Andrea Bartezzaghi

Nicholas Thumiger, Andrea Bartezzaghi, Mattia Rigotti et al. · ibm-research

Computational Fluid Dynamics (CFD) is central to race-car aerodynamic development, yet its cost -- tens of thousands of core-hours per high-fidelity evaluation -- severely limits the design space exploration feasible within realistic budgets. AI-based surrogate models promise to alleviate this bottleneck, but progress has been constrained by the limited complexity of public datasets, which are dominated by smoothed passenger-car shapes that fail to exercise surrogates on the thin, complex, highly loaded components governing motorsport performance. This work presents three primary contributions. First, we introduce a high-fidelity RANS dataset built on a parametric LMP2-class CAD model and spanning six operating conditions (map points) covering straight-line and cornering regimes, generated and validated by aerodynamics experts at Dallara to preserve features relevant to industrial motorsport. Second, we present the Gauge-Invariant Spectral Transformer (GIST), a graph-based neural operator whose spectral embeddings encode mesh connectivity to enhance predictions on tightly packed, complex geometries. GIST guarantees discretization invariance and scales linearly with mesh size, achieving state-of-the-art accuracy on both public benchmarks and the proposed race-car dataset. Third, we demonstrate that GIST achieves a level of predictive accuracy suitable for early-stage aerodynamic design, providing a first validation of the concept of interactive design-space exploration -- where engineers query a surrogate in place of the CFD solver -- within industrial motorsport workflows.

Maximilian Kimmich, Andrea Bartezzaghi, Jasmina Bogojeska et al.

Neural approaches have become very popular in Question Answering (QA), however, they require a large amount of annotated data. In this work, we propose a novel approach that combines data augmentation via question-answer generation with Active Learning to improve performance in low-resource settings, where the target domains are diverse in terms of difficulty and similarity to the source domain. We also investigate Active Learning for question answering in different stages, overall reducing the annotation effort of humans. For this purpose, we consider target domains in realistic settings, with an extremely low amount of annotated samples but with many unlabeled documents, which we assume can be obtained with little effort. Additionally, we assume a sufficient amount of labeled data from the source domain being available. We perform extensive experiments to find the best setup for incorporating domain experts. Our findings show that our novel approach, where humans are incorporated in a data generation approach, boosts performance in the low-resource, domain-specific setting, allowing for low-labeling-effort question answering systems in new, specialized domains. They further demonstrate how human annotation affects the performance of QA depending on the stage it is performed.

Brown Ebouky, Andrea Bartezzaghi, Mattia Rigotti · ibm-research

The recent advent of reasoning models like OpenAI's o1 was met with excited speculation by the AI community about the mechanisms underlying these capabilities in closed models, followed by a rush of replication efforts, particularly from the open source community. These speculations were largely settled by the demonstration from DeepSeek-R1 that chains-of-thought and reinforcement learning (RL) can effectively replicate reasoning on top of base LLMs. However, it remains valuable to explore alternative methods for theoretically eliciting reasoning that could help elucidate the underlying mechanisms, as well as providing additional methods that may offer complementary benefits. Here, we build on the long-standing literature in cognitive psychology and cognitive architectures, which postulates that reasoning arises from the orchestrated, sequential execution of a set of modular, predetermined cognitive operations. Crucially, we implement this key idea within a modern agentic tool-calling framework. In particular, we endow an LLM with a small set of "cognitive tools" encapsulating specific reasoning operations, each executed by the LLM itself. Surprisingly, this simple strategy results in considerable gains in performance on standard mathematical reasoning benchmarks compared to base LLMs, for both closed and open-weight models. For instance, providing our "cognitive tools" to GPT-4.1 increases its pass@1 performance on AIME2024 from 32% to 53%, even surpassing the performance of o1-preview. In addition to its practical implications, this demonstration contributes to the debate regarding the role of post-training methods in eliciting reasoning in LLMs versus the role of inherent capabilities acquired during pre-training, and whether post-training merely uncovers these latent abilities.

Brown Ebouky, Gabriele Carrino, Niccolo Avogaro et al.

Human visual reasoning is governed by active vision, a process where metacognitive control drives top-down goal-directed attention, dynamically routing foveal focus toward task-relevant details while maintaining peripheral awareness of the global scene. In contrast, modern Vision-Language Models (VLMs) process visual information passively, relying on the static accumulation of massive token contexts that dilute spatial reasoning and induce linguistic hallucinations. Here we propose the following paradigm shift: GazeVLM, a multimodal architecture that internalizes this metacognitive oversight over its deployment of attention resources directly into the reasoning loop. By empowering the VLM to autonomously generate gaze tokens ($\texttt{<LOOK>}$), GazeVLM establishes a top-down control mechanism over its own causal attention mask. The model dynamically dictates its focal intent, triggering a continuous suppression bias that dampens irrelevant visual features, implementing spatial selective attention and simulating foveal fixation. Once local reasoning concludes, the bias lifts, seamlessly restoring the global view. This architecture enables the model to fluidly transition between global spatial awareness and localized focal reasoning without relying on external agentic contraptions like cropping tools, or inflating the context window with additional visual tokens derived from localized visual patches. Trained with a bespoke Group Relative Policy Optimization (GRPO) procedure that rewards valid grounding, our 4B-parameter GazeVLM delivers strong high-resolution multimodal reasoning performance, surpassing state-of-the-art VLMs in its parameter class by nearly 4% and agentic multimodal pipelines built around thinking with images by more than 5% on HRBench-4k and HRBench-8k.

