Abstract

Contrastive vision-language models (VLMs) such as CLIP achieve strong zero-shot recognition yet remain vulnerable to spurious correlations, particularly background over-reliance. We introduce Cluster-based Concept Importance (CCI), a novel interpretability method that uses CLIP's own patch embeddings to group spatial patches into semantically coherent clusters, mask them, and evaluate relative changes in model predictions. CCI sets a new state of the art on faithfulness benchmarks, surpassing prior methods by large margins; for example, it yields more than a twofold improvement on the deletion-AUC metric for MS COCO retrieval. We further propose that CCI, when combined with GroundedSAM, automatically categorizes predictions as foreground- or background-driven, providing a crucial diagnostic ability. Existing benchmarks such as CounterAnimals, however, rely solely on accuracy and implicitly attribute all performance degradation to background correlations. Our analysis shows this assumption to be incomplete, since many errors arise from viewpoint variation, scale shifts, and fine-grained object confusions. To disentangle these effects, we introduce COVAR, a benchmark that systematically varies object foregrounds and backgrounds. Leveraging CCI with COVAR, we present a comprehensive evaluation of eighteen CLIP variants, offering methodological advances and empirical evidence that chart a path toward more robust VLMs

Abstract

Large-scale vision-language models such as CLIP achieve strong zero-shot recognition but struggle with classes that are rarely seen during pretraining, including newly emerging entities and culturally specific categories. We introduce LiteEmbed, a lightweight framework for few-shot personalization of CLIP that enables new classes to be added without retraining its encoders. LiteEmbed performs subspace-guided optimization of text embeddings within CLIP's vocabulary, leveraging a PCA-based decomposition that disentangles coarse semantic directions from fine-grained variations. Two complementary objectives, coarse alignment and fine separation, jointly preserve global semantic consistency while enhancing discriminability among visually similar classes. Once optimized, the embeddings are plug-and-play, seamlessly substituting CLIP's original text features across classification, retrieval, segmentation, and detection tasks. Extensive experiments demonstrate substantial gains over prior methods, establishing LiteEmbed as an effective approach for adapting CLIP to underrepresented, rare, or unseen classes.

Abstract

We propose an alignment-free, end-to-end Non-Audible Murmur (NAM)-to-Speech conversion model. Existing methods rely on large NAM-text pairs per speaker to generate
high-quality alignments for training non-autoregressive models.
However, alignment quality deteriorates when trained on multispeaker data, limiting their ability to generalize and effectively
utilize the available training data. To address this, we introduce
a streamlined autoregressive approach that eliminates the need
for explicit alignment learning. By leveraging multi-speaker
samples, synthetic training pairs, and multitask character recognition training, our method reduces the word error rate (WER)
by 59.19% compared to the state-of-the-art (SOTA) on two public datasets. We demonstrate the model's zero-shot capability and validate the effectiveness of multitask training through
ablation studies. Speech samples are available at https:
//noalignnam.github.io/autoregressiveNAM/

Abstract

Pretrained models are ubiquitous in the current deep learning landscape, offering strong
results on a broad range of tasks. Recent works have shown that models differing in various design choices exhibit categorically diverse generalization behavior, resulting in one
model grasping distinct data-specific insights unavailable to the other. In this paper, we
propose to leverage large publicly available model repositories as an auxiliary source of
model improvements. We introduce a data partitioning strategy where pretrained models
autonomously adopt either the role of a student, seeking knowledge, or that of a teacher,
imparting knowledge. Experiments across various tasks demonstrate the effectiveness of
our proposed approach. In image classification, we improved the performance of ViT-B
by approximately 1.4% through bidirectional knowledge transfer with ViT-T. For semantic
segmentation, our method boosted all evaluation metrics by enabling knowledge transfer
both within and across backbone architectures. In video saliency prediction, our approach
achieved a new state-of-the-art. We further extend our approach to knowledge transfer
between multiple models, leading to considerable performance improvements for all model
participants.

Abstract

To understand a prompt, Vision-Language models (VLMs) must perceive the image, comprehend the text, and build associations within and across both modalities. For instance, given an `image of a red toy car', the model should associate this image to phrases like `car', `red toy', `red object', etc. Prior work proposed the Binding ID mechanism in LLMs, suggesting that the entity and its corresponding attribute tokens share a Binding ID in the model activations. We investigate this for image-text binding in VLMs using a synthetic dataset and task that requires models to associate 3D objects in an image with their descriptions in the text. Our experiments demonstrate that VLMs assign a distinct Binding ID to an object's image tokens and its textual references, enabling in-context association.

