arXiv subjects

Alessandro Suglia

Publications and source records attributed to Alessandro Suglia.

At least 19 recordsLinked to original sources

arXiv2026

Investigating Knowledge Transfer Across Interactive Dialogue Games

Dialogue games represent a challenging setting where complex cognitive skills are required to accomplish tasks while coordinating with other players. Considering that language represents an interface for both understanding the game rules and executing actions, it is reasonable to assume that training on a specific language game will enhance specific capabilities that might be relevant for other tasks as well. Motivated by this rationale, in this paper, we investigate how knowledge transfers across different dialogue games. We study transferability by finetuning LLM models on games from the clembench suite (Chalamalasetti et al., 2023) and performing two analyses: i) we derive a task-transferability graph using a binary integer optimization program from Zamir et al. (2018), using task performance as the main metric; and ii) we compute task vectors (Ilharco et al., 2022) for each game to study similarities across finetuned models and their task transferability. In our first analysis, we find that some games benefit more from transfer than finetuning, and that the visuospatial family (e.g., exploration games) transfers best. With our task vector analysis instead, we find that similarity-based approaches capture game-role relationships but almost no transferability patterns, suggesting that more complex metrics are required.

cs.CL↗

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

Do Composed Image Retrieval Benchmarks Require Multimodal Composition?

Composed Image Retrieval (CIR) is a multimodal retrieval task where a query consists of a reference image and a textual modification, and the goal is to retrieve a target image satisfying both. In principle, strong performance on CIR benchmarks is assumed to require multimodal composition, i.e., combining complementary information from reference image and textual modification. In this work, we show that this assumption does not always hold. Across four widely used CIR benchmarks and eleven Generalist Multimodal Embedding models, a large fraction of queries can be solved using a single modality (from 32.2% to 83.6%), revealing pervasive unimodal shortcuts. Thus, high CIR performance can arise from unimodal signals rather than true multimodal composition. To better understand this issue, we perform a two-stage audit. First, we identify shortcut-solvable queries through cross-model analysis. Second, we conduct human validation on 4,741 shortcut-free queries, of which only 1,689 are well-formed, with common issues including ambiguous edits and mismatched targets. Re-evaluating models on this validated subset reveals qualitatively different behaviour: queries can no longer be solved with a single modality, and successful retrieval requires combining both inputs. While accuracy decreases, reliance on multimodal information increases. Overall, current CIR benchmarks conflate shortcut-solvable, noisy, and genuinely compositional queries, leading to an overestimation of model capability in multimodal composition.

cs.CV↗

arXiv2026

MIXAR: Scaling Autoregressive Pixel-based Language Models to Multiple Languages and Scripts

Pixel-based language models are gaining momentum as alternatives to traditional token-based approaches, promising to circumvent tokenization challenges. However, the inherent perceptual diversity across languages poses a significant hurdle for multilingual generalization in pixel space. This paper introduces MIXAR, the first generative pixel-based language model trained on eight different languages utilizing a range of different scripts. We empirically evaluate MIXAR against previous pixel-based models as well as comparable tokenizer-based models, demonstrating substantial performance improvement on discriminative and generative multilingual tasks. Additionally, we show how MIXAR is robust to languages never seen during the training. These results are further strengthened when scaling the model to 0.5B parameters which not only improves its capabilities in generative tasks like LAMBADA but also its robustness when challenged with input perturbations such as orthographic attacks.

cs.CL↗

arXiv2026

AgriPath: A Systematic Exploration of Architectural Trade-offs for Crop Disease Classification

