M. Jehanzeb Mirza

Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski et al. · ibm-research, mit

Although action recognition systems can achieve top performance when evaluated on in-distribution test points, they are vulnerable to unanticipated distribution shifts in test data. However, test-time adaptation of video action recognition models against common distribution shifts has so far not been demonstrated. We propose to address this problem with an approach tailored to spatio-temporal models that is capable of adaptation on a single video sample at a step. It consists in a feature distribution alignment technique that aligns online estimates of test set statistics towards the training statistics. We further enforce prediction consistency over temporally augmented views of the same test video sample. Evaluations on three benchmark action recognition datasets show that our proposed technique is architecture-agnostic and able to significantly boost the performance on both, the state of the art convolutional architecture TANet and the Video Swin Transformer. Our proposed method demonstrates a substantial performance gain over existing test-time adaptation approaches in both evaluations of a single distribution shift and the challenging case of random distribution shifts. Code will be available at \url{https://github.com/wlin-at/ViTTA}.

M. Jehanzeb Mirza, Marc Masana, Horst Possegger et al.

Although deep neural networks enable impressive visual perception performance for autonomous driving, their robustness to varying weather conditions still requires attention. When adapting these models for changed environments, such as different weather conditions, they are prone to forgetting previously learned information. This catastrophic forgetting is typically addressed via incremental learning approaches which usually re-train the model by either keeping a memory bank of training samples or keeping a copy of the entire model or model parameters for each scenario. While these approaches show impressive results, they can be prone to scalability issues and their applicability for autonomous driving in all weather conditions has not been shown. In this paper we propose DISC -- Domain Incremental through Statistical Correction -- a simple online zero-forgetting approach which can incrementally learn new tasks (i.e weather conditions) without requiring re-training or expensive memory banks. The only information we store for each task are the statistical parameters as we categorize each domain by the change in first and second order statistics. Thus, as each task arrives, we simply 'plug and play' the statistical vectors for the corresponding task into the model and it immediately starts to perform well on that task. We show the efficacy of our approach by testing it for object detection in a challenging domain-incremental autonomous driving scenario where we encounter different adverse weather conditions, such as heavy rain, fog, and snow.

Muhammad Jehanzeb Mirza, Pol Jané Soneira, Wei Lin et al.

Test-Time-Training (TTT) is an approach to cope with out-of-distribution (OOD) data by adapting a trained model to distribution shifts occurring at test-time. We propose to perform this adaptation via Activation Matching (ActMAD): We analyze activations of the model and align activation statistics of the OOD test data to those of the training data. In contrast to existing methods, which model the distribution of entire channels in the ultimate layer of the feature extractor, we model the distribution of each feature in multiple layers across the network. This results in a more fine-grained supervision and makes ActMAD attain state of the art performance on CIFAR-100C and Imagenet-C. ActMAD is also architecture- and task-agnostic, which lets us go beyond image classification, and score 15.4% improvement over previous approaches when evaluating a KITTI-trained object detector on KITTI-Fog. Our experiments highlight that ActMAD can be applied to online adaptation in realistic scenarios, requiring little data to attain its full performance.

M. Jehanzeb Mirza, Inkyu Shin, Wei Lin et al.

Our MATE is the first Test-Time-Training (TTT) method designed for 3D data, which makes deep networks trained for point cloud classification robust to distribution shifts occurring in test data. Like existing TTT methods from the 2D image domain, MATE also leverages test data for adaptation. Its test-time objective is that of a Masked Autoencoder: a large portion of each test point cloud is removed before it is fed to the network, tasked with reconstructing the full point cloud. Once the network is updated, it is used to classify the point cloud. We test MATE on several 3D object classification datasets and show that it significantly improves robustness of deep networks to several types of corruptions commonly occurring in 3D point clouds. We show that MATE is very efficient in terms of the fraction of points it needs for the adaptation. It can effectively adapt given as few as 5% of tokens of each test sample, making it extremely lightweight. Our experiments show that MATE also achieves competitive performance by adapting sparsely on the test data, which further reduces its computational overhead, making it ideal for real-time applications.

