arXiv subjects

Gautam Rajendrakumar Gare

Publications and source records attributed to Gautam Rajendrakumar Gare.

13 recordsLinked to original sources

arXiv2026

Your Model Already Knows Don't Teach It, Learn to Ask It: Soft Prompting for Few-Shot Adaptation of Vision-Language Models

We address few-shot object detection with vision-language models (VLMs) in out-of-domain settings such as aerial, industrial, and medical imagery, using only ten annotated images for supervision. Existing adaptation methods are discrete prompt optimization and LoRA fine-tuning. We revisit a third option: soft prompting, where a small number of continuous prompt tokens are optimized while the pretrained backbone remains frozen. We identify two key design choices. First, placing prompt tokens at the cross-modal boundary between visual and text tokens outperforms other placements (10.0 vs. 8.4 mAP). Second, initializing prompts from the empty space token outperforms semantic and random initialization. With these choices, one to three learned tokens (7,168 parameters on average) match the best LoRA configuration on Roboflow20-VL (14.2 mAP, 10-shot) while training over 20,000x fewer parameters. Soft prompting remains harder to optimize, exhibiting higher variance across random seeds. Unlike LoRA, however, it causes no forgetting: the LoRA rank matching our accuracy reduces NaturalBench VQA accuracy by 35% relative, rising to 56% at the largest rank, whereas soft prompting leaves pretrained performance unchanged. The learned tokens behave like prompts rather than weights. They transfer to a newer model without retraining (+0.8 mAP on Qwen3.5-9B) and can be verbalized into readable prompts competitive with prompt-search methods (matching DetPO and outperforming GEPA). The approach also extends beyond detection. On RoboCasa manipulation tasks, the frozen $π_{0.5}$ vision-language-action policy benefits from soft prompting, matching the LoRA baseline on two of three tasks when tokens are placed at the gradient bottleneck. These results suggest modern VLMs already encode much of what is needed for specialized domains; the challenge is learning how to ask.

cs.CV↗

arXiv2026

Ask Twice, Look Twice: Prompt Echoing Resolves the Question-First Paradox in Vision-Language Models

Where should the question go in a vision-language model (VLM) prompt: before the image or after it? Intuition says before: knowing what is asked should tell the model where to look. Yet across visual question answering benchmarks, question-first prompting consistently underperforms the image-first ordering recommended for frontier VLMs, a phenomenon we term the question-first paradox. We trace this paradox to a conflict between two stages of VLM computation. Logit-lens and attention probes show that question-first prompting steers perception, shifting image patch representations toward question-relevant concepts. But downstream, stranded behind hundreds of image tokens, the question is barely attended by the answer token, which instead commits to image-driven, often wrong answers. Causal attention knockout confirms that the answer reads the question only when it follows the image. This diagnosis yields a training-free fix: question echoing, restating the question on both sides of the image so one copy steers perception while the other is available at answer time. A similar division of labor appears in a fifty-year-old finding on human 'adjunct questions', where repeating a question before and after a passage improves comprehension. Echoing the image as well brings further gains by restoring the whole-image view otherwise lost by a causal decoder. The paradox holds across five open VLMs, costing up to 17.5 group-accuracy points. Echoed prompts recover most of the gap and, on NaturalBench and Winoground, surpass the best single-pass ordering by up to 19 group-accuracy points on Winoground, with no training, fine-tuning, or architecture change. The paradox reveals a tension between steering what a model sees and preserving access to what it was asked; echoing resolves this through prompt design. Project Page: https://rakshanda-cmu.github.io/ask-twice-look-twice/

cs.CV↗

arXiv2026

Attributes Should Come from Images, Not Class Names: Distribution-Conditioned Attribute Selection for Vision-Language Models

