Counting Circuits: Mechanistic Interpretability of Visual Reasoning in Large Vision-Language Models

Counting serves as a simple but powerful test of a Large Vision-Language Model's (LVLM's) reasoning; it forces the model to identify each individual object and then add them all up. In this study, we investigate how LVLMs implement counting using controlled synthetic and real-world benchmarks, combined with mechanistic analyses. Our results show that LVLMs display a human-like counting behavior, with precise performance on small numerosities and noisy estimation for larger quantities. We introduce two novel interpretability methods, Visual Activation Patching and HeadLens, and use them to uncover a structured "counting circuit" that is largely shared across a variety of visual reasoning tasks. Building on these insights, we propose a lightweight intervention strategy that exploits simple and abundantly available synthetic images to fine-tune arbitrary pretrained LVLMs exclusively on counting. Despite the narrow scope of this fine-tuning, the intervention not only enhances counting accuracy on in-distribution synthetic data, but also yields an average improvement of +8.36% on out-of-distribution counting benchmarks and an average gain of +1.54% on complex, general visual reasoning tasks for Qwen2.5-VL. These findings highlight the central, influential role of counting in visual reasoning and suggest a potential pathway for improving overall visual reasoning capabilities through targeted enhancement of counting mechanisms.

T3D: Few-Step Diffusion Language Models via Trajectory Self-Distillation with Direct Discriminative Optimization

Diffusion large language models (DLLMs) have the potential to enable fast text generation by decoding multiple tokens in parallel. However, in practice, their inference efficiency is constrained by the need for many refinement steps, while aggressively reducing the number of steps leads to a substantial degradation in generation quality. To alleviate this, we propose a trajectory self-distillation framework that improves few-step decoding by distilling the model's own generative trajectories. We incorporate Direct Discriminative Optimization (DDO), a reverse-KL objective that promotes mode-seeking distillation and encourages the student to concentrate on high-probability teacher modes. Across benchmarks, our approach consistently outperforms strong few-step baselines and standard training under tight step budgets. Although full-step decoding remains superior, we substantially narrow the gap, establishing a strong foundation towards practical few-step DLLMs. The source code is available at https://github.com/Tyrion58/T3D.

GenVP: Generating Visual Puzzles with Contrastive Hierarchical VAEs

Raven's Progressive Matrices (RPMs) is an established benchmark to examine the ability to perform high-level abstract visual reasoning (AVR). Despite the current success of algorithms that solve this task, humans can generalize beyond a given puzzle and create new puzzles given a set of rules, whereas machines remain locked in solving a fixed puzzle from a curated choice list. We propose Generative Visual Puzzles (GenVP), a framework to model the entire RPM generation process, a substantially more challenging task. Our model's capability spans from generating multiple solutions for one specific problem prompt to creating complete new puzzles out of the desired set of rules. Experiments on five different datasets indicate that GenVP achieves state-of-the-art (SOTA) performance both in puzzle-solving accuracy and out-of-distribution (OOD) generalization in 22 OOD scenarios. Compared to SOTA generative approaches, which struggle to solve RPMs when the feasible solution space increases, GenVP efficiently generalizes to these challenging setups. Moreover, our model demonstrates the ability to produce a wide range of complete RPMs given a set of abstract rules by effectively capturing the relationships between abstract rules and visual object properties.

EAZY: Eliminating Hallucinations in LVLMs by Zeroing out Hallucinatory Image Tokens

Despite their remarkable potential, Large Vision-Language Models (LVLMs) still face challenges with object hallucination, a problem where their generated outputs mistakenly incorporate objects that do not actually exist. Although most works focus on addressing this issue within the language-model backbone, our work shifts the focus to the image input source, investigating how specific image tokens contribute to hallucinations. Our analysis reveals a striking finding: a small subset of image tokens with high attention scores are the primary drivers of object hallucination. By removing these hallucinatory image tokens (only 1.5% of all image tokens), the issue can be effectively mitigated. This finding holds consistently across different models and datasets. Building on this insight, we introduce EAZY, a novel, training-free method that automatically identifies and Eliminates hAllucinations by Zeroing out hallucinatorY image tokens. We utilize EAZY for unsupervised object hallucination detection, achieving 15% improvement compared to previous methods. Additionally, EAZY demonstrates remarkable effectiveness in mitigating hallucinations while preserving model utility and seamlessly adapting to various LVLM architectures.

CASIM: Composite Aware Semantic Injection for Text to Motion Generation

Recent advances in generative modeling and tokenization have driven significant progress in text-to-motion generation, leading to enhanced quality and realism in generated motions. However, effectively leveraging textual information for conditional motion generation remains an open challenge. We observe that current approaches, primarily relying on fixed-length text embeddings (e.g., CLIP) for global semantic injection, struggle to capture the composite nature of human motion, resulting in suboptimal motion quality and controllability. To address this limitation, we propose the Composite Aware Semantic Injection Mechanism (CASIM), comprising a composite-aware semantic encoder and a text-motion aligner that learns the dynamic correspondence between text and motion tokens. Notably, CASIM is model and representation-agnostic, readily integrating with both autoregressive and diffusion-based methods. Experiments on HumanML3D and KIT benchmarks demonstrate that CASIM consistently improves motion quality, text-motion alignment, and retrieval scores across state-of-the-art methods. Qualitative analyses further highlight the superiority of our composite-aware approach over fixed-length semantic injection, enabling precise motion control from text prompts and stronger generalization to unseen text inputs.