Abhiram Kolli, Muhammad Jehanzeb Mirza, Horst Possegger et al.

Keyless entry systems in cars are adopting neural networks for localizing its operators. Using test-time adversarial defences equip such systems with the ability to defend against adversarial attacks without prior training on adversarial samples. We propose a test-time adversarial example detector which detects the input adversarial example through quantifying the localized intermediate responses of a pre-trained neural network and confidence scores of an auxiliary softmax layer. Furthermore, in order to make the network robust, we extenuate the non-relevant features by non-iterative input sample clipping. Using our approach, mean performance over 15 levels of adversarial perturbations is increased by 55.33% for the fast gradient sign method (FGSM) and 6.3% for both the basic iterative method (BIM) and the projected gradient method (PGD).

M. Jehanzeb Mirza, Leonid Karlinsky, Wei Lin et al.

Vision and Language Models (VLMs), such as CLIP, have enabled visual recognition of a potentially unlimited set of categories described by text prompts. However, for the best visual recognition performance, these models still require tuning to better fit the data distributions of the downstream tasks, in order to overcome the domain shift from the web-based pre-training data. Recently, it has been shown that it is possible to effectively tune VLMs without any paired data, and in particular to effectively improve VLMs visual recognition performance using text-only training data generated by Large Language Models (LLMs). In this paper, we dive deeper into this exciting text-only VLM training approach and explore ways it can be significantly further improved taking the specifics of the downstream task into account when sampling text data from LLMs. In particular, compared to the SOTA text-only VLM training approach, we demonstrate up to 8.4% performance improvement in (cross) domain-specific adaptation, up to 8.7% improvement in fine-grained recognition, and 3.1% overall average improvement in zero-shot classification compared to strong baselines.

Nimrod Shabtay, Felipe Maia Polo, Sivan Doveh et al.

The large-scale training of multi-modal models on data scraped from the web has shown outstanding utility in infusing these models with the required world knowledge to perform effectively on multiple downstream tasks. However, one downside of scraping data from the web can be the potential sacrifice of the benchmarks on which the abilities of these models are often evaluated. To safeguard against test data contamination and to truly test the abilities of these foundation models we propose LiveXiv: A scalable evolving live benchmark based on scientific ArXiv papers. LiveXiv accesses domain-specific manuscripts at any given timestamp and proposes to automatically generate visual question-answer pairs (VQA). This is done without any human-in-the-loop, using the multi-modal content in the manuscripts, like graphs, charts, and tables. Moreover, we introduce an efficient evaluation approach that estimates the performance of all models on the evolving benchmark using evaluations of only a subset of models. This significantly reduces the overall evaluation cost. We benchmark multiple open and proprietary Large Multi-modal Models (LMMs) on the first version of our benchmark, showing its challenging nature and exposing the models true abilities, avoiding contamination. Lastly, in our commitment to high quality, we have collected and evaluated a manually verified subset. By comparing its overall results to our automatic annotations, we have found that the performance variance is indeed minimal (<2.5%). Our dataset is available online on HuggingFace, and our code will be available here.

Saurav Jha, M. Jehanzeb Mirza, Wei Lin et al.

Vision-Language Models (VLMs) remain limited in spatial reasoning tasks that require multi-view understanding and embodied perspective shifts. Recent approaches such as MindJourney attempt to mitigate this gap through test-time scaling where a world model imagines action-conditioned trajectories and a heuristic verifier selects helpful views from such trajectories. In this work, we systematically examine how such test-time verifiers behave across benchmarks, uncovering both their promise and their pitfalls. Our uncertainty-based analyses show that MindJourney's verifier provides little meaningful calibration, and that random scoring often reduces answer entropy equally well, thus exposing systematic action biases and unreliable reward signals. To mitigate these, we introduce a Verification through Spatial Assertions (ViSA) framework that grounds the test-time reward in verifiable, frame-anchored micro-claims. This principled verifier consistently improves spatial reasoning on the SAT-Real benchmark and corrects trajectory-selection biases through more balanced exploratory behavior. However, on the challenging MMSI-Bench, none of the verifiers, including ours, achieve consistent scaling, suggesting that the current world models form an information bottleneck where imagined views fail to enrich fine-grained reasoning. Together, these findings chart the bad, good, and ugly aspects of test-time verification for world-model-based reasoning. Our code is available at https://github.com/chandar-lab/visa-for-mindjourney.