Abstract

We motivate weakly supervised learning as an effective learning paradigm for problems where curating perfectly annotated datasets is expensive and may require domain expertise such as fine-grained classification. We focus on Partial Label Learning (PLL), a weakly-supervised learning paradigm where each training instance is paired with a set of candidate labels (partial label), one of which is the true label. Noisy PLL (NPLL) relaxes this constraint by allowing some partial labels to not contain the true label, enhancing the practicality of the problem. Our work centres on NPLL and presents a framework that initially assigns pseudo-labels to images by exploiting the noisy partial labels through a weighted nearest neighbour algorithm. These pseudo-label and image pairs are then used to train a deep neural network classifier with label smoothing. The classifier's features and predictions are subsequently employed to refine and enhance the accuracy of pseudo-labels. We perform thorough experiments on seven datasets and compare against nine NPLL and PLL methods. We achieve state-of-the-art results in all studied settings from the prior literature, obtaining substantial gains in the simulated fine-grained benchmarks. Further, we show the promising generalisation capability of our framework in realistic, fine-grained, crowd-sourced datasets.

Abstract

We consider the problem of single-source domain generalization. Existing methods typically rely on extensive augmentations to synthetically cover diverse domains during training. However, they struggle with semantic shifts (e.g., background and viewpoint changes), as they often learn global features instead of local concepts that tend to be domain invariant. To address this gap, we propose an approach that compels models to leverage such local concepts
during prediction. Given no suitable dataset with per-class concepts and localization maps exists, we first develop a novel pipeline to generate annotations by exploiting the rich features of diffusion and large-language models. Our next innovation is TIDE, a novel training scheme with a concept saliency alignment loss that ensures model focus on the right per-concept regions and a local concept contrastive loss that promotes learning domain-invariant concept representations. This not only gives a robust model but also can be visually interpreted using the predicted concept saliency maps. Given these maps at test time, our final contribution is a new correction algorithm that uses the corresponding local concept representations to iteratively refine the prediction until it aligns with prototypical concept representations that we store at the end of model training. We evaluate our approach extensively on four standard DG benchmark datasets and substantially outperform the current state-of-the-art (12% improvement on average) while also demonstrating that our predictions can be visually interpreted.

Abstract

Recent evaluations of LLMs on coreference resolution have revealed that traditional output formats and evaluation metrics do not fully capture the models' referential understanding. To address this, we introduce IdentifyMe, a new benchmark for mention resolution presented in a multiple-choice question (MCQ) format, commonly used for evaluating LLMs. IdentifyMe features long narratives and employs heuristics to exclude easily identifiable mentions, creating a more challenging task. The benchmark also consists of a curated mixture of different mention types and corresponding entities, allowing for a fine-grained analysis of model performance. We evaluate both closed- and open source LLMs on IdentifyMe and observe a significant performance gap (20-30%) between the state-of-the-art sub-10B open models vs. closed ones. We observe that pronominal mentions, which have limited surface information, are typically much harder for models to resolve than nominal mentions. Additionally, we find that LLMs often confuse entities when their mentions overlap in nested structures. The highest-scoring model, GPT-4o, achieves 81.9% accuracy, highlighting the strong referential capabilities of state-of-the-art LLMs while also indicating room for further improvement

Abstract

A fundamental aspect of compositional reasoning in a video is associating people and their actions across time. Recent years have seen great progress in general-purpose vision/video models and a move towards long-video understanding. While exciting, we take a step back and ask: are today's models good at compositional reasoning on short videos? To this end, we introduce VELOCITI, a benchmark to study Video-LLMs by disentangling and assessing the comprehension of agents, actions, and their associations across multiple events. We adopt the Video-Language Entailment setup and propose StrictVLE that requires correct classification (rather than ranking) of the positive and negative caption. We evaluate several models and observe that even the best, LLaVA-OneVision (42.5%) and GPT-4o (44.3%), are far from human accuracy at 89.6%. Results show that action understanding lags behind agents, and negative captions created using entities appearing in the video perform worse than those obtained from pure text manipulation. We also present challenges with ClassicVLE and multiple-choice (MC) evaluation, strengthening our preference for StrictVLE. Finally, we validate that our benchmark requires visual inputs of multiple frames making it ideal to study video-language compositional reasoning.

Abstract

Pronunciation correction is crucial for Text-to-
Speech (TTS) systems in production. Traditional methods, which
rely on phoneme sequence manipulation, are often cumber-
some and error-prone. To address this, we propose Prompt-to-
Correct, an editing-based methodology for pronunciation cor-
rection in TTS systems using voice prompts. Our approach
enables accurate, granular corrections at test-time without the
need for additional training or fine-tuning. Unlike existing
speech editing methods, we eliminate the need for external
alignment to determine edit boundaries. By simply providing
a correctly-pronounced reading of a word in any voice or
accent, our system successfully corrects mispronunciations while
maintaining continuity. Experimental results demonstrate that
our method outperforms traditional baselines and state-of-the-
art speech editing techniques. Speech samples are available at:
https://prompt-to-correct.github.io/P2C