Reliable crop disease detection requires models that perform consistently across diverse acquisition conditions, yet existing evaluations often focus on single architectural families or lab-generated datasets. This work presents a systematic empirical comparison of three model paradigms for fine-grained crop disease classification: Convolutional Neural Networks (CNNs), contrastive Vision-Language Models (VLMs), and generative VLMs. To enable controlled analysis of domain effects, we introduce AgriPath-LF16, a benchmark of 111k images spanning 16 crops and 41 diseases with explicit separation between laboratory and field imagery, alongside a balanced 30k subset for standardised training and evaluation. We train and evaluate all models under unified protocols across full, lab-only, and field-only training regimes using macro-F1 and Parse Success Rate (PSR) to account for generative reliability (i.e., output parsability measured via PSR). The results reveal distinct performance profiles: CNNs achieve the highest accuracy on in-domain imagery but exhibit pronounced degradation under domain shift; contrastive VLMs provide a robust and parameter-efficient alternative with competitive cross-domain performance; generative VLMs demonstrate the strongest resilience to distributional variation, albeit with additional failure modes stemming from free-text generation. These findings highlight that architectural choice should be guided by deployment context rather than aggregate performance alone.

cs.CV↗

arXiv2026

VLM-RobustBench: A Comprehensive Benchmark for Robustness of Vision-Language Models

Vision-language models (VLMs) achieve strong performance on standard, high-quality datasets, but we still do not fully understand how they perform under real-world image distortions. We present VLM-RobustBench, a benchmark spanning 49 augmentation types across noise, blur, weather, digital, and geometric perturbations, evaluated under graded severities (low/mid/high) and binary transforms, yielding 133 corrupted settings. We evaluate VLMs from four families (Qwen, InternVL, Molmo, Gemma) on two complementary benchmarks: MMBench (visually grounded) and MMMU-Pro (reasoning-oriented). Our results reveal that visual severity is a weak predictor of difficulty: low-severity spatial perturbations often degrade performance more than visually severe photometric corruptions. In particular, low-severity glass_blur reduces MMBench accuracy by about 8 pp on average across models, while the largest drops arise from resampling and geometric distortions (e.g., upsample, elastic_transform), reaching up to 34 pp. Overall, our findings suggest current VLMs are semantically strong but spatially fragile, motivating the definition of novel robustness evaluation protocols and training regimes that emphasize resampling and geometric invariances.

cs.CV↗

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↗

arXiv2026

Same Answer, Different Representations: Hidden instability in VLMs

The robustness of Vision Language Models (VLMs) is commonly assessed through output-level invariance, implicitly assuming that stable predictions reflect stable multimodal processing. In this work, we argue that this assumption is insufficient. We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness (spatial consistency of vision tokens), alongside standard label-based metrics. Applying this framework to modern VLMs across the SEEDBench, MMMU, and POPE datasets reveals three distinct failure modes. First, models frequently preserve predicted answers while undergoing substantial internal representation drift; for perturbations such as text overlays, this drift approaches the magnitude of inter-image variability, indicating that representations move to regions typically occupied by unrelated inputs despite unchanged outputs. Second, robustness does not improve with scale; larger models achieve higher accuracy but exhibit equal or greater sensitivity, consistent with sharper yet more fragile decision boundaries. Third, we find that perturbations affect tasks differently: they harm reasoning when they disrupt how models combine coarse and fine visual cues, but on the hallucination benchmarks, they can reduce false positives by making models generate more conservative answers.

cs.AI↗

arXiv2025

FOSSIL: Harnessing Feedback on Suboptimal Samples for Data-Efficient Generalisation with Imitation Learning for Embodied Vision-and-Language Tasks

Current approaches to embodied AI tend to learn policies from expert demonstrations. However, without a mechanism to evaluate the quality of demonstrated actions, they are limited to learning from optimal behaviour, or they risk replicating errors and inefficiencies. While reinforcement learning offers one alternative, the associated exploration typically results in sacrificing data efficiency. This work explores how agents trained with imitation learning can learn robust representations from both optimal and suboptimal demonstrations when given access to constructive language feedback as a means to contextualise different modes of behaviour. We directly provide language feedback embeddings as part of the input sequence into a Transformer-based policy, and optionally complement the traditional next action prediction objective with auxiliary self-supervised learning objectives for feedback prediction. We test our approach on a range of embodied Vision-and-Language tasks in our custom BabyAI-XGen environment and show significant improvements in agents' compositional generalisation abilities and robustness, suggesting that our data-efficient method allows models to successfully convert suboptimal behaviour into learning opportunities. Overall, our results suggest that language feedback is a competitive and intuitive alternative to intermediate scalar rewards for language-specified embodied tasks.

