Hi, I'm Mir Md Mohiuddin Abrar

LLM Quantization & Efficient Deployment

I am researching reduced-precision quantization, model compression, inference efficiency, and deployment trade-offs for large language models. The goal is to understand how LLMs can become more practical for memory-constrained and latency-sensitive environments while preserving useful model quality.

Recovery Coverage Audit — LLM Unlearning

Studied whether “forgotten” information in large language models is really gone. Built a controlled synthetic knowledge world plus a public benchmark check, and showed that how much knowledge appears to return depends heavily on the recovery test—the later training data, update budget, and comparison baseline. The study argues that unlearning results should be reported as conditional on the recovery procedure, rather than treated as direct evidence of deletion.

Evaluation study on whether machine unlearning in LLMs removes knowledge or merely hides it.

Evaluation Conventions in Perplexity and Model Compression

Investigated a hidden implementation choice in language-model perplexity evaluation and showed that, for one model family, it changes reported values by more than an order of magnitude and can mask real degradation in compressed models. Audited widely used evaluation codebases and published compression research to show the choice is almost never reported, then proposed a reporting standard and an inexpensive two-pass check.

Measurement-validity study showing that an unreported convention distorts language-model and compression results.