visually Grounded Language Learning: a review of language games, datasets, tasks, and models

In recent years, several machine learning models have been proposed. They are trained with a language modelling objective on large-scale text-only data. With such pretraining, they can achieve impressive results on many Natural Language Understanding and Generation tasks. However, many facets of meaning cannot be learned by ``listening to the radio" only. In the literature, many Vision+Language (V+L) tasks have been defined with the aim of creating models that can ground symbols in the visual modality. In this work, we provide a systematic literature review of several tasks and models proposed in the V+L field. We rely on Wittgenstein's idea of `language games' to categorise such tasks into 3 different families: 1) discriminative games, 2) generative games, and 3) interactive games. Our analysis of the literature provides evidence that future work should be focusing on interactive games where communication in Natural Language is important to resolve ambiguities about object referents and action plans and that physical embodiment is essential to understand the semantics of situations and events. Overall, these represent key requirements for developing grounded meanings in neural models.

cs.CL↗

arXiv2023

Multitask Multimodal Prompted Training for Interactive Embodied Task Completion

Interactive and embodied tasks pose at least two fundamental challenges to existing Vision & Language (VL) models, including 1) grounding language in trajectories of actions and observations, and 2) referential disambiguation. To tackle these challenges, we propose an Embodied MultiModal Agent (EMMA): a unified encoder-decoder model that reasons over images and trajectories, and casts action prediction as multimodal text generation. By unifying all tasks as text generation, EMMA learns a language of actions which facilitates transfer across tasks. Different to previous modular approaches with independently trained components, we use a single multitask model where each task contributes to goal completion. EMMA performs on par with similar models on several VL benchmarks and sets a new state-of-the-art performance (36.81% success rate) on the Dialog-guided Task Completion (DTC), a benchmark to evaluate dialog-guided agents in the Alexa Arena

cs.LG↗

arXiv2023

Mind the Labels: Describing Relations in Knowledge Graphs With Pretrained Models

Pretrained language models (PLMs) for data-to-text (D2T) generation can use human-readable data labels such as column headings, keys, or relation names to generalize to out-of-domain examples. However, the models are well-known in producing semantically inaccurate outputs if these labels are ambiguous or incomplete, which is often the case in D2T datasets. In this paper, we expose this issue on the task of descibing a relation between two entities. For our experiments, we collect a novel dataset for verbalizing a diverse set of 1,522 unique relations from three large-scale knowledge graphs (Wikidata, DBPedia, YAGO). We find that although PLMs for D2T generation expectedly fail on unclear cases, models trained with a large variety of relation labels are surprisingly robust in verbalizing novel, unseen relations. We argue that using data with a diverse set of clear and meaningful labels is key to training D2T generation systems capable of generalizing to novel domains.

cs.CL↗

arXiv2023

No that's not what I meant: Handling Third Position Repair in Conversational Question Answering

The ability to handle miscommunication is crucial to robust and faithful conversational AI. People usually deal with miscommunication immediately as they detect it, using highly systematic interactional mechanisms called repair. One important type of repair is Third Position Repair (TPR) whereby a speaker is initially misunderstood but then corrects the misunderstanding as it becomes apparent after the addressee's erroneous response. Here, we collect and publicly release Repair-QA, the first large dataset of TPRs in a conversational question answering (QA) setting. The data is comprised of the TPR turns, corresponding dialogue contexts, and candidate repairs of the original turn for execution of TPRs. We demonstrate the usefulness of the data by training and evaluating strong baseline models for executing TPRs. For stand-alone TPR execution, we perform both automatic and human evaluations on a fine-tuned T5 model, as well as OpenAI's GPT-3 LLMs. Additionally, we extrinsically evaluate the LLMs' TPR processing capabilities in the downstream conversational QA task. The results indicate poor out-of-the-box performance on TPR's by the GPT-3 models, which then significantly improves when exposed to Repair-QA.

cs.CL↗