Lisa Weijler, Muhammad Jehanzeb Mirza, Leon Sick et al.

Test-Time Training (TTT) proposes to adapt a pre-trained network to changing data distributions on-the-fly. In this work, we propose the first TTT method for 3D semantic segmentation, TTT-KD, which models Knowledge Distillation (KD) from foundation models (e.g. DINOv2) as a self-supervised objective for adaptation to distribution shifts at test-time. Given access to paired image-pointcloud (2D-3D) data, we first optimize a 3D segmentation backbone for the main task of semantic segmentation using the pointclouds and the task of 2D $\to$ 3D KD by using an off-the-shelf 2D pre-trained foundation model. At test-time, our TTT-KD updates the 3D segmentation backbone for each test sample, by using the self-supervised task of knowledge distillation, before performing the final prediction. Extensive evaluations on multiple indoor and outdoor 3D segmentation benchmarks show the utility of TTT-KD, as it improves performance for both in-distribution (ID) and out-of-distribution (ODO) test datasets. We achieve a gain of up to 13% mIoU (7% on average) when the train and test distributions are similar and up to 45% (20% on average) when adapting to OOD test samples.

Assaf Ben-Kish, Itamar Zimerman, M. Jehanzeb Mirza et al.

A recent trend in LLMs is developing recurrent sub-quadratic models that improve long-context processing efficiency. We investigate leading large long-context models, focusing on how their fixed-size recurrent memory affects their performance. Our experiments reveal that, even when these models are trained for extended contexts, their use of long contexts remains underutilized. Specifically, we demonstrate that a chunk-based inference procedure, which identifies and processes only the most relevant portion of the input can mitigate recurrent memory failures and be effective for many long-context tasks: On LongBench, our method improves the overall performance of Falcon3-Mamba-Inst-7B by 14%, Falcon-Mamba-Inst-7B by 28%, RecurrentGemma-IT-9B by 50%, and RWKV6-Finch-7B by 51%. Surprisingly, this simple approach also leads to state-of-the-art results in the challenging LongBench v2 benchmark, showing competitive performance with equivalent size Transformers. Furthermore, our findings raise questions about whether recurrent models genuinely exploit long-range dependencies, as our single-chunk strategy delivers stronger performance - even in tasks that presumably require cross-context relations.

Weihan Xu, Kan Jen Cheng, Koichi Saito et al.

Joint editing of audio and visual content is crucial for precise and controllable content creation. This new task poses challenges due to the limitations of paired audio-visual data before and after targeted edits, and the heterogeneity across modalities. To address the data and modeling challenges in joint audio-visual editing, we introduce SAVEBench, a paired audiovisual dataset with text and mask conditions to enable object-grounded source-to-target learning. With SAVEBench, we train the Schrodinger Audio-Visual Editor (SAVE), an end-to-end flow-matching model that edits audio and video in parallel while keeping them aligned throughout processing. SAVE incorporates a Schrodinger Bridge that learns a direct transport from source to target audiovisual mixtures. Our evaluation demonstrates that the proposed SAVE model is able to remove the target objects in audio and visual content while preserving the remaining content, with stronger temporal synchronization and audiovisual semantic correspondence compared with pairwise combinations of an audio editor and a video editor.

Lukas Selch, Yufang Hou, M. Jehanzeb Mirza et al.

