Aashish Pokhrel आशिष

News

• Mar 2026 Paper accepted at ICPR 2026 — Negation in Vision-Language Models: A Survey
• Mar 2026 Paper accepted at GRAIL-V @ CVPR 2026 — Negation Matters: Training-Free Negation-Aware Image Retrieval

Publications

• 

  Negation in Vision-Language Models: A Survey

  Aashish Pokhrel, Bipin Ghimire, Prashanna Mani Paudel and Shivanand Venkanna Sheshappanavar

  ICPR 2026

  pdf [comming soon] abstract

  Vision-language models (VLMs) have achieved strong performance by aligning visual and textual representations, enabling success in tasks such as image retrieval, captioning, and visual recognition. However, recent studies show that VLMs struggle with negation, i.e., expressions describing absence or falsity. As a result, models often show affirmation bias, where negated and affirmative statements are treated similarly, leading to semantic collapse and poor reasoning. Providing essential background on VLMs and negation in natural language processing (NLP), this survey reviews prior work on negation in VLMs and organizes it into four main components. First, it examines key challenges that limit negation understanding, including semantic collapse, the NO Syndrome, scaling effects, data scarcity, and the cognitive complexity of negation. Second, it discusses how negation is defined and evaluated in multimodal settings. Third, it reviews methods for improving negation awareness, including data-centric and architectural approaches. Finally, it highlights the practical impact of negation-aware modeling and outlines future research directions.
• 

  Negation Matters: Training-Free Negation-Aware Image Retrieval

  Aashish Pokhrel and Shivanand Venkanna Sheshappanavar

  GRAIL-V @ CVPR 2026

  pdf [comming soon] abstract

  Negation—the linguistic ability to assert the absence of a concept—is a critical bottleneck in vision-language understanding. Vision-language models have demonstrated remarkable success in aligning visual and textual representation across domains such as content moderation, medical image retrieval, and natural language-guided search. Yet, these models consistently fail to handle negation robustly, often retrieving images that contain the very concept that was explicitly excluded. Existing methods address this limitation through fine-tuning on synthetic negation corpora—an approach that is computationally expensive, dataset-dependent, and prone to compromising generalization to unseen distributions. We propose SpaceVLM-DRC, which introduces Dynamic Repulsion with Context Anchoring (DRC) into the SpaceVLM framework as a training-free inference time negation resolution mechanism over a frozen CLIP backbone. This framework first decomposes queries into affirmative, negated, and counterfactual components, then applies dynamic repulsion to push negated concepts away in the embedding space, and finally anchors retrieval within the full-caption context to preserve semantic coherence. SpaceVLM-DRC surpasses state-of-the-art results on the MSRVTT negation retrieval benchmark and achieves performance comparable to fully fine-tuned approaches on the COCO negated retrieval dataset. Crucially, it requires no model retraining while preserving zero-shot generalization on non-negated queries. Our code is available at https://github.com/aashishpokhrel27/spacevlm-drc.

Research Experience

Industry Experience

Education

Skills