Selected work

Representative Papers

Exploration-Driven Personalized Federated Reinforcement Learning via Intrinsic Motivation

Aug 11, 2026

This work addresses the limitations of existing personalized federated reinforcement learning methods, which overly rely on extrinsic rewards and suffer from insufficient exploration in non-stationary or sparse-reward environments, leading to weak policy personalization and low sample efficiency. To overcome these challenges, we propose the first exploration-driven personalized federated reinforcement learning framework that integrates intrinsic motivation. Specifically, clients leverage Random Network Distillation (RND) to generate intrinsic rewards that enhance local exploration, while the server aggregates only minimal novelty summaries and broadcasts a global exploration prior to coordinate diverse cross-client exploration without compromising privacy. Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods on standard benchmarks as well as in sparse and delayed reward settings, achieving both stronger policy personalization and higher sample efficiency.

0 citationsRead paper

Beyond Global Editing: Per-Instance Disentangled Subspaces for Training-Free Hallucination Mitigation in LVLMs

Aug 10, 2026

This work addresses the prevalent issue of hallucination in large vision-language models (LVLMs), where generated text often contradicts visual input, undermining model reliability. The authors propose a training-free, test-time hallucination suppression framework that dynamically constructs a disentangled hallucination subspace for each input instance and selectively suppresses the most likely hallucinatory directions via adaptive weighted projection, while preserving image-relevant semantics. Moving beyond conventional global editing strategies, this approach achieves instance-level dynamic modeling of hallucinations for the first time, substantially improving both suppression accuracy and generalization. Extensive experiments demonstrate consistent performance gains across diverse LVLM architectures and vision-language benchmarks, confirming the framework’s robustness, versatility, and efficiency.

0 citationsRead paper

Time Present and Time Past: Benchmarking Large Language Models on Temporally Evolving Document Understanding

Aug 09, 2026

This work addresses the challenge that large language models (LLMs) struggle to accurately retrieve and reason over the correct temporal version of evolving official documents—such as legal or tax regulations—corresponding to a given query date. To this end, the authors introduce TIDE, a novel benchmark comprising 644 Bangladeshi customs documents spanning 1969–2025 and 3,050 expert-validated question-answer pairs, which explicitly frames version resolution as a core task in time-aware question answering. Evaluation of nine prominent LLMs across three settings—parametric memory, golden context, and retrieval-augmented generation—employs a rigorous three-judge LLM scoring mechanism with hard date gating and multi-calendar date parsing. Results reveal that even the best-performing model achieves only a 68.5% macro-averaged accuracy, with implicit version resolution at 59.7% and an alarmingly low 26.7% accuracy in recognizing when a provided document version is inapplicable, underscoring models’ overreliance on parametric knowledge rather than authoritative textual sources.

0 citationsRead paper

QuanTiMedAI: Quantum-Enhanced Time-Series Model guided by Agentic AI for Cardiac Arrest Mortality Prediction

Aug 06, 2026

This study addresses the limitation of existing cardiac arrest mortality prediction methods, which predominantly rely on static admission data and fail to capture the dynamic evolution of patients’ physiological states during ICU stays. To overcome this, the authors propose a novel dynamic prediction framework that integrates an agentic large language model (Agentic LLM) for automated extraction of salient clinical features with a low-parameter quantum recurrent neural network for efficient nonlinear temporal modeling. This approach represents the first fusion of AI-driven feature selection and a compact quantum time-series architecture. Evaluated on the MIMIC-IV dataset, the model achieves an AUROC of 0.852 using only 605 parameters—outperforming current state-of-the-art baselines by 2.9%—demonstrating a significant balance between predictive accuracy and model efficiency.

0 citationsRead paper

BnBERT-iPET: Sparse Few-Shot Language Modeling for Bengali via Lottery Ticket Pruning

Aug 05, 2026

This work addresses the challenge of deploying large pre-trained language models for low-resource languages like Bengali, which face prohibitive computational costs and carbon footprints that hinder deployment on resource-constrained devices. The study proposes BnBERT-iPET, a lightweight few-shot language model that integrates the Lottery Ticket Hypothesis-based pruning with iterative pattern-exploiting training (iPET), retaining only 10% of the original BERT parameters. Evaluated on standard Bengali benchmarks, the model achieves up to 90% sparsity while matching the performance of state-of-the-art models such as Bangla Electra, Indic-BERT, and XLM-RoBERTa. This approach significantly reduces computational requirements without sacrificing accuracy, offering a novel and sustainable paradigm for green natural language processing in low-resource linguistic settings.

