Papers by Souvik Kundu
Improving the Throughput of Diffusion-based Large Language Models via a Training-Free Confidence-Aware Calibration (2026.findings-acl)
Copied to clipboard
Jucheng Shen, Gaurav Sarkar, Yeonju Ro, Sharath Nittur Sridhar, Zhangyang Wang, Aditya Akella, Souvik Kundu
| Challenge: | CadLLM is a plug-and-play model-agnostic with KV caching based dLLMs. |
| Approach: | They propose a lightweight adaptive method that can control the generation block size, step size, and threshold based on the average confidence score of unmasked tokens. |
| Outcome: | The proposed method can increase throughput by up to 1.1-2.28x over the state-of-the-art model with competitive accuracy. |
Activation Steering for Chain-of-Thought Compression (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models produce intermediate explanations, commonly referred to as chains of thought (CoTs), but the generated rationales are typically verbose, consuming many additional tokens, and thus degrading throughput and increasing inference energy consumption. |
| Approach: | They propose to generate concise reasoning traces by directly adjusting internal representations via activation steering. |
| Outcome: | The proposed method reduces generated token length by 69.4% across five reasoning benchmarks while maintaining accuracy. |
A Co-Attentive Cross-Lingual Neural Model for Dialogue Breakdown Detection (2020.coling-main)
Copied to clipboard
| Challenge: | Existing models for dialogue breakdown detection do not focus on preventing dialogue breakdowns. |
| Approach: | They propose a model that integrates a pretrained cross-lingual language model and a co-attention network for dialogue breakdown detection. |
| Outcome: | The proposed model outperforms all previous approaches on evaluation metrics in Japanese and English tracks in Dialogue Breakdown Detection Challenge 4 . |
A Nil-Aware Answer Extraction Framework for Question Answering (D18-1)
Copied to clipboard
| Challenge: | Recent research suggests that reading comprehension-based question answering systems assume that every question has a valid answer in the associated passage. |
| Approach: | They propose a novel nil-aware answer span extraction framework that can return Nil or a text span from the associated passage as an answer in a single step. |
| Outcome: | The proposed framework outperforms baseline approaches on a newsQA dataset. |
AFLoRA: Adaptive Freezing of Low Rank Adaptation in Parameter Efficient Fine-Tuning of Large Models (2024.acl-short)
Copied to clipboard
| Challenge: | Pre-trained language models have demonstrated commendable performance on various NLP tasks. |
| Approach: | They propose a Parameter-Efficient Fine-Tuning (PEFT) method that incrementally freezes low-rank matrices during fine-tuning to reduce computation and alleviate over-fitting. |
| Outcome: | The proposed method achieves state-of-the-art performance with an average improvement of 0.85% on the GLUE benchmark while yielding up to 1.86 improvement as opposed to similar PEFT alternatives. |
Evaluation Benchmarks for Spanish Sentence Representations (2022.lrec-1)
Copied to clipboard
Vladimir Araujo, Andrés Carvallo, Souvik Kundu, José Cañete, Marcelo Mendoza, Robert E. Mercer, Felipe Bravo-Marquez, Marie-Francine Moens, Alvaro Soto
| Challenge: | Existing and newly constructed datasets address different tasks from various domains. |
| Approach: | They propose to use Spanish SentEval and Spanish DiscoEval to evaluate stand-alone and discourse-aware sentence representations. |
| Outcome: | The proposed benchmarks evaluate the capabilities of stand-alone and discourse-aware sentence representations in Spanish and show that they are more robust and comparable than previous benchmarks. |
Exploiting Explicit Paths for Multi-hop Reading Comprehension (P19-1)
Copied to clipboard
| Challenge: | Existing approaches to multi-hop reading comprehension do not include multiple sentences or passages. |
| Approach: | They propose a path-based reasoning approach for a multi-hop reading comprehension task . they propose to extract paths from text and compose them to encode them . |