cs.CL↗

arXiv2025

Improving Cooperation in Collaborative Embodied AI

The integration of Large Language Models (LLMs) into multiagent systems has opened new possibilities for collaborative reasoning and cooperation with AI agents. This paper explores different prompting methods and evaluates their effectiveness in enhancing agent collaborative behaviour and decision-making. We enhance CoELA, a framework designed for building Collaborative Embodied Agents that leverage LLMs for multi-agent communication, reasoning, and task coordination in shared virtual spaces. Through systematic experimentation, we examine different LLMs and prompt engineering strategies to identify optimised combinations that maximise collaboration performance. Furthermore, we extend our research by integrating speech capabilities, enabling seamless collaborative voice-based interactions. Our findings highlight the effectiveness of prompt optimisation in enhancing collaborative agent performance; for example, our best combination improved the efficiency of the system running with Gemma3 by 22% compared to the original CoELA system. In addition, the speech integration provides a more engaging user interface for iterative system development and demonstrations.

cs.AI↗

arXiv2025

Triangulating LLM Progress through Benchmarks, Games, and Cognitive Tests

We examine three evaluation paradigms: standard benchmarks (e.g., MMLU and BBH), interactive games (e.g., Signalling Games or Taboo), and cognitive tests (e.g., for working memory or theory of mind). First, we investigate which of the former two-benchmarks or games-is most effective at discriminating LLMs of varying quality. Then, inspired by human cognitive assessments, we compile a suite of targeted tests that measure cognitive abilities deemed essential for effective language use, and we investigate their correlation with model performance in benchmarks and games. Our analyses reveal that interactive games are superior to standard benchmarks in discriminating models. Causal and logical reasoning correlate with both static and interactive tests, while differences emerge regarding core executive functions and social/emotional skills, which correlate more with games. We advocate for the development of new interactive benchmarks and targeted cognitive tasks inspired by assessing human abilities but designed specifically for LLMs.

cs.CL↗

arXiv2025

Playpen: An Environment for Exploring Learning Through Conversational Interaction

Interaction between learner and feedback-giver has come into focus recently for post-training of Large Language Models (LLMs), through the use of reward models that judge the appropriateness of a model's response. In this paper, we investigate whether Dialogue Games -- goal-directed and rule-governed activities driven predominantly by verbal actions -- can also serve as a source of feedback signals for learning. We introduce Playpen, an environment for off- and online learning through Dialogue Game self-play, and investigate a representative set of post-training methods: supervised fine-tuning; direct alignment (DPO); and reinforcement learning with GRPO. We experiment with post-training a small LLM (Llama-3.1-8B-Instruct), evaluating performance on unseen instances of training games as well as unseen games, and on standard benchmarks. We find that imitation learning through SFT improves performance on unseen instances, but negatively impacts other skills, while interactive learning with GRPO shows balanced improvements without loss of skills. We release the framework and the baseline training setups to foster research in the promising new direction of learning in (synthetic) interaction.

cs.CL↗

arXiv2025

VoyagerVision: Investigating the Role of Multi-modal Information for Open-ended Learning Systems

