arXiv subjects

Taras Rumezhak

Publications and source records attributed to Taras Rumezhak.

5 recordsLinked to original sources

arXiv2026

Running the Gauntlet: Challenging Agentic Tasks

As agentic systems continue to evolve and are widely deployed in real-world scenarios, there is a growing demand to faithfully evaluate their capabilities. However, current benchmarks are typically built on popular applications with relatively simple tasks and focus on a narrow set of capabilities while overlooking broader dimensions, resulting in saturated performance on modern agents and failing to probe their limitations. To this end, we introduce GauntletBench, a web-based benchmark for evaluating agent generalisation in challenging scenarios, focusing on three underexplored capabilities (temporal perception, graphical understanding, and 3D reasoning), across five less-covered professional applications (Video Editor, Workflow Builder, 3D Modeller, Flight Analyser, and Circuit Designer), each with 27 vision-intensive tasks (135 in total). Our benchmark provides a modular pipeline that comprises an environment compatible with both open- and closed-source agent frameworks, a controlled web-based application, a well-structured task suite, and an automated evaluation engine with diverse metrics. Contrary to widespread expectations, our empirical results reveal that frontier agentic systems remain far from achieving human-level performance. Even the state-of-the-art agent achieves only a 28.2% success rate on our GauntletBench, highlighting the limitations in these overlooked capabilities and generalisation. By comparison, non-expert human annotators achieve over 80% success on our challenging yet feasible tasks, revealing the substantial gap between current agent capabilities and those required for complex real-world scenarios.

cs.LG↗

arXiv2026

Just on Time: Token-Level Early Stopping for Diffusion Language Models

Diffusion language models generate text through iterative refinement, a process that is often computationally inefficient because many tokens reach stability long before the final denoising step. We introduce a training-free, token-level early stopping approach that identifies convergence independently at each position. Our method leverages lightweight signals derived from the model's predictions and local context to dynamically determine when individual tokens can be finalized. This yields adaptive per-token freezing without task-specific fine-tuning, substantially reducing the total number of diffusion steps required. Across diverse benchmarks, spanning mathematical reasoning, general question answering, and scientific understanding, our approach achieves substantial efficiency gains while preserving generation quality.

cs.LG↗

arXiv2026

Critic-Guided Reinforcement Unlearning in Text-to-Image Diffusion

Machine unlearning in text-to-image diffusion models aims to remove targeted concepts while preserving overall utility. Prior diffusion unlearning methods typically rely on supervised weight edits or global penalties; reinforcement-learning (RL) approaches, while flexible, often optimize sparse end-of-trajectory rewards, yielding high-variance updates and weak credit assignment. We present a general RL framework for diffusion unlearning that treats denoising as a sequential decision process and introduces a timestep-aware critic with noisy-step rewards. Concretely, we train a CLIP-based reward predictor on noisy latents and use its per-step signal to compute advantage estimates for policy-gradient updates of the reverse diffusion kernel. Our algorithm is simple to implement, supports off-policy reuse, and plugs into standard text-to-image backbones. Across multiple concepts, the method achieves better or comparable forgetting to strong baselines while maintaining image quality and benign prompt fidelity; ablations show that (i) per-step critics and (ii) noisy-conditioned rewards are key to stability and effectiveness. We release code and evaluation scripts to facilitate reproducibility and future research on RL-based diffusion unlearning.

cs.LG↗

arXiv2023

Epistemic Parity: Reproducibility as an Evaluation Metric for Differential Privacy

Differential privacy (DP) data synthesizers support public release of sensitive information, offering theoretical guarantees for privacy but limited evidence of utility in practical settings. Utility is typically measured as the error on representative proxy tasks, such as descriptive statistics, accuracy of trained classifiers, or performance over a query workload. The ability for these results to generalize to practitioners' experience has been questioned in a number of settings, including the U.S. Census. In this paper, we propose an evaluation methodology for synthetic data that avoids assumptions about the representativeness of proxy tasks, instead measuring the likelihood that published conclusions would change had the authors used synthetic data, a condition we call epistemic parity. Our methodology consists of reproducing empirical conclusions of peer-reviewed papers on real, publicly available data, then re-running these experiments a second time on DP synthetic data, and comparing the results. We instantiate our methodology over a benchmark of recent peer-reviewed papers that analyze public datasets in the ICPSR repository. We model quantitative claims computationally to automate the experimental workflow, and model qualitative claims by reproducing visualizations and comparing the results manually. We then generate DP synthetic datasets using multiple state-of-the-art mechanisms, and estimate the likelihood that these conclusions will hold. We find that state-of-the-art DP synthesizers are able to achieve high epistemic parity for several papers in our benchmark. However, some papers, and particularly some specific findings, are difficult to reproduce for any of the synthesizers. We advocate for a new class of mechanisms that favor stronger utility guarantees and offer privacy protection with a focus on application-specific threat models and risk-assessment.

cs.CR↗

arXiv2022

Towards realistic symmetry-based completion of previously unseen point clouds

3D scanning is a complex multistage process that generates a point cloud of an object typically containing damaged parts due to occlusions, reflections, shadows, scanner motion, specific properties of the object surface, imperfect reconstruction algorithms, etc. Point cloud completion is specifically designed to fill in the missing parts of the object and obtain its high-quality 3D representation. The existing completion approaches perform well on the academic datasets with a predefined set of object classes and very specific types of defects; however, their performance drops significantly in the real-world settings and degrades even further on previously unseen object classes. We propose a novel framework that performs well on symmetric objects, which are ubiquitous in man-made environments. Unlike learning-based approaches, the proposed framework does not require training data and is capable of completing non-critical damages occurring in customer 3D scanning process using e.g. Kinect, time-of-flight, or structured light scanners. With thorough experiments, we demonstrate that the proposed framework achieves state-of-the-art efficiency in point cloud completion of real-world customer scans. We benchmark the framework performance on two types of datasets: properly augmented existing academic dataset and the actual 3D scans of various objects.

cs.CV↗