A popular route to interpretable zero-shot classification asks a large language model (LLM) to describe each class name and prompts CLIP with the resulting descriptors. We show that these descriptors carry little visual evidence of their own: removing the class name from the prompt collapses ImageNet accuracy from 59.5% to 15.5%. The diagnosis is that the descriptors are conditioned on the label rather than on the images, so they describe the concept in general and mislead exactly when the data shifts; an LLM insists that strawberries are red, but every strawberry in ImageNet-Sketch is a colorless line drawing. We therefore select attributes from the target image collection instead: we score a large attribute pool against the images in CLIP's joint embedding space and keep the top-scoring attributes per class. Selected this way, class-name-free attribute prompts reach 23.8% on ImageNet (against 15.5% for LLM descriptors), the gain holds on four shifted ImageNet variants, and reselecting from the LLM's own pool isolates the selection mechanism as the cause. With one image per class, the selected attributes outperform the prompt-tuning method CoOp by 3 points while fitting in under a minute instead of 14 hours, with no learned soft prompt to obscure the decision. Because the attribute set is chosen by the data, it doubles as a readable summary of a dataset, which we use to describe distribution shift in words. Our code and results are available on our project page: https://ggare-cmu.github.io/AttributeSelect/

cs.CV↗

arXiv2026

DetPO: In-Context Learning with Multi-Modal LLMs for Few-Shot Object Detection

Multi-Modal LLMs (MLLMs) demonstrate strong visual grounding capabilities on popular object detection benchmarks like OdinW-13 and RefCOCO. However, state-of-the-art models still struggle to generalize to out-of-distribution classes, tasks and imaging modalities not typically found in their pre-training. While in-context prompting is a common strategy to improve performance across diverse tasks, we find that it often yields lower detection accuracy than prompting with class names alone. This suggests that current MLLMs cannot yet effectively leverage few-shot visual examples and rich textual descriptions for object detection. Since frontier MLLMs are typically only accessible via APIs, and state-of-the-art open-weights models are prohibitively expensive to fine-tune on consumer-grade hardware, we instead explore black-box prompt optimization for few-shot object detection. To this end, we propose Detection Prompt Optimization (DetPO), a gradient-free test-time optimization approach that refines text-only prompts by maximizing detection accuracy on few-shot visual training examples while calibrating prediction confidence. Our proposed approach yields consistent improvements across generalist MLLMs on Roboflow20-VL and LVIS, outperforming prior black-box approaches by up to 9.7 mAP. Our code and optimized prompts are available at https://ggare-cmu.github.io/DetPO/

cs.CV↗

arXiv2025

Label Uncertainty for Ultrasound Segmentation

In medical imaging, inter-observer variability among radiologists often introduces label uncertainty, particularly in modalities where visual interpretation is subjective. Lung ultrasound (LUS) is a prime example-it frequently presents a mixture of highly ambiguous regions and clearly discernible structures, making consistent annotation challenging even for experienced clinicians. In this work, we introduce a novel approach to both labeling and training AI models using expert-supplied, per-pixel confidence values. Rather than treating annotations as absolute ground truth, we design a data annotation protocol that captures the confidence that radiologists have in each labeled region, modeling the inherent aleatoric uncertainty present in real-world clinical data. We demonstrate that incorporating these confidence values during training leads to improved segmentation performance. More importantly, we show that this enhanced segmentation quality translates into better performance on downstream clinically-critical tasks-specifically, estimating S/F oxygenation ratio values, classifying S/F ratio change, and predicting 30-day patient readmission. While we empirically evaluate many methods for exposing the uncertainty to the learning model, we find that a simple approach that trains a model on binarized labels obtained with a (60%) confidence threshold works well. Importantly, high thresholds work far better than a naive approach of a 50% threshold, indicating that training on very confident pixels is far more effective. Our study systematically investigates the impact of training with varying confidence thresholds, comparing not only segmentation metrics but also downstream clinical outcomes. These results suggest that label confidence is a valuable signal that, when properly leveraged, can significantly enhance the reliability and clinical utility of AI in medical imaging.

