Mark James Carman

Loris Giulivi, Mark James Carman, Giacomo Boracchi

Artificial intelligence (AI) systems power the world we live in. Deep neural networks (DNNs) are able to solve tasks in an ever-expanding landscape of scenarios, but our eagerness to apply these powerful models leads us to focus on their performance and deprioritises our ability to understand them. Current research in the field of explainable AI tries to bridge this gap by developing various perturbation or gradient-based explanation techniques. For images, these techniques fail to fully capture and convey the semantic information needed to elucidate why the model makes the predictions it does. In this work, we develop a new form of explanation that is radically different in nature from current explanation methods, such as Grad-CAM. Perception visualization provides a visual representation of what the DNN perceives in the input image by depicting what visual patterns the latent representation corresponds to. Visualizations are obtained through a reconstruction model that inverts the encoded features, such that the parameters and predictions of the original models are not modified. Results of our user study demonstrate that humans can better understand and predict the system's decisions when perception visualizations are available, thus easing the debugging and deployment of deep models as trusted systems.

Andrea Sassella, Andrea Chizzola, Tommaso Bianchi et al.

This report benchmarks the performance of ENGINEERING Ingegneria Informatica S.p.A.'s EngGPT2MoE-16B-A3B LLM, a 16B parameter Mixture of Experts (MoE) model with 3B active parameters. Performance is investigated across a wide variety of representative benchmarks, and is compared against comparably-sized open-source MoE and dense models. In comparison with popular Italian models, namely FastwebMIIA-7B, Minerva-7B, Velvet-14B, and LLaMAntino-3-ANITA-8B, EngGPT2MoE-16B-A3B performs as well or better on international benchmarks: ARC-Challenge, GSM8K, AIME24, AIME25, MMLU, and HumanEval (HE). It achieves the best performance for the longest context setting (32k) of the RULER benchmark. On the Italian benchmark dataset ITALIC, the model performs as well or better than the other models except for Velvet-14B, which outperforms it. Compared with popular MoE models of comparable size, the new model reports higher values than DeepSeek-MoE-16B-Chat on all considered benchmarks. It has higher values than Moonlight-16B-A3B on HE, MMLU, AIME24, AIME25, GSM8K, and the 32k RULER setting, but lower on BFCL and some ARC and ITALIC settings. Finally it has lower values than GPT-OSS-20B on most benchmarks, including HE, MMLU, AIME24, AIME25, GSM8K, ARC, BFCL, and the RULER 32k. When compared with popular dense models, EngGPT2MoE-16B-A3B reports higher values on AIME24 and AIME25 than Llama-3.1-8B-Instruct, Gemma-3-12b-it, and Ministral-3-8BInstruct-2512-BF16, but lower values on ITALIC, BFCL, and RULER with a 32k context. When performance is aggregated across all benchmark metrics, EngGPT2MoE-16B-A3B shows higher performance than the Italian models under evaluation while achieving lower results than some of the most performant international models, in particular GPT-5 nano and Qwen3-8B. Taken together, our findings find the new model to be a step forward for native Italian Large Language Models.

Gabriele Carrino, Andrea Sassella, Nicolo Brunello et al.

Recent advances in large language models (LLMs) highlight the importance of post training techniques for improving reasoning and mathematical ability. Group Relative Policy Optimization (GRPO) has shown promise in this domain by combining group relative advantage estimation, PPO style clipping, and KL regularization. However, its complexity raises the question of whether all components are necessary for fostering reasoning behaviors. We conduct a systematic analysis of GRPO and identify two key findings: (1) incorporating negative feedback is essential training solely on actions above a baseline limits learning; and (2) PPO style constraints, such as policy ratio clipping, are not required to improve mathematical reasoning or performance. Building on these insights, we propose REINFORCE with Group Relative Advantage (RGRA), a simplified variant that retains group relative advantage estimation but removes PPO style clipping and policy ratio terms. Experiments across standard mathematical benchmarks indicate that RGRA has the potential to achieve stronger performance than GRPO. Our results suggest that simpler REINFORCE based approaches can effectively enhance reasoning in LLMs, offering a more transparent and efficient alternative to GRPO.

Federico Toschi, Nicolò Brunello, Andrea Sassella et al.