Physics-Based Dynamic Models Hybridisation Using Physics-Informed Neural Networks

Physics-based dynamic models (PBDMs) are simplified representations of complex dynamical systems. PBDMs take specific processes within a complex system and assign a fragment of variables and an accompanying set of parameters to depict the processes. As this often leads to suboptimal parameterisation of the system, a key challenge requires refining the empirical parameters and variables to reduce uncertainties while maintaining the model s explainability and enhancing its predictive accuracy. We demonstrate that a hybrid mosquito population dynamics model, which integrates a PBDM with Physics-Informed Neural Networks (PINN), retains the explainability of the PBDM by incorporating the PINN-learned model parameters in place of its empirical counterparts. Specifically, we address the limitations of traditional PBDMs by modelling the parameters of larva and pupa development rates using a PINN that encodes complex, learned interactions of air temperature, precipitation and humidity. Our results demonstrate improved mosquito population simulations including the difficult-to-predict mosquito population peaks. This opens the possibility of hybridisation concept application on other complex systems based on PBDMs such as cancer growth to address the challenges posed by scarce and noisy data, and to numerical weather prediction and climate modelling to overcome the gap between physics-based and data-driven weather prediction models.

TrajDiffuse: A Conditional Diffusion Model for Environment-Aware Trajectory Prediction

Accurate prediction of human or vehicle trajectories with good diversity that captures their stochastic nature is an essential task for many applications. However, many trajectory prediction models produce unreasonable trajectory samples that focus on improving diversity or accuracy while neglecting other key requirements, such as collision avoidance with the surrounding environment. In this work, we propose TrajDiffuse, a planning-based trajectory prediction method using a novel guided conditional diffusion model. We form the trajectory prediction problem as a denoising impaint task and design a map-based guidance term for the diffusion process. TrajDiffuse is able to generate trajectory predictions that match or exceed the accuracy and diversity of the SOTA, while adhering almost perfectly to environmental constraints. We demonstrate the utility of our model through experiments on the nuScenes and PFSD datasets and provide an extensive benchmark analysis against the SOTA methods.

Box2Flow: Instance-based Action Flow Graphs from Videos

A large amount of procedural videos on the web show how to complete various tasks. These tasks can often be accomplished in different ways and step orderings, with some steps able to be performed simultaneously, while others are constrained to be completed in a specific order. Flow graphs can be used to illustrate the step relationships of a task. Current task-based methods try to learn a single flow graph for all available videos of a specific task. The extracted flow graphs tend to be too abstract, failing to capture detailed step descriptions. In this work, our aim is to learn accurate and rich flow graphs by extracting them from a single video. We propose Box2Flow, an instance-based method to predict a step flow graph from a given procedural video. In detail, we extract bounding boxes from videos, predict pairwise edge probabilities between step pairs, and build the flow graph with a spanning tree algorithm. Experiments on MM-ReS and YouCookII show our method can extract flow graphs effectively.

CIC-BART-SSA: Controllable Image Captioning with Structured Semantic Augmentation

Controllable Image Captioning (CIC) aims at generating natural language descriptions for an image, conditioned on information provided by end users, e.g., regions, entities or events of interest. However, available image--language datasets mainly contain captions that describe the entirety of an image, making them ineffective for training CIC models that can potentially attend to any subset of regions or relationships. To tackle this challenge, we propose a novel, fully automatic method to sample additional focused and visually grounded captions using a unified structured semantic representation built on top of the existing set of captions associated with an image. We leverage Abstract Meaning Representation (AMR), a cross-lingual graph-based semantic formalism, to encode all possible spatio-semantic relations between entities, beyond the typical spatial-relations-only focus of current methods. We use this Structured Semantic Augmentation (SSA) framework to augment existing image--caption datasets with the grounded controlled captions, increasing their spatial and semantic diversity and focal coverage. We then develop a new model, CIC-BART-SSA, specifically tailored for the CIC task, that sources its control signals from SSA-diversified datasets. We empirically show that, compared to SOTA CIC models, CIC-BART-SSA generates captions that are superior in diversity and text quality, are competitive in controllability, and, importantly, minimize the gap between broad and highly focused controlled captioning performance by efficiently generalizing to the challenging highly focused scenarios. Code is available at https://github.com/SamsungLabs/CIC-BART-SSA.

ALWOD: Active Learning for Weakly-Supervised Object Detection

Object detection (OD), a crucial vision task, remains challenged by the lack of large training datasets with precise object localization labels. In this work, we propose ALWOD, a new framework that addresses this problem by fusing active learning (AL) with weakly and semi-supervised object detection paradigms. Because the performance of AL critically depends on the model initialization, we propose a new auxiliary image generator strategy that utilizes an extremely small labeled set, coupled with a large weakly tagged set of images, as a warm-start for AL. We then propose a new AL acquisition function, another critical factor in AL success, that leverages the student-teacher OD pair disagreement and uncertainty to effectively propose the most informative images to annotate. Finally, to complete the AL loop, we introduce a new labeling task delegated to human annotators, based on selection and correction of model-proposed detections, which is both rapid and effective in labeling the informative images. We demonstrate, across several challenging benchmarks, that ALWOD significantly narrows the gap between the ODs trained on few partially labeled but strategically selected image instances and those that rely on the fully-labeled data. Our code is publicly available on https://github.com/seqam-lab/ALWOD.