Large Multimodal Models (LMMs) are increasingly applied to scientific research, yet it remains unclear whether they can reliably understand and reason over the multimodal complexity of papers. A central challenge lies in detecting and resolving inconsistencies across text, figures, tables, and equations, issues that are often subtle, domain-specific, and ultimately undermine clarity, reproducibility, and trust. Existing benchmarks overlook this issue, either isolating single modalities or relying on synthetic errors that fail to capture real-world complexity. We introduce PRISMM-Bench (Peer-Review-sourced Inconsistency Set for Multimodal Models), the first benchmark grounded in real reviewer-flagged inconsistencies in scientific papers. Through a multi-stage pipeline of review mining, LLM-assisted filtering and human verification, we curate 262 inconsistencies from 242 papers. Based on this set, we design three tasks, namely inconsistency identification, remedy and pair matching, which assess a model's capacity to detect, correct, and reason over inconsistencies across different modalities. Furthermore, to address the notorious problem of choice-only shortcuts in multiple-choice evaluation, where models exploit answer patterns without truly understanding the question, we further introduce structured JSON-based answer representations that minimize linguistic biases by reducing reliance on superficial stylistic cues. We benchmark 21 leading LMMs, including large open-weight models (GLM-4.5V 106B, InternVL3 78B) and proprietary models (Gemini 2.5 Pro, GPT-5 with high reasoning). Results reveal strikingly low performance (26.1-54.2%), underscoring the challenge of multimodal scientific reasoning and motivating progress towards trustworthy scientific assistants.

Akshit Singh, Shyam Marjit, Wei Lin et al.

Existing methods for extracting reward signals in Reinforcement Learning typically rely on labeled data and dedicated training splits, a setup that contrasts with how humans learn directly from their environment. In this work, we propose TTRV to enhance vision language understanding by adapting the model on the fly at inference time, without the need for any labeled data. Concretely, we enhance the Group Relative Policy Optimization (GRPO) framework by designing rewards based on the frequency of the base model's output, while inferring on each test sample multiple times. Further, we also propose to control the diversity of the model's output by simultaneously rewarding the model for obtaining low entropy of the output empirical distribution. Our approach delivers consistent gains across both object recognition and visual question answering (VQA), with improvements of up to 52.4% and 29.8%, respectively, and average boosts of 24.6% and 10.0% across 16 datasets.Remarkably, on image recognition, TTRV applied to InternVL 8B surpasses GPT-4o by an average of 2.3% over 8 benchmarks, while remaining highly competitive on VQA, demonstrating that test-time reinforcement learning can match or exceed the strongest proprietary models. Finally, we find many interesting properties of test-time RL for VLMs: for example, even in extremely data-constrained scenarios, where adaptation is performed on a single randomly chosen unlabeled test example, TTRV still yields non-trivial improvements of up to 5.5% in recognition tasks.

Paul Gavrikov, Wei Lin, M. Jehanzeb Mirza et al.

Is basic visual understanding really solved in state-of-the-art VLMs? We present VisualOverload, a slightly different visual question answering (VQA) benchmark comprising 2,720 question-answer pairs, with privately held ground-truth responses. Unlike prior VQA datasets that typically focus on near global image understanding, VisualOverload challenges models to perform simple, knowledge-free vision tasks in densely populated (or, overloaded) scenes. Our dataset consists of high-resolution scans of public-domain paintings that are populated with multiple figures, actions, and unfolding subplots set against elaborately detailed backdrops. We manually annotated these images with questions across six task categories to probe for a thorough understanding of the scene. We hypothesize that current benchmarks overestimate the performance of VLMs, and encoding and reasoning over details is still a challenging task for them, especially if they are confronted with densely populated scenes. Indeed, we observe that even the best model (o3) out of 37 tested models only achieves 19.6% accuracy on our hardest test split and overall 69.5% accuracy on all questions. Beyond a thorough evaluation, we complement our benchmark with an error analysis that reveals multiple failure modes, including a lack of counting skills, failure in OCR, and striking logical inconsistencies under complex tasks. Altogether, VisualOverload exposes a critical gap in current vision models and offers a crucial resource for the community to develop better models. Benchmark: http://paulgavrikov.github.io/visualoverload

Wei Lin, Muhammad Jehanzeb Mirza, Sivan Doveh et al.

