Papers by Usman Naseem
When the Model Said ‘No Comment’, We Knew Helpfulness Was Dead, Honesty Was Alive, and Safety Was Terrified (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing work uses SFT and MoE to align Large Language Models, but these work face challenges in multi-objective settings. |
| Approach: | They propose a framework that uses prompt-injected fine-tuning to extract axis-specific task features . it deploys a MoCaE module that calibrates expert routing using fractal and natural geometry . |
| Outcome: | The proposed framework achieves significant gains on Alpaca, BeaverTails, TruthfulQA and TruthfulQ with +171.5% win rate and +110.1% truthfulness-informativeness. |
ReflectDiffu: Reflect between Emotion-intent Contagion and Mimicry for Empathetic Response Generation via a RL-Diffusion Framework (2025.acl-long)
Copied to clipboard
| Challenge: | Existing models for empathetic dialogue generation neglect the intricate interplay between emotion and intent, leading to suboptimal controllability of empathy. |
| Approach: | They propose a framework that integrates emotion contagion and intent mimicry to enhance empathetic response generation. |
| Outcome: | The proposed framework outperforms existing models in relevance, controllability, and informativeness. |
Multiview Clickbait Detection via Jointly Modeling Subjective and Objective Preference (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing clickbait detection models rely on analyzing the objective semantics of posts or correlating posts with article content only, but fail to identify and exploit the manipulation intention of clickbaiting from a user’s subjective perspective. |
| Approach: | They propose a multiview clickbait detection model to model subjective and objective preferences simultaneously to capture clickbaiting from a user's subjective perspective. |
| Outcome: | The proposed model outperforms state-of-the-art models on two real-world datasets and shows that it integrates subjective and objective preferences simultaneously. |
Do Personality Traits Interfere? Geometric Limitations of Steering in Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing studies on personality steering in large language models rely on injecting trait-specific steering vectors into the residual stream to control the strength of trait expression. |
| Approach: | They examine the geometric relationships between Big Five personality steering directions by applying geometric conditioning schemes to their steering vectors. |
| Outcome: | The proposed model can be used to steer personality traits in large language models. |
MULSUM: A Multimodal Summarization System with Vis-Aligner and Diversity-Aware Image Selection (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing systems that condense text and images into concise, faithful digests are inefficient and require large fusion transformers. |
| Approach: | They propose a framework that uses image embeddings to generate a visually informed text summary and a Diversity-Aware Image Selector to maximize images-relevance to the summary. |
| Outcome: | The proposed framework outperforms baselines on automatic metrics such as ROUGE and human evaluation shows that selected images act as explanatory evidence rather than ornamental add-ons. |
Reversal of Thought: Enhancing Large Language Models with Preference-Guided Reverse Reasoning Warm-up (2025.acl-long)
Copied to clipboard
| Challenge: | Existing methods to improve LLMs’ logical capabilities involve traceable or verifiable logical sequences that generate more reliable responses yet increase computational costs, or introduce rigid logic template rules, reducing flexibility. |
| Approach: | They propose a plug-and-play reasoning framework that enhances LLMs' logical reasoning abilities during the warm-up phase prior to batch inference. |
| Outcome: | The proposed framework surpasses baselines in both reasoning accuracy and efficiency. |
VITAL: A New Dataset for Benchmarking Pluralistic Alignment in Healthcare (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to align Large Language Models with human values model an averaged or monolithic preference, despite progress in pluralistic alignment, no prior work has focused on health . |
| Approach: | They propose a benchmark dataset to assess and benchmark pluralistic alignment methodologies. |
| Outcome: | The proposed model can model pluralistic views within health domains. |
AlignCultura: Towards Culturally Aligned Large Language Models? (2026.acl-long)
Copied to clipboard
| Challenge: | Existing benchmarks represent early steps toward cultural alignment, yet no benchmarks currently enables systematic evaluation of cultural alignment in line with UNESCO’s principles of cultural diversity w.r.t HHH paradigm. |
| Approach: | Align-Cultura aims to evaluate cultural alignment in large language models . it uses a Query Construction pipeline to reclassify prompts and expand underrepresented domains . response generation pairs prompts with culturally grounded responses . |
| Outcome: | Empirically, culturally fine-tuned models improve joint HHH by 4%–6%, reduce cultural failures by 18%, achieve 10%–12% efficiency gains, and limit leakage to 0.3%. |
POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization (2026.findings-acl)
Copied to clipboard
Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Alexandro Garrido Veliz, P Sam Sahil, Yiran Zhang, Idris Abdulmumin, Marco Antonio Stranisci, Özge Alacam, Cengiz Acarturk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Simona Frenda, Alessandra Teresa Cignarella, Elena Tutubalina, Oleg Rogov, Aung Kyaw Htet, Xintong Wang, Surendrabikram Thapa, Kritesh Rauniyar, Tanmoy Chakraborty, MD Arfeen Zeeshan, Dheeraj Kodati, Satya Keerthi, Sahar Moradizeyveh, Firoj Alam, Md Arid Hasan, Syed Ishtiaque Ahmed, Ye Kyaw Thu, Shantipriya Parida, Ihsan Ayyub Qazi, Lilian Diana Awuor Wanzare, Nelson Odhiambo Onyango, Clemencia Siro, Jane Wanjiru Kimani, Ibrahim Said Ahmad, Adem Chanie Ali, Martin Semmann, Chris Biemann, Shamsuddeen Hassan Muhammad, Seid Muhie Yimam
| Challenge: | polarization is a pervasive threat to democratic institutions, civil discourse, and social cohesion worldwide . most existing datasets focus on English or high-resource languages, reflecting a widespread trend across NLP tasks . |
| Approach: | They propose a multilingual, multicultural, and multi-event dataset with over 110K instances in 22 languages drawn from diverse online platforms and real-world events. |
| Outcome: | The proposed dataset analyzes polarization detection, type, and manifestation using a variety of annotation platforms adapted to each cultural context. |
Fairness Evaluation and Inference Level Mitigation in LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models display undesirable behaviors embedded in their internal representations, undermining fairness, inconsistency drift, and the propagation of unwanted patterns during extended dialogues. |
| Approach: | They propose a pruning-based framework that detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. |
| Outcome: | The proposed framework detects context-aware neuron activations and applies adaptive masking to modulate their influence during generation. |
Can LLM Agents Maintain a Persona in Discourse? (2025.emnlp-main)
Copied to clipboard
Pranav Bhandari, Nicolas Fay, Michael J Wise, Amitava Datta, Stephanie Meek, Usman Naseem, Mehwish Nasim
| Challenge: | Large language models are often subjected to context-shifting behaviour, resulting in a lack of consistent and interpretable personality-aligned interactions. |
| Approach: | They propose to use two conversation agents to generate a discourse with an assigned personality from the OCEAN framework and then use multiple judge agents to infer original traits. |
| Outcome: | The proposed model is based on two conversation agents with a personality assigned from the OCEAN framework and then multiple judge agents to infer the original traits assigned. |
Activation-Space Personality Steering: Hybrid Layer Selection for Stable Trait Control in LLMs (2026.eacl-long)
Copied to clipboard
| Challenge: | Personality-aware LLMs exhibit implicit personalities in their generation, but reliably controlling or aligning these traits to meet specific needs remains an open challenge. |
| Approach: | They propose a pipeline that extracts hidden state activations from transformer layers using the Big Five Personality Traits framework. |
| Outcome: | The proposed model extracts hidden state activations from transformer layers using the Big Five personality traits (Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism) |
Pluralistic Alignment for Healthcare: A Role-Driven Framework (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches to align large language models fail to reflect diversity in sensitive domains like healthcare, where personal, cultural, and situational factors shape pluralism. |
| Approach: | They propose a lightweight, generalizable, pluralistic alignment approach to model diverse perspectives and values in open and closed models. |
| Outcome: | The proposed approach advances the pluralistic alignment for all three modes across seven varying-sized open and closed models. |
Are Large Language Models Economically Viable for Industry Deployment? (2026.acl-industry)
Copied to clipboard
Abdullah Mohammad, Sushant Kumar Ray, Pushkar Arora, Rafiq Ali, Ebad Shabbir, Gautam Siddharth Kashyap, Jiechao Gao, Usman Naseem
| Challenge: | Generative AI is increasingly deployed in healthcare, financial analytics, and conversational automation. |
| Approach: | They propose a framework that evaluates large language models across their full lifecycle on legacy GPUs. |
| Outcome: | The proposed framework evaluates LLMs across their full lifecycle on legacy GPUs. |
Activation Decomposition and Steering for LLM Backdoor Remediation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to defending against LLM backdoors rely on auxiliary models or safety-related datasets. |
| Approach: | They propose a method which contrasts benign and poisoned settings to decompose feature vectors for steering without auxiliary models or datasets. |
| Outcome: | The proposed method achieves better defense qualities than existing steering strategies. |
SafeConstellations: Mitigating Over-Refusals in LLMs Through Task-Aware Representation Steering (2026.acl-long)