Open-endedness is an active field of research in the pursuit of capable Artificial General Intelligence (AGI), allowing models to pursue tasks of their own choosing. Simultaneously, recent advancements in Large Language Models (LLMs) such as GPT-4o [9] have allowed such models to be capable of interpreting image inputs. Implementations such as OMNI-EPIC [4] have made use of such features, providing an LLM with pixel data of an agent's POV to parse the environment and allow it to solve tasks. This paper proposes that providing these visual inputs to a model gives it greater ability to interpret spatial environments, and as such, can increase the number of tasks it can successfully perform, extending its open-ended potential. To this aim, this paper proposes VoyagerVision -- a multi-modal model capable of creating structures within Minecraft using screenshots as a form of visual feedback, building on the foundation of Voyager. VoyagerVision was capable of creating an average of 2.75 unique structures within fifty iterations of the system, as Voyager was incapable of this, it is an extension in an entirely new direction. Additionally, in a set of building unit tests VoyagerVision was successful in half of all attempts in flat worlds, with most failures arising in more complex structures. Project website is available at https://esmyth-dev.github.io/VoyagerVision.github.io/

cs.AI↗

arXiv2025

Movie Facts and Fibs (MF$^2$): A Benchmark for Long Movie Understanding

Despite recent progress in vision-language models (VLMs), holistic understanding of long-form video content remains a significant challenge, partly due to limitations in current benchmarks. Many focus on peripheral, ``needle-in-a-haystack'' details, encouraging context-insensitive retrieval over deep comprehension. Others rely on large-scale, semi-automatically generated questions (often produced by language models themselves) that are easier for models to answer but fail to reflect genuine understanding. In this paper, we introduce MF$^2$, a new benchmark for evaluating whether models can comprehend, consolidate, and recall key narrative information from full-length movies (50-170 minutes long). MF$^2$ includes over 50 full-length, open-licensed movies, each paired with manually constructed sets of claim pairs -- one true (fact) and one plausible but false (fib), totalling over 850 pairs. These claims target core narrative elements such as character motivations and emotions, causal chains, and event order, and refer to memorable moments that humans can recall without rewatching the movie. Instead of multiple-choice formats, we adopt a binary claim evaluation protocol: for each pair, models must correctly identify both the true and false claims. This reduces biases like answer ordering and enables a more precise assessment of reasoning. Our experiments demonstrate that both open-weight and closed state-of-the-art models fall well short of human performance, underscoring the relative ease of the task for humans and their superior ability to retain and reason over critical narrative information -- an ability current VLMs lack.

cs.CV↗

arXiv2025

LLMs instead of Human Judges? A Large Scale Empirical Study across 20 NLP Evaluation Tasks

There is an increasing trend towards evaluating NLP models with LLMs instead of human judgments, raising questions about the validity of these evaluations, as well as their reproducibility in the case of proprietary models. We provide JUDGE-BENCH, an extensible collection of 20 NLP datasets with human annotations covering a broad range of evaluated properties and types of data, and comprehensively evaluate 11 current LLMs, covering both open-weight and proprietary models, for their ability to replicate the annotations. Our evaluations show substantial variance across models and datasets. Models are reliable evaluators on some tasks, but overall display substantial variability depending on the property being evaluated, the expertise level of the human judges, and whether the language is human or model-generated. We conclude that LLMs should be carefully validated against human judgments before being used as evaluators.

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

Repairs in a Block World: A New Benchmark for Handling User Corrections with Multi-Modal Language Models

In dialogue, the addressee may initially misunderstand the speaker and respond erroneously, often prompting the speaker to correct the misunderstanding in the next turn with a Third Position Repair (TPR). The ability to process and respond appropriately to such repair sequences is thus crucial in conversational AI systems. In this paper, we first collect, analyse, and publicly release BlockWorld-Repairs: a dataset of multi-modal TPR sequences in an instruction-following manipulation task that is, by design, rife with referential ambiguity. We employ this dataset to evaluate several state-of-the-art Vision and Language Models (VLM) across multiple settings, focusing on their capability to process and accurately respond to TPRs and thus recover from miscommunication. We find that, compared to humans, all models significantly underperform in this task. We then show that VLMs can benefit from specialised losses targeting relevant tokens during fine-tuning, achieving better performance and generalising better to new scenarios. Our results suggest that these models are not yet ready to be deployed in multi-modal collaborative settings where repairs are common, and highlight the need to design training regimes and objectives that facilitate learning from interaction. Our code and data are available at www.github.com/JChiyah/blockworld-repairs

cs.CL↗