Comparing two images in terms of Commonalities and Differences (CaD) is a fundamental human capability that forms the basis of advanced visual reasoning and interpretation. It is essential for the generation of detailed and contextually relevant descriptions, performing comparative analysis, novelty detection, and making informed decisions based on visual data. However, surprisingly, little attention has been given to these fundamental concepts in the best current mimic of human visual intelligence - Large Multimodal Models (LMMs). We develop and contribute a new two-phase approach CaD-VI for collecting synthetic visual instructions, together with an instruction-following dataset CaD-Inst containing 349K image pairs with CaD instructions collected using CaD-VI. Our approach significantly improves the CaD spotting capabilities in LMMs, advancing the SOTA on a diverse set of related tasks by up to 17.5%. It is also complementary to existing difference-only instruction datasets, allowing automatic targeted refinement of those resources increasing their effectiveness for CaD tuning by up to 10%. Additionally, we propose an evaluation benchmark with 7.5K open-ended QAs to assess the CaD understanding abilities of LMMs.

Irene Huang, Wei Lin, M. Jehanzeb Mirza et al.

Compositional Reasoning (CR) entails grasping the significance of attributes, relations, and word order. Recent Vision-Language Models (VLMs), comprising a visual encoder and a Large Language Model (LLM) decoder, have demonstrated remarkable proficiency in such reasoning tasks. This prompts a crucial question: have VLMs effectively tackled the CR challenge? We conjecture that existing CR benchmarks may not adequately push the boundaries of modern VLMs due to the reliance on an LLM-only negative text generation pipeline. Consequently, the negatives produced either appear as outliers from the natural language distribution learned by VLMs' LLM decoders or as improbable within the corresponding image context. To address these limitations, we introduce ConMe -- a compositional reasoning benchmark and a novel data generation pipeline leveraging VLMs to produce `hard CR Q&A'. Through a new concept of VLMs conversing with each other to collaboratively expose their weaknesses, our pipeline autonomously generates, evaluates, and selects challenging compositional reasoning questions, establishing a robust CR benchmark, also subsequently validated manually. Our benchmark provokes a noteworthy, up to 33%, decrease in CR performance compared to preceding benchmarks, reinstating the CR challenge even for state-of-the-art VLMs.

Nadezda Kirillova, M. Jehanzeb Mirza, Horst Bischof et al.

State-of-the-art Multiple Object Tracking (MOT) approaches have shown remarkable performance when trained and evaluated on current benchmarks. However, these benchmarks primarily consist of clear weather scenarios, overlooking adverse atmospheric conditions such as fog, haze, smoke and dust. As a result, the robustness of trackers against these challenging conditions remains underexplored. To address this gap, we introduce physics-based volumetric fog simulation method for arbitrary MOT datasets, utilizing frame-by-frame monocular depth estimation and a fog formation optical model. We enhance our simulation by rendering both homogeneous and heterogeneous fog and propose to use the dark channel prior method to estimate atmospheric light, showing promising results even in night and indoor scenes. We present the leading benchmark MOTChallenge (third release) augmented with fog (smoke for indoor scenes) of various intensities and conduct a comprehensive evaluation of MOT methods, revealing their limitations under fog and fog-like challenges.

Sivan Doveh, Shaked Perek, M. Jehanzeb Mirza et al.

State-of-the-art Vision-Language Models (VLMs) ground the vision and the language modality primarily via projecting the vision tokens from the encoder to language-like tokens, which are directly fed to the Large Language Model (LLM) decoder. While these models have shown unprecedented performance in many downstream zero-shot tasks (eg image captioning, question answers, etc), still little emphasis has been put on transferring one of the core LLM capability of In-Context Learning (ICL). ICL is the ability of a model to reason about a downstream task with a few examples demonstrations embedded in the prompt. In this work, through extensive evaluations, we find that the state-of-the-art VLMs somewhat lack the ability to follow ICL instructions. In particular, we discover that even models that underwent large-scale mixed modality pre-training and were implicitly guided to make use of interleaved image and text information (intended to consume helpful context from multiple images) under-perform when prompted with few-shot demonstrations (in an ICL way), likely due to their lack of direct ICL instruction tuning. To enhance the ICL abilities of the present VLM, we propose a simple yet surprisingly effective multi-turn curriculum-based learning methodology with effective data mixes, leading up to a significant 21.03% (and 11.3% on average) ICL performance boost over the strongest VLM baselines and a variety of ICL benchmarks. Furthermore, we also contribute new benchmarks for ICL evaluation in VLMs and discuss their advantages over the prior art.