Maximilian Schmidt, Andrea Bartezzaghi, Ngoc Thang Vu

Although language models (LMs) have boosted the performance of Question Answering, they still need plenty of data. Data annotation, in contrast, is a time-consuming process. This especially applies to Question Answering, where possibly large documents have to be parsed and annotated with questions and their corresponding answers. Furthermore, Question Answering models often only work well for the domain they were trained on. Since annotation is costly, we argue that domain-agnostic knowledge from LMs, such as linguistic understanding, is sufficient to create a well-curated dataset. With this motivation, we show that using large language models can improve Question Answering performance on various datasets in the few-shot setting compared to state-of-the-art approaches. For this, we perform data generation leveraging the Prompting framework, suggesting that language models contain valuable task-agnostic knowledge that can be used beyond the common pre-training/fine-tuning scheme. As a result, we consistently outperform previous approaches on few-shot Question Answering.

Niccolo Avogaro, Thomas Frick, Mattia Rigotti et al. · ibm-research

Large Vision-Language Models (VLMs) are increasingly being regarded as foundation models that can be instructed to solve diverse tasks by prompting, without task-specific training. We examine the seemingly obvious question: how to effectively prompt VLMs for semantic segmentation. To that end, we systematically evaluate the segmentation performance of several recent models guided by either text or visual prompts on the out-of-distribution MESS dataset collection. We introduce a scalable prompting scheme, few-shot prompted semantic segmentation, inspired by open-vocabulary segmentation and few-shot learning. It turns out that VLMs lag far behind specialist models trained for a specific segmentation task, by about 30% on average on the Intersection-over-Union metric. Moreover, we find that text prompts and visual prompts are complementary: each one of the two modes fails on many examples that the other one can solve. Our analysis suggests that being able to anticipate the most effective prompt modality can lead to a 11% improvement in performance. Motivated by our findings, we propose PromptMatcher, a remarkably simple training-free baseline that combines both text and visual prompts, achieving state-of-the-art results outperforming the best text-prompted VLM by 2.5%, and the top visual-prompted VLM by 3.5% on few-shot prompted semantic segmentation.

Brown Ebouky, Ajad Chhatkuli, Cristiano Malossi et al.

Self-supervised learning (SSL) has emerged as a central paradigm for training foundation models by leveraging large-scale unlabeled datasets, often producing representations with strong generalization capabilities. These models are typically pre-trained on general-purpose datasets such as ImageNet and subsequently adapted to various downstream tasks through finetuning. While recent advances have explored parameter-efficient strategies for adapting pre-trained models, extending SSL pre-training itself to new domains - particularly under limited data regimes and for dense prediction tasks - remains underexplored. In this work, we address the problem of adapting vision foundation models to new domains in an unsupervised and data-efficient manner, specifically targeting downstream semantic segmentation. We propose GLARE (Global Local and Regional Enforcement), a novel continual self-supervised pre-training task designed to enhance downstream segmentation performance. GLARE introduces patch-level augmentations to encourage local consistency and incorporates a regional consistency constraint that leverages spatial semantics in the data. For efficient continual pre-training, we initialize Vision Transformers (ViTs) with weights from existing SSL models and update only lightweight adapter modules - specifically UniAdapter - while keeping the rest of the backbone frozen. Experiments across multiple semantic segmentation benchmarks on different domains demonstrate that GLARE consistently improves downstream performance with minimal computational and parameter overhead.

Niccolo Avogaro, Thomas Frick, Yagmur G. Cinar et al. · ibm-research

Large-scale pretrained vision backbones have transformed computer vision by providing powerful feature extractors that enable various downstream tasks, including training-free approaches like visual prompting for semantic segmentation. Despite their success in generic scenarios, these models often fall short when applied to specialized technical domains where the visual features differ significantly from their training distribution. To bridge this gap, we introduce VP Lab, a comprehensive iterative framework that enhances visual prompting for robust segmentation model development. At the core of VP Lab lies E-PEFT, a novel ensemble of parameter-efficient fine-tuning techniques specifically designed to adapt our visual prompting pipeline to specific domains in a manner that is both parameter- and data-efficient. Our approach not only surpasses the state-of-the-art in parameter-efficient fine-tuning for the Segment Anything Model (SAM), but also facilitates an interactive, near-real-time loop, allowing users to observe progressively improving results as they experiment within the framework. By integrating E-PEFT with visual prompting, we demonstrate a remarkable 50\% increase in semantic segmentation mIoU performance across various technical datasets using only 5 validated images, establishing a new paradigm for fast, efficient, and interactive model deployment in new, challenging domains. This work comes in the form of a demonstration.

Jason Tsay, Andrea Bartezzaghi, Aleke Nolte et al.

Artificial Intelligence (AI) development is inherently iterative and experimental. Over the course of normal development, especially with the advent of automated AI, hundreds or thousands of experiments are generated and are often lost or never examined again. There is a lost opportunity to document these experiments and learn from them at scale, but the complexity of tracking and reproducing these experiments is often prohibitive to data scientists. We present the Lifelong Database of Experiments (LDE) that automatically extracts and stores linked metadata from experiment artifacts and provides features to reproduce these artifacts and perform meta-learning across them. We store context from multiple stages of the AI development lifecycle including datasets, pipelines, how each is configured, and training runs with information about their runtime environment. The standardized nature of the stored metadata allows for querying and aggregation, especially in terms of ranking artifacts by performance metrics. We exhibit the capabilities of the LDE by reproducing an existing meta-learning study and storing the reproduced metadata in our system. Then, we perform two experiments on this metadata: 1) examining the reproducibility and variability of the performance metrics and 2) implementing a number of meta-learning algorithms on top of the data and examining how variability in experimental results impacts recommendation performance. The experimental results suggest significant variation in performance, especially depending on dataset configurations; this variation carries over when meta-learning is built on top of the results, with performance improving when using aggregated results. This suggests that a system that automatically collects and aggregates results such as the LDE not only assists in implementing meta-learning but may also improve its performance.