Nirmal Thomas

MELLM Workshop @ ACL 2026 (Accepted, Non-Archival)

A post-hoc method (LCD) that attaches per-language rank-2 LoRA corrections to every linear layer of a quantized model—adding only 0.12% parameters per language and training in under 20 minutes on a single GPU. Across Qwen2.5-3B and Llama-3.2-3B, LCD recovers 70–83% of the non-Latin-script perplexity gap and 17–28% of the GlobalMMLU accuracy gap introduced by INT3 quantization.

Tyler A. Chang, Catherine Arnett, and Authors at the 5th Multilingual Representation Learning (MRL) Workshop

The 5th Workshop on Multi-lingual Representation Learning(MRL), EMNLP-2025

Contributed to the construction of dataset samples in Hindi