A Framework for a Capability-driven Evaluation of Scenario Understanding for Multimodal Large Language Models in Autonomous Driving

Multimodal large language models (MLLMs) hold the potential to enhance autonomous driving by combining domain-independent world knowledge with context-specific language guidance. Their integration into autonomous driving systems shows promising results in isolated proof-of-concept applications, while their performance is evaluated on selective singular aspects of perception, reasoning, or planning. To leverage their full potential a systematic framework for evaluating MLLMs in the context of autonomous driving is required. This paper proposes a holistic framework for a capability-driven evaluation of MLLMs in autonomous driving. The framework structures scenario understanding along the four core capability dimensions semantic, spatial, temporal, and physical. They are derived from the general requirements of autonomous driving systems, human driver cognition, and language-based reasoning. It further organises the domain into context layers, processing modalities, and downstream tasks such as language-based interaction and decision-making. To illustrate the framework's applicability, two exemplary traffic scenarios are analysed, grounding the proposed dimensions in realistic driving situations. The framework provides a foundation for the structured evaluation of MLLMs' potential for scenario understanding in autonomous driving.

Measuring Physical Plausibility of 3D Human Poses Using Physics Simulation

Modeling humans in physical scenes is vital for understanding human-environment interactions for applications involving augmented reality or assessment of human actions from video (e.g. sports or physical rehabilitation). State-of-the-art literature begins with a 3D human pose, from monocular or multiple views, and uses this representation to ground the person within a 3D world space. While standard metrics for accuracy capture joint position errors, they do not consider physical plausibility of the 3D pose. This limitation has motivated researchers to propose other metrics evaluating jitter, floor penetration, and unbalanced postures. Yet, these approaches measure independent instances of errors and are not representative of balance or stability during motion. In this work, we propose measuring physical plausibility from within physics simulation. We introduce two metrics to capture the physical plausibility and stability of predicted 3D poses from any 3D Human Pose Estimation model. Using physics simulation, we discover correlations with existing plausibility metrics and measuring stability during motion. We evaluate and compare the performances of two state-of-the-art methods, a multi-view triangulated baseline, and ground truth 3D markers from the Human3.6m dataset.

VITRO: Vocabulary Inversion for Time-series Representation Optimization

Although LLMs have demonstrated remarkable capabilities in processing and generating textual data, their pre-trained vocabularies are ill-suited for capturing the nuanced temporal dynamics and patterns inherent in time series. The discrete, symbolic nature of natural language tokens, which these vocabularies are designed to represent, does not align well with the continuous, numerical nature of time series data. To address this fundamental limitation, we propose VITRO. Our method adapts textual inversion optimization from the vision-language domain in order to learn a new time series per-dataset vocabulary that bridges the gap between the discrete, semantic nature of natural language and the continuous, numerical nature of time series data. We show that learnable time series-specific pseudo-word embeddings represent time series data better than existing general language model vocabularies, with VITRO-enhanced methods achieving state-of-the-art performance in long-term forecasting across most datasets.

Explainable Procedural Mistake Detection

Automated task guidance has recently attracted attention from the AI research community. Procedural mistake detection (PMD) is a challenging sub-problem of classifying whether a human user (observed through egocentric video) has successfully executed the task at hand (specified by a procedural text). Despite significant efforts in building resources and models for PMD, machine performance remains nonviable, and the reasoning processes underlying this performance are opaque. As such, we recast PMD to an explanatory self-dialog of questions and answers, which serve as evidence for a decision. As this reformulation enables an unprecedented transparency, we leverage a fine-tuned natural language inference (NLI) model to formulate two automated coherence metrics for generated explanations. Our results show that while open-source VLMs struggle with this task off-the-shelf, their accuracy, coherence, and dialog efficiency can be vastly improved by incorporating these coherence metrics into common inference and fine-tuning methods. Furthermore, our multi-faceted metrics can visualize common outcomes at a glance, highlighting areas for improvement.

Instructional Video Generation

Despite the recent strides in video generation, state-of-the-art methods still struggle with elements of visual detail. One particularly challenging case is the class of egocentric instructional videos in which the intricate motion of the hand coupled with a mostly stable and non-distracting environment is necessary to convey the appropriate visual action instruction. To address these challenges, we introduce a new method for instructional video generation. Our diffusion-based method incorporates two distinct innovations. First, we propose an automatic method to generate the expected region of motion, guided by both the visual context and the action text. Second, we introduce a critical hand structure loss to guide the diffusion model to focus on smooth and consistent hand poses. We evaluate our method on augmented instructional datasets based on EpicKitchens and Ego4D, demonstrating significant improvements over state-of-the-art methods in terms of instructional clarity, especially of the hand motion in the target region, across diverse environments and actions.Video results can be found on the project webpage: https://excitedbutter.github.io/Instructional-Video-Generation/

Class-wise Autoencoders Measure Classification Difficulty And Detect Label Mistakes