Stefan Leitner, M. Jehanzeb Mirza, Wei Lin et al.

In autonomous driving scenarios, current object detection models show strong performance when tested in clear weather. However, their performance deteriorates significantly when tested in degrading weather conditions. In addition, even when adapted to perform robustly in a sequence of different weather conditions, they are often unable to perform well in all of them and suffer from catastrophic forgetting. To efficiently mitigate forgetting, we propose Domain-Incremental Learning through Activation Matching (DILAM), which employs unsupervised feature alignment to adapt only the affine parameters of a clear weather pre-trained network to different weather conditions. We propose to store these affine parameters as a memory bank for each weather condition and plug-in their weather-specific parameters during driving (i.e. test time) when the respective weather conditions are encountered. Our memory bank is extremely lightweight, since affine parameters account for less than 2% of a typical object detector. Furthermore, contrary to previous domain-incremental learning approaches, we do not require the weather label when testing and propose to automatically infer the weather condition by a majority voting linear classifier.

M. Jehanzeb Mirza, Leonid Karlinsky, Wei Lin et al.

Recently, large-scale pre-trained Vision and Language (VL) models have set a new state-of-the-art (SOTA) in zero-shot visual classification enabling open-vocabulary recognition of potentially unlimited set of categories defined as simple language prompts. However, despite these great advances, the performance of these zeroshot classifiers still falls short of the results of dedicated (closed category set) classifiers trained with supervised fine tuning. In this paper we show, for the first time, how to reduce this gap without any labels and without any paired VL data, using an unlabeled image collection and a set of texts auto-generated using a Large Language Model (LLM) describing the categories of interest and effectively substituting labeled visual instances of those categories. Using our label-free approach, we are able to attain significant performance improvements over the zero-shot performance of the base VL model and other contemporary methods and baselines on a wide variety of datasets, demonstrating absolute improvement of up to 11.7% (3.8% on average) in the label-free setting. Moreover, despite our approach being label-free, we observe 1.3% average gains over leading few-shot prompting baselines that do use 5-shot supervision.

M. Jehanzeb Mirza, Jakub Micorek, Horst Possegger et al.

Domain adaptation is crucial to adapt a learned model to new scenarios, such as domain shifts or changing data distributions. Current approaches usually require a large amount of labeled or unlabeled data from the shifted domain. This can be a hurdle in fields which require continuous dynamic adaptation or suffer from scarcity of data, e.g. autonomous driving in challenging weather conditions. To address this problem of continuous adaptation to distribution shifts, we propose Dynamic Unsupervised Adaptation (DUA). By continuously adapting the statistics of the batch normalization layers we modify the feature representations of the model. We show that by sequentially adapting a model with only a fraction of unlabeled data, a strong performance gain can be achieved. With even less than 1% of unlabeled data from the target domain, DUA already achieves competitive results to strong baselines. In addition, the computational overhead is minimal in contrast to previous approaches. Our approach is simple, yet effective and can be applied to any architecture which uses batch normalization as one of its components. We show the utility of DUA by evaluating it on a variety of domain adaptation datasets and tasks including object recognition, digit recognition and object detection.

Muhammad Jehanzeb Mirza, Cornelius Buerkle, Julio Jarquin et al.

State-of-the-art object detection systems for autonomous driving achieve promising results in clear weather conditions. However, such autonomous safety critical systems also need to work in degrading weather conditions, such as rain, fog and snow. Unfortunately, most approaches evaluate only on the KITTI dataset, which consists only of clear weather scenes. In this paper we address this issue and perform one of the most detailed evaluation on single and dual modality architectures on data captured in real weather conditions. We analyse the performance degradation of these architectures in degrading weather conditions. We demonstrate that an object detection architecture performing good in clear weather might not be able to handle degrading weather conditions. We also perform ablation studies on the dual modality architectures and show their limitations.