eess.IV↗

arXiv2025

Activation Reward Models for Few-Shot Model Alignment

Aligning Large Language Models (LLMs) and Large Multimodal Models (LMMs) to human preferences is a central challenge in improving the quality of the models' generative outputs for real-world applications. A common approach is to use reward modeling to encode preferences, enabling alignment via post-training using reinforcement learning. However, traditional reward modeling is not easily adaptable to new preferences because it requires a separate reward model, commonly trained on large preference datasets. To address this, we introduce Activation Reward Models (Activation RMs) -- a novel few-shot reward modeling method that leverages activation steering to construct well-aligned reward signals using minimal supervision and no additional model finetuning. Activation RMs outperform existing few-shot reward modeling approaches such as LLM-as-a-judge with in-context learning, voting-based scoring, and token probability scoring on standard reward modeling benchmarks. Furthermore, we demonstrate the effectiveness of Activation RMs in mitigating reward hacking behaviors, highlighting their utility for safety-critical applications. Toward this end, we propose PreferenceHack, a novel few-shot setting benchmark, the first to test reward models on reward hacking in a paired preference format. Finally, we show that Activation RM achieves state-of-the-art performance on this benchmark, surpassing even GPT-4o.

cs.CV↗

arXiv2025

Improving Model's Interpretability and Reliability using Biomarkers

Accurate and interpretable diagnostic models are crucial in the safety-critical field of medicine. We investigate the interpretability of our proposed biomarker-based lung ultrasound diagnostic pipeline to enhance clinicians' diagnostic capabilities. The objective of this study is to assess whether explanations from a decision tree classifier, utilizing biomarkers, can improve users' ability to identify inaccurate model predictions compared to conventional saliency maps. Our findings demonstrate that decision tree explanations, based on clinically established biomarkers, can assist clinicians in detecting false positives, thus improving the reliability of diagnostic models in medicine.

cs.HC↗

arXiv2022

Learning Generic Lung Ultrasound Biomarkers for Decoupling Feature Extraction from Downstream Tasks

Contemporary artificial neural networks (ANN) are trained end-to-end, jointly learning both features and classifiers for the task of interest. Though enormously effective, this paradigm imposes significant costs in assembling annotated task-specific datasets and training large-scale networks. We propose to decouple feature learning from downstream lung ultrasound tasks by introducing an auxiliary pre-task of visual biomarker classification. We demonstrate that one can learn an informative, concise, and interpretable feature space from ultrasound videos by training models for predicting biomarker labels. Notably, biomarker feature extractors can be trained from data annotated with weak video-scale supervision. These features can be used by a variety of downstream Expert models targeted for diverse clinical tasks (Diagnosis, lung severity, S/F ratio). Crucially, task-specific expert models are comparable in accuracy to end-to-end models directly trained for such target tasks, while being significantly lower cost to train.

eess.IV↗

arXiv2022

Dense Pixel-Labeling for Reverse-Transfer and Diagnostic Learning on Lung Ultrasound for COVID-19 and Pneumonia Detection

We propose using a pre-trained segmentation model to perform diagnostic classification in order to achieve better generalization and interpretability, terming the technique reverse-transfer learning. We present an architecture to convert segmentation models to classification models. We compare and contrast dense vs sparse segmentation labeling and study its impact on diagnostic classification. We compare the performance of U-Net trained with dense and sparse labels to segment A-lines, B-lines, and Pleural lines on a custom dataset of lung ultrasound scans from 4 patients. Our experiments show that dense labels help reduce false positive detection. We study the classification capability of the dense and sparse trained U-Net and contrast it with a non-pretrained U-Net, to detect and differentiate COVID-19 and Pneumonia on a large ultrasound dataset of about 40k curvilinear and linear probe images. Our segmentation-based models perform better classification when using pretrained segmentation weights, with the dense-label pretrained U-Net performing the best.

eess.IV↗

arXiv2022

The Role of Pleura and Adipose in Lung Ultrasound AI

In this paper, we study the significance of the pleura and adipose tissue in lung ultrasound AI analysis. We highlight their more prominent appearance when using high-frequency linear (HFL) instead of curvilinear ultrasound probes, showing HFL reveals better pleura detail. We compare the diagnostic utility of the pleura and adipose tissue using an HFL ultrasound probe. Masking the adipose tissue during training and inference (while retaining the pleural line and Merlin's space artifacts such as A-lines and B-lines) improved the AI model's diagnostic accuracy.