Copied to clipboard
| Challenge: | Current safety alignment methods fail to identify intended benign task before refusing to respond. |
| Approach: | They propose a method that uses inference-time trajectory-shifting to guide model behavior . they show that LLMs persist in refusing inputs containing harmful content . |
| Outcome: | The proposed approach reduces over-refusals with minimal impact on utility. |
From Generation to Detection: A Multimodal Multi-Task Dataset for Benchmarking Health Misinformation (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Infodemics and health misinformation have significant negative impact on individuals and society . generative AI has significantly accelerated the spread and expanded the reach of health misinfo . |
| Approach: | MM-Health is a large scale multimodal misinformation dataset in the health domain . it includes human-generated multimodal information and AI-generated multiplemodal information . |
| Outcome: | MM-Health is a large scale misinformation dataset in the health domain . it includes human-generated multimodal information and AI-generated content . |
GameTox: A Comprehensive Dataset and Analysis for Enhanced Toxicity Detection in Online Gaming Communities (2025.naacl-short)
Copied to clipboard
| Challenge: | Existing methods to detect toxic behavior in online gaming environments are limited by utterance-level annotation. |
| Approach: | They propose to annotate game chat utterances for toxicity detection through intent classification and slot filling. |
| Outcome: | The proposed model improves the detection of toxic speech in online gaming environments and reveals limitations of current models. |
Causal Intervention for Abstractive Related Work Generation (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing models ignore the inherent causality during related work generation, leading to spurious correlations which downgrade the models’ generation quality and generalizability. |
| Approach: | They propose a Causal Intervention Module for Related Work Generation (CaM) that captures causal relationships in related work generation and implements causal interventions to mitigate the negative impact of spurious correlations. |
| Outcome: | The proposed framework improves the quality and coherence of generated related work by capturing causalities in the generation process. |
Accuracy meets Diversity in a News Recommender System (2022.coling-1)
Copied to clipboard
| Challenge: | Existing news recommender systems use news stories that users have read in the past to infer their interests and preferences. |
| Approach: | They propose a two-tower architecture that learns news representation through a news item tower and users’ representations through s query towers. |
| Outcome: | The proposed architecture achieves a balance between accuracy and diversity on two news datasets. |
Revealing the Truth with ConLLM for Detecting Multi-Modal Deepfakes (2026.findings-eacl)
Copied to clipboard
Gautam Siddharth Kashyap, Harsh Joshi, Niharika Jain, Ebad Shabbir, Jiechao Gao, Nipun Joshi, Usman Naseem
| Challenge: | Existing methods for deepfake detection suffer from two limitations: modality fragmentation and shallow inter-modal reasoning. |
| Approach: | They propose a framework for multimodal deepfake detection that uses contrastive learning and large language models to mitigate modality fragmentation and refine embeddings to address shallow inter-modal reasoning. |
| Outcome: | ConLLM reduces audio deepfake EER by 50%, improves video accuracy by 8%, and achieves approximately 9% accuracy gains in audio-visual tasks. |
Too Helpful, Too Harmless, Too Honest or Just Right? (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods optimize for individual alignment dimensions in isolation, leading to trade-offs and inconsistent behavior. |
| Approach: | They propose a modular alignment framework that integrates a Mixture of Calibrated Experts (MoCaE) within the Transformer architecture. |
| Outcome: | The proposed framework outperforms baselines on three alignment benchmarks, achieving 32.5% win rate, 33.9% safety score, and 28.4% truthfulness. |
Framing Political Bias in Multilingual LLMs Across Pakistani Languages (2026.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) shape public discourse, yet most evaluations of economic and political bias focus on high-resource Western languages and contexts. |
| Approach: | They propose to use a culturally adapted Political Compass Test to evaluate political bias in 13 state-of-the-art LLMs across five Pakistani languages. |
| Outcome: | The proposed framework captures ideological stance (economic/social axes) and stylistic framing (content, tone, emphasis) in 13 state-of-the-art LLMs across five Pakistani languages. |
Benchmarking Large Language Models for Cryptanalysis and Side-Channel Vulnerabilities (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Recent advances in Large Language Models (LLMs) have transformed natural language understanding and generation, leading to extensive benchmarking across diverse tasks. |
