Projects

Selected code releases from my research. The full list is on GitHub.

Latent-Space Intervention for Cross-Lingual Factual Consistency

Activation-level interventions in a multilingual model's latent space that make its factual answers more consistent across languages, with no drop in accuracy.

Code · EMNLP 2026 (Findings)

LLM Sensitivity to Long Inputs

Experiments measuring how the safety behavior of instruction-tuned LLMs changes as inputs grow longer.

Code · EMNLP 2026 (Main)

Label-Specific Soft Prompt Tuning

Parameter-efficient transfer learning with per-label soft prompts and adapters that match full fine-tuning while training under 1% of a model's parameters.

Code · NAACL 2025

Cross-Lingual Nearest-Neighbor Retrieval for Data-Efficient Classification

A retrieval-based cross-lingual pipeline that reaches supervised-level performance with roughly 50 labeled examples per language.

Code · EMNLP 2025