Publications

Bold denotes my own name. ^*denotes equal contribution (shared first authorship). Full list on Google Scholar.

Journal and Conference Papers

• CAT-GS: Balanced Multimodal Learning via Calibrated Gating and Fusion Surgery
  Mahir Shahriar Tamim, S. Khan, M. S. Alim, T. A. Khan, S. Rahman, N. Mohammed
  A training method that stops one input type from dominating multimodal learning, so the model makes better use of every modality without any change to its architecture.
  Neurocomputing (Q1, IF 6.7) [Paper] [arXiv] ▪ Multimodal
• Can LLMs Follow the Pulse of a Crisis? Evaluating Crisis Sentiment in Bangladesh's July Uprising
  M. S. Alim^*, Mahir Shahriar Tamim^*, T. A. Khan, S. Khan, R. F. Duti, S. Z. Ridoy, M. A. Moni
  A dataset of about 200K Bangla social media comments from the 2024 uprising, labelled by native speakers, showing that even strong LLMs struggle to read public sentiment during a fast-moving crisis.
  AACL 2026 Hengqin, China [arXiv] [PDF] ▪ Dataset ▪ NLP ▪ LLM Evaluation
• Ensemble and Temporal Feature-Based Framework for Rainfall Classification in Bangladesh
  Mahir Shahriar Tamim, M. S. Alim, M. Rahman, T. A. Khan, M. M. Anwar
  Combines ensemble learning with engineered temporal features to classify rainfall regimes across Bangladesh from historical meteorological records.
  PLOS ONE (Q1, IF 2.6) [Paper] ▪ Applied ML
• Multi-Teacher Knowledge Distillation and Ensemble Algorithm for Efficient Brain Tumor Classification in Resource-Constrained Environments with Explainable AI
  M. S. Alim, Mahir Shahriar Tamim, S. Sarkar, M. A. Yousuf, A. S. Al-Moisheer, S. A. Alyami, M. A. Moni
  Distils an ensemble of teacher networks into a compact student for brain tumour classification on constrained hardware, with explainability retained end to end.
  Knowledge-Based Systems (Q1, IF 8.0) [Paper] ▪ Medical AI
• When a Nation Speaks: Machine Learning and NLP in People's Sentiment Analysis During Bangladesh's 2024 Mass Uprising
  M. S. Alim, Mahir Shahriar Tamim, M. Rahman, T. A. Khan, M. M. Anwar
  Large-scale sentiment and discourse analysis of social media during Bangladesh's 2024 mass uprising.
  IEEE ICCIT 2025 [arXiv] ▪ NLP

Workshop Papers

• Harness Recession: When Small Models Outgrow Their Scaffolding
  M. S. Alim^*, Mahir Shahriar Tamim^*, S. Z. Ridoy, S. Khan, A. T. Wasi
  Studies when scaffolding around a small language model improves agentic task performance and when it holds back a model that has outgrown it.
  SLM-Agents: 1st NeurIPS Workshop on SLMs for Agentic Systems, 2026 [Paper] ▪ Agentic AI ▪ Small Language Models

Under Review and in Preparation

• Retrieval Geometry Shapes Cache-Based Clip Adaptation
  Mahir Shahriar Tamim, M. S. Alim, A. T. Wasi, S. Z. Ridoy, M. Nesa, M. A. Yousuf, A. Lamb (Tsinghua University), M. A. Moni
  Shows that the choice of retrieval encoder, not the memory bank itself, determines how much cache-based test-time adaptation helps CLIP, and proposes MARC, a training-free pairing of frozen CLIP predictions with DINOv2-B retrieval.
  ICLR 2027 (CORE A*) · Under review [arXiv] [Paper] ▪ Test-Time Adaptation ▪ Vision–Language
• SphereAD: Hyperspherical Self-Reference for Training-Free Anomaly Localization Across Diverse Domains
  Training-free anomaly localisation using hyperspherical self-reference, transferring across domains without per-domain fitting.
  In preparation ▪ Computer Vision
• When Bytes Take It All: CLIPByte for Tokenizer-Free Captioning via Byte-Level Vision–Language Bridging
  Bridges a CLIP visual encoder to a byte-level decoder, removing tokenizer-induced bottlenecks for low-resource scripts in image captioning.
  In preparation ▪ Vision–Language