| Approach: | They evaluate the cryptanalytic potential of stateoftheart LLMs on ciphertexts produced by a range of cryptographic algorithms. |
| Outcome: | The proposed model can decrypt plaintexts produced by a range of cryptographic algorithms using zeroshot and fewshot settings along with chainofthought prompting. |
Towards Visually Grounded Multimodal Summarization via Cross-Modal Transformer and Gated Attention (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for multimodal summarization often inject shallow visual features into deep models, leading to representational mismatches and weak cross-modal grounding. |
| Approach: | They propose a framework that performs text summarization and representative image selection . a deep visual processor aligns the visual encoder with the language model at corresponding depths . |
| Outcome: | The proposed framework produces more accurate, visually grounded summaries and selects more representative images. |
LLMs on a Budget? Say HOLA (2025.emnlp-industry)
Copied to clipboard
Zohaib Hasan Siddiqui, Jiechao Gao, Ebad Shabbir, Mohammad Anas Azeez, Rafiq Ali, Gautam Siddharth Kashyap, Usman Naseem
| Challenge: | Current solutions such as quantization, pruning, and Retrieval-Augmented Generation (RAG) offer only partial optimizations and often sacrifice accuracy, speed, or generality. |
| Approach: | They propose an end-to-end optimization framework for efficient LLM deployment . it leverages Hierarchical Speculative Decoding (HSD) for faster inference without quality loss. |
| Outcome: | HOLA delivers +17.6% EMA on GSM8K, +10.5% MCA on ARC, and reduced latency and memory on edge devices like Jetson Nano. |
TurnBench-MS: A Benchmark for Evaluating Multi-Turn, Multi-Step Reasoning in Large Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing benchmarks focus on single-turn or single-step tasks, failing to capture iterative reasoning in real-world settings. |
| Approach: | They propose a benchmark that evaluates multi-turn, multi-step reasoning through an interactive code-breaking task inspired by the "Turing Machine Board Game" the best model achieves 84% accuracy in Classic mode, but performance drops to 18% in Nightmare mode. |
| Outcome: | The new benchmark evaluates multi-turn, multi-step reasoning through an interactive code-breaking task inspired by the "Turing Machine Board Game" the best model achieves 84% accuracy in Classic mode, but performance drops to 18% in Nightmare mode. |
Do Large Language Models Reflect Demographic Pluralism in Safety? (2026.findings-eacl)
Copied to clipboard
Usman Naseem, Gautam Siddharth Kashyap, Sushant Kumar Ray, Rafiq Ali, Ebad Shabbir, Abdullah Mohammad
| Challenge: | Existing datasets that focus on demographics and safety are narrow in their annotator pools. |
| Approach: | They propose to decouple value framing from responses by modeling pluralism directly at the prompt level. |
| Outcome: | Demo-SafetyBench decouples value framing from responses to model pluralism at the prompt level. |
Do Clinical Question Answering Systems Really Need Specialised Medical Fine Tuning? (2026.eacl-industry)
Copied to clipboard
Sushant Kumar Ray, Gautam Siddharth Kashyap, Sahil Tripathi, Nipun Joshi, Vijay Govindarajan, Rafiq Ali, Jiechao Gao, Usman Naseem
| Challenge: | Clinical Question-Answering (CQA) industry systems rely on Large Language Models (LLMs). |
| Approach: | They propose a framework that applies alignment at inference time rather than through SFT to help CQA users achieve consistent reasoning. |
| Outcome: | MEDASSESS-X improves Accuracy, Factual Consistency and Safety by up to 50%. |
Measuring What Matters Beyond Text: Evaluating Multimodal Summaries by Quality, Alignment, and Diversity (2026.findings-acl)
Copied to clipboard
| Challenge: | MLLMs have facilitated multimodal summarization with multimodal outputs, but their evaluation is fragmented . MM-Eval integrates assessments of textual quality, cross-modal alignment, and visual diversity . |
| Approach: | They propose a unified evaluation framework that integrates assessments of textual quality, cross-modal alignment, and visual diversity. |
| Outcome: | The proposed framework improves over heuristic aggregation baselines and provides an interpretable, reference-weak framework for comparative evaluation of multimodal summaries. |
ExtremeAIGC: Benchmarking LMM Vulnerability to AI-Generated Extremist Content (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing datasets for evaluating LMM robustness lack exploration of extremist content . existing models lack diverse image generation models and comprehensive coverage of historical events . |
| Approach: | They propose a benchmark dataset to assess LMM models against extremist content . ExtremeAIGC simulates real-world events and malicious use cases . |
| Outcome: | a new benchmark dataset and evaluation framework assesses LMM models against extremist content. |
BeefBot: Harnessing Advanced LLM and RAG Techniques for Providing Scientific and Technology Solutions to Beef Producers (2025.coling-demos)
Copied to clipboard
| Challenge: | Generic Large Language Models (LLMs) are useful for information retrieval but often hallucinate and fail to deliver tailored solutions to the specific needs of beef producers. |