0 citationsRead paper

Recent publications

Latest Papers

Exploration-Driven Personalized Federated Reinforcement Learning via Intrinsic Motivation

Aug 11, 2026

This work addresses the limitations of existing personalized federated reinforcement learning methods, which overly rely on extrinsic rewards and suffer from insufficient exploration in non-stationary or sparse-reward environments, leading to weak policy personalization and low sample efficiency. To overcome these challenges, we propose the first exploration-driven personalized federated reinforcement learning framework that integrates intrinsic motivation. Specifically, clients leverage Random Network Distillation (RND) to generate intrinsic rewards that enhance local exploration, while the server aggregates only minimal novelty summaries and broadcasts a global exploration prior to coordinate diverse cross-client exploration without compromising privacy. Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods on standard benchmarks as well as in sparse and delayed reward settings, achieving both stronger policy personalization and higher sample efficiency.

0 citationsRead paper

Beyond Global Editing: Per-Instance Disentangled Subspaces for Training-Free Hallucination Mitigation in LVLMs

Aug 10, 2026

This work addresses the prevalent issue of hallucination in large vision-language models (LVLMs), where generated text often contradicts visual input, undermining model reliability. The authors propose a training-free, test-time hallucination suppression framework that dynamically constructs a disentangled hallucination subspace for each input instance and selectively suppresses the most likely hallucinatory directions via adaptive weighted projection, while preserving image-relevant semantics. Moving beyond conventional global editing strategies, this approach achieves instance-level dynamic modeling of hallucinations for the first time, substantially improving both suppression accuracy and generalization. Extensive experiments demonstrate consistent performance gains across diverse LVLM architectures and vision-language benchmarks, confirming the framework’s robustness, versatility, and efficiency.

0 citationsRead paper

Time Present and Time Past: Benchmarking Large Language Models on Temporally Evolving Document Understanding

Aug 09, 2026

This work addresses the challenge that large language models (LLMs) struggle to accurately retrieve and reason over the correct temporal version of evolving official documents—such as legal or tax regulations—corresponding to a given query date. To this end, the authors introduce TIDE, a novel benchmark comprising 644 Bangladeshi customs documents spanning 1969–2025 and 3,050 expert-validated question-answer pairs, which explicitly frames version resolution as a core task in time-aware question answering. Evaluation of nine prominent LLMs across three settings—parametric memory, golden context, and retrieval-augmented generation—employs a rigorous three-judge LLM scoring mechanism with hard date gating and multi-calendar date parsing. Results reveal that even the best-performing model achieves only a 68.5% macro-averaged accuracy, with implicit version resolution at 59.7% and an alarmingly low 26.7% accuracy in recognizing when a provided document version is inapplicable, underscoring models’ overreliance on parametric knowledge rather than authoritative textual sources.

0 citationsRead paper

QuanTiMedAI: Quantum-Enhanced Time-Series Model guided by Agentic AI for Cardiac Arrest Mortality Prediction

Aug 06, 2026

This study addresses the limitation of existing cardiac arrest mortality prediction methods, which predominantly rely on static admission data and fail to capture the dynamic evolution of patients’ physiological states during ICU stays. To overcome this, the authors propose a novel dynamic prediction framework that integrates an agentic large language model (Agentic LLM) for automated extraction of salient clinical features with a low-parameter quantum recurrent neural network for efficient nonlinear temporal modeling. This approach represents the first fusion of AI-driven feature selection and a compact quantum time-series architecture. Evaluated on the MIMIC-IV dataset, the model achieves an AUROC of 0.852 using only 605 parameters—outperforming current state-of-the-art baselines by 2.9%—demonstrating a significant balance between predictive accuracy and model efficiency.

0 citationsRead paper

BnBERT-iPET: Sparse Few-Shot Language Modeling for Bengali via Lottery Ticket Pruning

Aug 05, 2026

This work addresses the challenge of deploying large pre-trained language models for low-resource languages like Bengali, which face prohibitive computational costs and carbon footprints that hinder deployment on resource-constrained devices. The study proposes BnBERT-iPET, a lightweight few-shot language model that integrates the Lottery Ticket Hypothesis-based pruning with iterative pattern-exploiting training (iPET), retaining only 10% of the original BERT parameters. Evaluated on standard Bengali benchmarks, the model achieves up to 90% sparsity while matching the performance of state-of-the-art models such as Bangla Electra, Indic-BERT, and XLM-RoBERTa. This approach significantly reduces computational requirements without sacrificing accuracy, offering a novel and sustainable paradigm for green natural language processing in low-resource linguistic settings.

0 citationsRead paper