Multilingual LLM Alignment & Evaluation | NLP Research | Applied AI

🎓 MSc Speech and Natural Language Processing · University of Sheffield (2025–2026; final award pending, predicted Distinction)

I investigate when we can trust LLM behaviour across languages and tasks—from the transfer of representation-level interventions to the reliability of multilingual safety evaluation. I also build reproducible NLP evaluation pipelines and evidence-grounded applied AI systems.

🌐 Research homepage · CV · LinkedIn · Email

Why: English-derived steering interventions may not have comparable behavioural effects in other languages, even when representations appear similar.

• MSc dissertation: Evaluated Contrastive Activation Addition (CAA) across three open-weight LLM families and five languages (English, Italian, Chinese, Korean and Kazakh), studying sycophancy, truthfulness and representation–behaviour transfer.
• Finding: High cross-lingual representational similarity did not reliably predict equivalent behavioural effects. A descriptive cross-task comparison showed differing patterns across Qwen and OLMo.
• How: Layer-wise evaluations, held-out sets, bootstrap confidence intervals, and Gaussian and permutation random-vector baselines on HPC.
• Follow-up: Built a full 5×5 source–target transfer matrix with norm-matched controls to distinguish source-vector effects from target-language responses. Joint manuscript in preparation (2027).

Research repository →

Why: Improvements in safety behaviour measured in English may not transfer to Korean, and automated safety judges may not be equally reliable across languages.

• Evaluation: 200 matched English/Korean prompts per language across OLMo 3 post-training checkpoints (SFT, DPO and RLVR).
• Finding: Scorer-measured harmful-request refusal rates were approximately 80% in English versus 10% in Korean, with a substantial gap persisting under human validation.
• Validation: Stratified human review of 70 responses, bootstrap and Bayesian uncertainty analysis; identified opposing patterns of safety-judge errors across languages.
• Engineering: Reproducible, configuration-driven evaluation components, typed schemas, deterministic validation and GitHub Actions CI.

Safety evaluation repository →

• Multilingual LLM safety, evaluation and cross-lingual transfer
• Activation steering and representation engineering
• Generalisation of alignment interventions across languages and tasks
• Reliability of behavioural evaluation and LLM-as-judge systems

Programming: Python · SQL

ML & NLP: PyTorch · Hugging Face Transformers · scikit-learn

LLM Alignment & Evaluation: CAA · Representation analysis · Cross-lingual and cross-task evaluation · Post-training safety evaluation · LLM-as-judge validation · Human annotation and error analysis

Experimental Methods: Controlled experiments · Bootstrap confidence intervals · Bayesian uncertainty analysis · Random-vector baselines

Retrieval & Applied AI: BM25 · Dense retrieval · Reciprocal Rank Fusion · Cross-encoder reranking · ChromaDB · Evidence-grounded LLM verification

Engineering & Research Infrastructure: HPC/SLURM · Linux · Weights & Biases · Git · GitHub Actions CI · Docker · FastAPI · Flask · Pydantic · pytest · Streamlit

A reproducible research framework for comparing refusal behaviour across English and Korean after SFT, DPO and RLVR. Combines matched-prompt evaluation with human validation to examine both model behaviour and safety-judge reliability.

LLM Evaluation · Multilingual Safety · Human Validation · PyTorch · GitHub Actions

MSc dissertation and follow-up evaluation of activation-steering generalisation across three model families and five languages. Investigates the distinction between representation similarity and reliable behavioural transfer.

Activation Steering · Representation Analysis · HPC/SLURM · Statistical Evaluation

Evidence-grounded verification of corporate sustainability reports, mapping requirements to retrieved passages and traceable judgements over 11 reports. In the evaluation reported in my current CV, hybrid retrieval with reranking improved Hit@5 from 83.3% to 100% and MRR from 0.628 to 0.819 compared with BM25-only retrieval.

Built with hybrid retrieval, reranking, a manually annotated gold set, FastAPI, Streamlit, Docker and automated tests. Retrieval and verifier errors were analysed separately.

RAG · Information Retrieval · LLM Verification · FastAPI · Docker

First-author LSTM forecasting research that reduced RMSE by approximately 71% relative to the baseline. Published at ACK 2023.

Time Series · LSTM · Python

• Kim, M., Kim, Y., Kim, E., Lee, B., Han, Y., & Lee, G. (2023). Implementation of Container Volume Prediction Technology using Deep Learning. ACK Conference Proceedings, 30(2). First author.
• Cross-lingual and cross-task generalisation of representation-level interventions in multilingual LLMs. Manuscript in preparation (2027).

Korean (Native) · English (Fluent) · Spanish (Intermediate) · Kazakh (Intermediate)