The recent advancements introduced by Large Language Models (LLMs) have transformed how Artificial Intelligence (AI) can support complex, real world tasks, pushing research outside the text boundaries towards multi modal contexts and leading to Multimodal Large Language Models (MLMs). Given the current adoption of LLM based assistants in solving technical or domain specific problems, the natural continuation of this trend is to extend the input domains of these assistants exploiting MLMs. Ideally, these MLMs should be used as real time assistants in procedural tasks, hopefully integrating a view of the environment where the user being assisted is, or even better sharing the same point of view via Virtual Reality (VR) or Augmented Reality (AR) supports, to reason over the same scenario the user is experiencing. With this work, we aim at evaluating the quality of currently openly available MLMs to provide this kind of assistance on technical tasks. To this end, we annotated a data set of furniture assembly with step by step labels and manual references: the Manual to Action Dataset (M2AD). We used this dataset to assess (1) to which extent the reasoning abilities of MLMs can be used to reduce the need for detailed labelling, allowing for more efficient, cost effective annotation practices, (2) whether MLMs are able to track the progression of assembly steps (3) and whether MLMs can refer correctly to the instruction manual pages. Our results showed that while some models understand procedural sequences, their performance is limited by architectural and hardware constraints, highlighting the need for multi image and interleaved text image reasoning.

Nicolò Brunello, Davide Rigamonti, Andrea Sassella et al.

The reasoning capabilities of Large Language Models (LLMs) have increased greatly over the last few years, as have their size and complexity. Nonetheless, the use of LLMs in production remains challenging due to their unpredictable nature and discrepancies that can exist between their desired behavior and their actual model output. In this paper, we introduce a new tool, InTraVisTo (Inside Transformer Visualisation Tool), designed to enable researchers to investigate and trace the computational process that generates each token in a Transformer-based LLM. InTraVisTo provides a visualization of both the internal state of the Transformer model (by decoding token embeddings at each layer of the model) and the information flow between the various components across the different layers of the model (using a Sankey diagram). With InTraVisTo, we aim to help researchers and practitioners better understand the computations being performed within the Transformer model and thus to shed some light on internal patterns and reasoning processes employed by LLMs.

Raul Singh, Nicolo Brunello, Vincenzo Scotti et al.

The ability of Large Language Models (LLMs) to solve complex tasks has made them crucial in the development of AI-based applications. However, the high computational requirements to fine-tune these LLMs on downstream tasks pose significant challenges, particularly when resources are limited. In response to this challenge, we introduce L1RA, a novel technique aimed at dynamically distributing the rank of low-rank adapters during fine-tuning using LoRA. Given a rank budget (i.e., total sum of adapters rank), L1RA leverages L1 regularisation to prune redundant ranks and redistribute them across adapters, thereby optimising resource utilisation. Through a series of comprehensive experiments, we empirically demonstrate that L1RA maintains comparable or even reduced computational overhead compared to other LoRA variants, including the vanilla approach, while achieving same or better performances. Moreover, the post-training analysis of rank distribution unveiled insights into the specific model components requiring the most adaptation to align with the task objective: the feed-forward layers and the attention output projection. These results highlight the efficacy of L1RA in not only enhancing the efficiency of LLM fine-tuning, but also in providing valuable diagnostic information for model refinement and customisation. In conclusion, L1RA stands as a promising technique for advancing the performance and interpretability of LLM adaptation, particularly in scenarios where computational resources are constrained.

Samuele Pino, Mark James Carman, Paolo Bestagini

Deepfakes are computer manipulated videos where the face of an individual has been replaced with that of another. Software for creating such forgeries is easy to use and ever more popular, causing serious threats to personal reputation and public security. The quality of classifiers for detecting deepfakes has improved with the releasing of ever larger datasets, but the understanding of why a particular video has been labelled as fake has not kept pace. In this work we develop, extend and compare white-box, black-box and model-specific techniques for explaining the labelling of real and fake videos. In particular, we adapt SHAP, GradCAM and self-attention models to the task of explaining the predictions of state-of-the-art detectors based on EfficientNet, trained on the Deepfake Detection Challenge (DFDC) dataset. We compare the obtained explanations, proposing metrics to quantify their visual features and desirable characteristics, and also perform a user survey collecting users' opinions regarding the usefulness of the explainers.

Aditya Joshi, Pushpak Bhattacharyya, Mark James Carman

Automatic sarcasm detection is the task of predicting sarcasm in text. This is a crucial step to sentiment analysis, considering prevalence and challenges of sarcasm in sentiment-bearing text. Beginning with an approach that used speech-based features, sarcasm detection has witnessed great interest from the sentiment analysis community. This paper is the first known compilation of past work in automatic sarcasm detection. We observe three milestones in the research so far: semi-supervised pattern extraction to identify implicit sentiment, use of hashtag-based supervision, and use of context beyond target text. In this paper, we describe datasets, approaches, trends and issues in sarcasm detection. We also discuss representative performance values, shared tasks and pointers to future work, as given in prior works. In terms of resources that could be useful for understanding state-of-the-art, the survey presents several useful illustrations - most prominently, a table that summarizes past papers along different dimensions such as features, annotation techniques, data forms, etc.