| Approach: | They propose to use Retrieval-Augmented Generation and fine-tuning to build a chatbot for beef producers that retrieves latest agricultural technologies and scientific insights. |
| Outcome: | The proposed chatbot retrieves latest agricultural technologies, practices and scientific insights to provide rapid, domain-specific advice. |
Leveraging Taxonomy and LLMs for Improved Multimodal Hierarchical Classification (2025.coling-main)
Copied to clipboard
Shijing Chen, Mohamed Reda Bouadjenek, Usman Naseem, Basem Suleiman, Shoaib Jameel, Flora Salim, Hakim Hacid, Imran Razzak
| Challenge: | Multi-level Hierarchical Classification (MLHC) is a critical tool in modern data analysis. |
| Approach: | They propose a taxonomy-embedded transitional LLM-agnostic framework for multimodality classification that leverages large language models to enforce consistency across hierarchical levels. |
| Outcome: | The proposed framework improves on the MEP-3M dataset with various hierarchical levels compared to conventional models. |
Truth, Trust, and Trouble: Medical AI on the Edge (2025.emnlp-industry)
Copied to clipboard
Mohammad Anas Azeez, Rafiq Ali, Ebad Shabbir, Zohaib Hasan Siddiqui, Gautam Siddharth Kashyap, Jiechao Gao, Usman Naseem
| Challenge: | Large Language Models (LLMs) are promising for transforming digital health applications . but ensuring they meet industry standards for factual accuracy, usefulness, and safety remains a challenge . |
| Approach: | They present a framework to assess large language models' accuracy, usefulness, and safety . they assess models' honesty, helpfulness, harmlessness and domain-specific tuning . |
| Outcome: | The proposed framework assesses models across honesty, helpfulness, and harmlessness . AlpaCare-13B achieves highest accuracy (91.7%) and harmlessity (0.92) . |
LLM-Based Multi-Task Bangla Hate Speech Detection: Type, Severity, and Target (2026.acl-long)
Copied to clipboard
| Challenge: | Existing work on social media platforms is limited in its ability to detect hate speech . a lack of reliable and scalable automated hate speech detection systems is a challenge for low-resource languages like Bangla. |
| Approach: | They propose to use a single-task, single-targeted, single language dataset to identify hate speech in Bangla. |
| Outcome: | The proposed dataset is the largest manually annotated Bangla hate-speech dataset to date. |
XGUARD: A Graded Benchmark for Evaluating Safety Failures of Large Language Models on Extremist Content (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing safety evaluations rely on binary labels, overlooking the nuanced risk these outputs pose. |
| Approach: | They propose a framework to assess the severity of extremist content generated by Large Language Models (LLMs) it categorizes model responses into five danger levels (0–4) defined by degree of extremism endorsement . |
| Outcome: | The proposed framework categorizes model responses into five danger levels (0–4) defined by degree of extremist endorsement, enabling nuanced analysis of failure frequency and severity. |
Health-ORSC-Bench: A Benchmark for Measuring Over-Refusal and Safety Completion in Health Context (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing safety alignment benchmarks fail to evaluate Safe Completion: the model’s ability to maximise helpfulness on dual-use or borderline queries without crossing into actionable harm. |
| Approach: | They propose a large-scale benchmark to measure Over-Refusal and Safe Completion quality in healthcare. |
| Outcome: | The framework evaluates 30 state-of-the-art LLMs including GPT-5 and Claude-4. |
Can Large Language Models Enhance Predictions of Disease Progression? Investigating Through Disease Network Link Prediction (2024.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have made significant strides in various tasks, yet their effectiveness in predicting disease progression remains relatively unexplored. |
| Approach: | They propose a large language model with graph prompting and Retrieval-augmented generation to enhance the prediction performance of disease comorbidity within disease networks. |
| Outcome: | The proposed model outperforms Graph Neural Networks and Graph Prompts and Retrieval-Augmented Generation models in disease progression prediction tasks. |
Analyzing the Dynamics of Climate Change Discourse on Twitter: A New Annotated Corpus and Multi-Aspect Classification (2024.lrec-main)
Copied to clipboard
Shuvam Shiwakoti, Surendrabikram Thapa, Kritesh Rauniyar, Akshyat Shah, Aashish Bhandari, Usman Naseem
| Challenge: | a lack of data on climate change discourse has highlighted the need for further advancement . a new study examines the discourse on social media platforms that ignores climate change . |
| Approach: | They analyze climate change discourse on Twitter using a meticulously annotated dataset . they find relevance, stance, hate speech, direction of hate, humor and humor are key aspects . |
| Outcome: | The proposed method combines annotated tweets with a dataset of 15,309 tweets . it reveals tweet distribution patterns, stance prevalence, and hate speech trends . |