We introduce a new framework for analyzing classification datasets based on the ratios of reconstruction errors between autoencoders trained on individual classes. This analysis framework enables efficient characterization of datasets on the sample, class, and entire dataset levels. We define reconstruction error ratios (RERs) that probe classification difficulty and allow its decomposition into (1) finite sample size and (2) Bayes error and decision-boundary complexity. Through systematic study across 19 popular visual datasets, we find that our RER-based dataset difficulty probe strongly correlates with error rate for state-of-the-art (SOTA) classification models. By interpreting sample-level classification difficulty as a label mistakenness score, we further find that RERs achieve SOTA performance on mislabel detection tasks on hard datasets under symmetric and asymmetric label noise. Our code is publicly available at https://github.com/voxel51/reconstruction-error-ratios.

Zero-Shot Coreset Selection: Efficient Pruning for Unlabeled Data

Deep learning increasingly relies on massive data with substantial costs for storage, annotation, and model training. To reduce these costs, coreset selection aims to find a representative subset of data to train models while ideally performing on par with the full data training. State-of-the-art coreset methods use carefully-designed criteria to quantify the importance of each data example via ground truth labels and dataset-specific training, then select examples whose scores lie in a certain range to construct a coreset. These methods work well in their respective settings, however, they cannot select data that are unlabeled, which is the majority of real-world data. To that end, this paper motivates and formalizes the problem of unlabeled coreset selection to enable greater scale and reduce annotation costs for deep learning. As a solution, we develop Zero-Shot Coreset Selection (ZCore), a method that efficiently selects coresets without ground truth labels or training on candidate data. Instead, ZCore uses existing foundation models to generate a zero-shot embedding space for unlabeled data, then quantifies the relative importance of each example based on overall coverage and redundancy within the embedding distribution. We evaluate ZCore on four datasets and outperform several state-of-the-art label-based methods, leading to a strong baseline for future research in unlabeled coreset selection. On ImageNet, ZCore selections achieve a downstream model accuracy of 53.99% with only 10% training data, which outperforms label-based methods while removing annotation requirements for 1.15 million images. Our code is publicly available at https://github.com/voxel51/zcore.

Learning to Estimate External Forces of Human Motion in Video

Analyzing sports performance or preventing injuries requires capturing ground reaction forces (GRFs) exerted by the human body during certain movements. Standard practice uses physical markers paired with force plates in a controlled environment, but this is marred by high costs, lengthy implementation time, and variance in repeat experiments; hence, we propose GRF inference from video. While recent work has used LSTMs to estimate GRFs from 2D viewpoints, these can be limited in their modeling and representation capacity. First, we propose using a transformer architecture to tackle the GRF from video task, being the first to do so. Then we introduce a new loss to minimize high impact peaks in regressed curves. We also show that pre-training and multi-task learning on 2D-to-3D human pose estimation improves generalization to unseen motions. And pre-training on this different task provides good initial weights when finetuning on smaller (rarer) GRF datasets. We evaluate on LAAS Parkour and a newly collected ForcePose dataset; we show up to 19% decrease in error compared to prior approaches.

Q-TART: Quickly Training for Adversarial Robustness and in-Transferability

Raw deep neural network (DNN) performance is not enough; in real-world settings, computational load, training efficiency and adversarial security are just as or even more important. We propose to simultaneously tackle Performance, Efficiency, and Robustness, using our proposed algorithm Q-TART, Quickly Train for Adversarial Robustness and in-Transferability. Q-TART follows the intuition that samples highly susceptible to noise strongly affect the decision boundaries learned by DNNs, which in turn degrades their performance and adversarial susceptibility. By identifying and removing such samples, we demonstrate improved performance and adversarial robustness while using only a subset of the training data. Through our experiments we highlight Q-TART's high performance across multiple Dataset-DNN combinations, including ImageNet, and provide insights into the complementary behavior of Q-TART alongside existing adversarial training approaches to increase robustness by over 1.3% while using up to 17.9% less training time.

Come Again? Re-Query in Referring Expression Comprehension

To build a shared perception of the world, humans rely on the ability to resolve misunderstandings by requesting and accepting clarifications. However, when evaluating visiolinguistic models, metrics such as accuracy enforce the assumption that a decision must be made based on a single piece of evidence. In this work, we relax this assumption for the task of referring expression comprehension by allowing the model to request help when its confidence is low. We consider two ways in which this help can be provided: multimodal re-query, where the user is allowed to point or click to provide additional information to the model, and rephrase re-query, where the user is only allowed to provide another referring expression. We demonstrate the importance of re-query by showing that providing the best referring expression for all objects can increase accuracy by up to 21.9% and that this accuracy can be matched by re-querying only 12% of initial referring expressions. We further evaluate re-query functions for both multimodal and rephrase re-query across three modern approaches and demonstrate combined replacement for rephrase re-query, which improves average single-query performance by up to 6.5% and converges to as close as 1.6% of the upper bound of single-query performance.