Publications
Also on Google Scholar.
Papers
A Removal Based Approach to Improve LLM Faithfulness at Test-Time
LLM explanations can leave out what actually influenced the answer. We remove from the input the concepts the explanation doesn’t credit and ask the model again, so the new answer can’t depend on them. It needs no retraining or access to the model’s weights. Across two datasets and two model families, it made explanations more faithful.
Graphlet- and motif-based link prediction in large networks
BLANT-Predict ranks likely missing links using only a network’s structure, from the small subgraphs around each pair of nodes. On 12 real-world networks with up to about a million nodes, it was more precise than 13 other methods, including when checked against future versions of a network. I helped design the study and ran experiments.
A Graph Theory of Asset Cyber Risk Management in ICS Networks
We model an industrial control system’s network as a graph of cyber risk and use centrality to find its single points of failure. We prove these rankings can flip under small changes in the risk estimates, though they held up well in simulations on realistic networks.
A Cyber Operations Management Theory for Software Supply Chains
We model software dependencies, from packages to pretrained models and datasets, as a graph of cyber risk, prove that choosing which ones to harden on a fixed budget is NP-hard, and give a randomized approximation algorithm.
Sampling methods for graphlet estimation
Counting small subgraphs (graphlets) exactly is infeasible on large networks, and the usual Markov chain Monte Carlo (MCMC) sampler draws dependent samples and is hard to parallelize. I redesigned two samplers, Node-Based Expansion (NBE) and Edge-Based Expansion (EBE), that draw independent samples and run up to 70 times faster on 100 threads than on one. I also cut MCMC’s slowest setup step from hours to seconds. The samplers are in BLANT.
* Equal contribution. † Equal senior authorship.
Writing in Bengali
Written for Prothom Alo, a Bangladeshi daily newspaper, and Somantoral, a student-run math magazine. My name in Bengali is এস এম এ নাহিয়ান.
- ফ্যালাসি সমাচার An article on logical fallacies · Somantoral, Issue 2, December 2024 (PDF)
- ব্রোঞ্জ জয়ের অভিজ্ঞতা My account of winning bronze at IMO 2024, printed alongside a teammate’s · Prothom Alo, August 2024
- আন্তর্জাতিক গণিত অলিম্পিয়াড – ২০২৩ (সমস্যা ৬-এর সমাধান) Solution to IMO 2023 Problem 6 · Prothom Alo, August 2023
- এশিয়া প্যাসিফিক গণিত অলিম্পিয়াড-২০২২ (সমস্যা ৩–এর সমাধান) Solution to APMO 2022 Problem 3 · Prothom Alo, March 2023
- আন্তর্জাতিক গণিত অলিম্পিয়াড-২০২২ (সমস্যা ১ এর সমাধান) Solution to IMO 2022 Problem 1 · Prothom Alo, August 2022
Olympiad notes in Bengali
Notes I wrote for Gonitzoggo’s olympiad courses.
- Trigs aren’t Bad Trigonometry for olympiad geometry · February 2023 (PDF)
- No equality Olympiad inequalities · January 2023 (PDF)
Talks in Bengali
- Beyond the Matrix, Episode 5 A podcast interview about olympiads and applying to MIT · Notre Dame Math Club, Dhaka, January 2026
- Learn from an MIT student A session on competitive programming, olympiads, and applying to universities abroad · Dhaka College IT Club, March 2025