| Outcome: | The proposed model outperforms previous models on the multi-hop Wikihop dataset and can be generalized to the OpenBookQA dataset. |
Learning to Identify Follow-Up Questions in Conversational Question Answering (2020.acl-main)
Copied to clipboard
| Challenge: | Recent work on conversational question answering does not focus on follow-up questions . a practical conversational QA system must understand the conversation history well . |
| Approach: | They propose a three-way attentive pooling network that determines suitability of a follow-up question by capturing pair-wise interactions between associated passage, conversation history, and a candidate follow- up question. |
| Outcome: | The proposed model outperforms baseline systems by significant margins in the follow-up question identification task. |
Mitigating Hallucinations in Vision-Language Models through Image-Guided Head Suppression (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for reducing hallucinations incur a significant increase in latency. |
| Approach: | They propose a task-agnostic attention-guided head suppression strategy that can be seamlessly integrated during inference without incurring significant compute or latency overhead. |
| Outcome: | The proposed approach reduces hallucinations by 2.7x while maintaining F1 and improves throughput by 1.8% compared to existing methods. |
LAWCAT: Efficient Distillation from Quadratic to Linear Attention with Convolution across Tokens for Long Context Modeling (2025.findings-emnlp)
Copied to clipboard
Zeyu Liu, Souvik Kundu, Lianghao Jiang, Anni Li, Srikanth Ronanki, Sravan Babu Bodapati, Gourav Datta, Peter Anthony Beerel
| Challenge: | a novel linearization framework is proposed to reduce the cost of training transformers from scratch. |
| Approach: | They propose a linear attention framework that integrates pre-trained transformers into a performant linear attention architecture. |
| Outcome: | The proposed framework improves performance on mistral-7B with 1K-length sequences and BABILong benchmarks. |
LaMDA: Large Model Fine-Tuning via Spectrally Decomposed Low-Dimensional Adaptation (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Low-rank adaptation (LoRA) fine-tunes large language models due to its significant reduction in trainable parameters, but its backward updates require storing high-dimensional intermediate activations and optimizer states, requiring high peak GPU memory. |
| Approach: | They propose a low-dimensional adaptation approach to fine-tune large language models which freezes a first projection matrix while introducing a lower-dimensional trainable square matrix. |
| Outcome: | The proposed approach reduces trainable parameters and peak GPU memory footprint while preserving low-dimensional trainable square matrix. |
LVLM-Compress-Bench: Benchmarking the Broader Impact of Large Vision-Language Model Compression (2025.findings-naacl)
Copied to clipboard
Souvik Kundu, Anahita Bhiwandiwalla, Sungduk Yu, Phillip Howard, Tiep Le, Sharath Nittur Sridhar, David Cobbley, Hao Kang, Vasudev Lal
| Challenge: | LVLMs have been shown to perform well on simple uni-modal benchmarks, but their detailed study on multi-modal models is still lacking. |
| Approach: | They propose a framework to analyze the impact of compression on LVLMs on multi-modal input driven tasks. |
| Outcome: | The proposed framework analyzes the impact of compression on generative performance of large vision language models on multi-modal input driven tasks. |
LAMB: A Training-Free Method to Enhance the Long-Context Understanding of SSMs via Attention-Guided Token Filtering (2025.acl-short)
Copied to clipboard
Zhifan Ye, Zheng Wang, Kejing Xia, Jihoon Hong, Leshu Li, Lexington Whalen, Cheng Wan, Yonggan Fu, Yingyan Celine Lin, Souvik Kundu
| Challenge: | Recent work attributes performance degradation to an exponential decay in hidden-state memory. |
| Approach: | They propose a token filtering strategy that is training-free and attention-guided . they propose 'LAMB' to preserve critical tokens during inference . |
| Outcome: | The proposed token filtering improves long-context performance by 30.35% over state-of-the-art methods on benchmarks. |