Soumyadeep Jana

Soumyadeep Jana, Sagar Nishad, Sanasam Ranbir Singh

Key-Value (KV) cache remains a major bottleneck for deploying Large Language Models (LLMs) in long-generation tasks. Prior work often applies uniform compression across both prefill and decoding caches, but compressing the prefill cache degrades performance by corrupting critical context. While preserving the prefill cache is essential, decoding-phase compression remains underexplored, with existing methods relying on rigid recency windows or instantaneous attention. Our analysis of attention dynamics reveals strong temporal patterns: critical tokens receive sustained attention over long horizons, while local reasoning involves short-lived bursts. Static heuristics fail to capture this behavior, leading to premature eviction of important tokens or retention of stale ones. We propose Moment-KV, a decoding-time KV cache compression method based on momentum-driven temporal attention aggregation. Our method models token importance as a continuously evolving state, where attention is aggregated with decay, capturing both long-term influence and recent relevance. Experiments show that Moment-KV significantly improves generation fidelity in long-generation tasks (2.3-3.2 %) while maintaining decoding latency.

Soumyadeep Jana, Pulkit Mittal, Sanasam Ranbir Singh

Large vision-language models (LVLMs) often hallucinate objects that are not present in the input image, largely because visual grounding weakens as decoding progresses. Existing inference-time mitigation methods modify logits or hidden states throughout generation, but they suffer from three key limitations: they lack an explicit grounding objective, intervene even when the model is already well-grounded, and use fixed correction strengths that do not adapt to the severity of grounding failure. We propose BRACS (Barrier-Regulated Adaptive Closed-form Steering), a training-free steering framework that addresses these issues through barrier-regulated adaptive closed-form steering. BRACS monitors the model's own attention to measure visual grounding and applies corrections to the hidden states only when grounding deteriorates. The corrective update is computed analytically in closed form, requiring no training of auxiliary networks or model retraining. Experiments on LLaVA-1.5-7B and Qwen-VL-Chat show that BRACS consistently outperforms prior methods on hallucination benchmarks, reducing CHAIR$_s$ by 9.4 points and improving POPE F1 by 2.7 points, while matching or improving performance on four general multimodal benchmarks. BRACS also remains efficient, operating at 80% of greedy decoding throughput and achieving 1.3 times higher speed on average than the baselines.

Soumyadeep Jana, Abhrajyoti Kundu, Sanasam Ranbir Singh

Multimodal sarcasm detection has attracted growing interest due to the rise of multimedia posts on social media. Understanding sarcastic image-text posts often requires external contextual knowledge, such as cultural references or commonsense reasoning. However, existing models struggle to capture the deeper rationale behind sarcasm, relying mainly on shallow cues like image captions or object-attribute pairs from images. To address this, we propose \textbf{MiDRE} (\textbf{Mi}xture of \textbf{D}ual \textbf{R}easoning \textbf{E}xperts), which integrates an internal reasoning expert for detecting incongruities within the image-text pair and an external reasoning expert that utilizes structured rationales generated via Chain-of-Thought prompting to a Large Vision-Language Model. An adaptive gating mechanism dynamically weighs the two experts, selecting the most relevant reasoning path. Unlike prior methods that treat external knowledge as static input, MiDRE selectively adapts to when such knowledge is beneficial, mitigating the risks of hallucinated or irrelevant signals from large models. Experiments on two benchmark datasets show that MiDRE achieves superior performance over baselines. Various qualitative analyses highlight the crucial role of external rationales, revealing that even when they are occasionally noisy, they provide valuable cues that guide the model toward a better understanding of sarcasm.

Soumyadeep Jana, Abhrajyoti Kundu, Sanasam Ranbir Singh

The widespread use of multimodal content on social media has heightened the need for effective sarcasm detection to improve opinion mining. However, existing models rely heavily on large annotated datasets, making them less suitable for real-world scenarios where labeled data is scarce. This motivates the need to explore the problem in a few-shot setting. To this end, we introduce DMDP (Deep Modality-Disentangled Prompt Tuning), a novel framework for few-shot multimodal sarcasm detection. Unlike prior methods that use shallow, unified prompts across modalities, DMDP employs gated, modality-specific deep prompts for text and visual encoders. These prompts are injected across multiple layers to enable hierarchical feature learning and better capture diverse sarcasm types. To enhance intra-modal learning, we incorporate a prompt-sharing mechanism across layers, allowing the model to aggregate both low-level and high-level semantic cues. Additionally, a cross-modal prompt alignment module enables nuanced interactions between image and text representations, improving the model's ability to detect subtle sarcastic intent. Experiments on two public datasets demonstrate DMDP's superior performance in both few-shot and extremely low-resource settings. Further cross-dataset evaluations show that DMDP generalizes well across domains, consistently outperforming baseline methods.

Soumyadeep Jana, Sahil Danayak, Sanasam Ranbir Singh

The growing prevalence of multimodal image-text sarcasm on social media poses challenges for opinion mining systems. Existing approaches rely on full fine-tuning of large models, making them unsuitable to adapt under resource-constrained settings. While recent parameter-efficient fine-tuning (PEFT) methods offer promise, their off-the-shelf use underperforms on complex tasks like sarcasm detection. We propose AdS-CLIP (Adapter-state Sharing in CLIP), a lightweight framework built on CLIP that inserts adapters only in the upper layers to preserve low-level unimodal representations in the lower layers and introduces a novel adapter-state sharing mechanism, where textual adapters guide visual ones to promote efficient cross-modal learning in the upper layers. Experiments on two public benchmarks demonstrate that AdS-CLIP not only outperforms standard PEFT methods but also existing multimodal baselines with significantly fewer trainable parameters.