iiinova — where your digital self keeps evolving

Sekh (Sk) Mainul Islam

@sekh-copenlu.bsky.social

152 followers 653 following 34 posts

PhD Fellow at the CopeNLU Group, University of Copenhagen; working on explainable automatic fact-checking . Prev: NYU Abu Dhabi, IIT Kharagpur. mainuliitkgp.github.io

Posts Replies Media

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 16/10/2026

Do Mixture-of-Experts LLMs find unique routing paths for different user preferences? Our new preprint, “Explaining the Roles of Experts in Mixture-of-Experts LLMs for Personalization,” shows they do.

100 1

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 06/11/2025

💡How is the CoT mechanism aligned with the knowledge interaction subspace? 📊 CoT maintains similar CK alignment compared to standard prompting for all the datasets, and also reduces PK alignment.

110 1

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 06/11/2025

💡 Can we find reasons for hallucinations based on PK-CK interactions? 📊 The gap between PK and CK is much higher for the examples with hallucinated spans than for the examples with no hallucinated spans across the sequence steps.

110 1

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 06/11/2025

💡 How do individual PK and CK contributions change over the NLE generation steps for different knowledge interactions? 📊 During most of the NLE generations, the model slightly prioritizes PK.

110 1

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 06/11/2025

🪛 We propose a novel rank-2 projection subspace that disentangles PK and CK contributions more accurately and use it for the first multi-step analysis of knowledge interactions across longer NLE sequences.

110 1

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 06/11/2025

💡 Is a rank-1 projection subspace enough for disentangling PK and CK contributions in all types of knowledge interaction scenarios? 📊 Different knowledge interactions are poorly captured by the rank-1 projection subspace in LLM model parameter

110 1

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 15/08/2025

📊 Key Takeaways: 3️⃣ Real & Fictional Bias Mitigation: Reduces both real-world stereotypes (e.g., “Italians are reckless drivers”) and fictional associations (e.g., “citizens of a fictional country have blue skin”), making it useful for both safety and interpretability research.

120 2

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 15/08/2025

📊 Key Takeaways: 2️⃣ Strong Generalization: Works on unseen biases during token-based fine-tuning.

110 1

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 15/08/2025

📊 Key Takeaways: 1️⃣ Consistent Bias Elicitation: BiasGym reliably surfaces biases for mechanistic analysis, enabling targeted debiasing without hurting downstream performance.

110 1

Sekh (Sk) Mainul Islam @sekh-copenlu.bsky.social · 15/08/2025

🚀 Excited to share our new preprint: BiasGym: Fantastic LLM Biases and How to Find (and Remove) Them 📄 Read the paper: arxiv.org/abs/2508.08855

192 9