eess.IV↗

arXiv2022

Weakly Supervised Contrastive Learning for Better Severity Scoring of Lung Ultrasound

With the onset of the COVID-19 pandemic, ultrasound has emerged as an effective tool for bedside monitoring of patients. Due to this, a large amount of lung ultrasound scans have been made available which can be used for AI based diagnosis and analysis. Several AI-based patient severity scoring models have been proposed that rely on scoring the appearance of the ultrasound scans. AI models are trained using ultrasound-appearance severity scores that are manually labeled based on standardized visual features. We address the challenge of labeling every ultrasound frame in the video clips. Our contrastive learning method treats the video clip severity labels as noisy weak severity labels for individual frames, thus requiring only video-level labels. We show that it performs better than the conventional cross-entropy loss based training. We combine frame severity predictions to come up with video severity predictions and show that the frame based model achieves comparable performance to a video based TSM model, on a large dataset combining public and private sources.

eess.IV↗

arXiv2021

Exploiting Class Similarity for Machine Learning with Confidence Labels and Projective Loss Functions

Class labels used for machine learning are relatable to each other, with certain class labels being more similar to each other than others (e.g. images of cats and dogs are more similar to each other than those of cats and cars). Such similarity among classes is often the cause of poor model performance due to the models confusing between them. Current labeling techniques fail to explicitly capture such similarity information. In this paper, we instead exploit the similarity between classes by capturing the similarity information with our novel confidence labels. Confidence labels are probabilistic labels denoting the likelihood of similarity, or confusability, between the classes. Often even after models are trained to differentiate between classes in the feature space, the similar classes' latent space still remains clustered. We view this type of clustering as valuable information and exploit it with our novel projective loss functions. Our projective loss functions are designed to work with confidence labels with an ability to relax the loss penalty for errors that confuse similar classes. We use our approach to train neural networks with noisy labels, as we believe noisy labels are partly a result of confusability arising from class similarity. We show improved performance compared to the use of standard loss functions. We conduct a detailed analysis using the CIFAR-10 dataset and show our proposed methods' applicability to larger datasets, such as ImageNet and Food-101N.

cs.CV↗

arXiv2020

W-Net: Dense Semantic Segmentation of Subcutaneous Tissue in Ultrasound Images by Expanding U-Net to Incorporate Ultrasound RF Waveform Data

We present W-Net, a novel Convolution Neural Network (CNN) framework that employs raw ultrasound waveforms from each A-scan, typically referred to as ultrasound Radio Frequency (RF) data, in addition to the gray ultrasound image to semantically segment and label tissues. Unlike prior work, we seek to label every pixel in the image, without the use of a background class. To the best of our knowledge, this is also the first deep-learning or CNN approach for segmentation that analyses ultrasound raw RF data along with the gray image. International patent(s) pending [PCT/US20/37519]. We chose subcutaneous tissue (SubQ) segmentation as our initial clinical goal since it has diverse intermixed tissues, is challenging to segment, and is an underrepresented research area. SubQ potential applications include plastic surgery, adipose stem-cell harvesting, lymphatic monitoring, and possibly detection/treatment of certain types of tumors. A custom dataset consisting of hand-labeled images by an expert clinician and trainees are used for the experimentation, currently labeled into the following categories: skin, fat, fat fascia/stroma, muscle and muscle fascia. We compared our results with U-Net and Attention U-Net. Our novel \emph{W-Net}'s RF-Waveform input and architecture increased mIoU accuracy (averaged across all tissue classes) by 4.5\% and 4.9\% compared to regular U-Net and Attention U-Net, respectively. We present analysis as to why the Muscle fascia and Fat fascia/stroma are the most difficult tissues to label. Muscle fascia in particular, the most difficult anatomic class to recognize for both humans and AI algorithms, saw mIoU improvements of 13\% and 16\% from our W-Net vs U-Net and Attention U-Net respectively.

eess.IV↗