Misun Kim

Dissertation: Cross-Lingual Activation Steering in Large Language Models: Representational Alignment and Behavioural Transfer

Kazakh language and regional studies · First Distinction Scholarship (2023).

• Investigated the robustness of activation-level interventions in multilingual LLMs using Contrastive Activation Addition (CAA), testing whether English-derived steering vectors transfer consistently across three model families and five languages.
• Identified a representation–behaviour gap: high steering-vector similarity did not reliably predict the magnitude or direction of behavioural transfer across models and tasks, concluding that high representational similarity is not sufficient to guarantee comparable behavioural effects, and leaving the question of necessity unresolved.
• Designed layer-wise evaluations with bootstrap confidence intervals, held-out sets, and Gaussian and permutation random-vector baselines to characterise failure modes and distinguish systematic effects from noise.
• Conducted a separate cross-task evaluation using English-derived truthfulness vectors on TruthfulQA, finding a descriptive cross-task contrast across models and behaviours: Qwen3.5-9B transferred sycophancy but not truthfulness, OLMo3-7B-Instruct the reverse.
• Follow-up study (post-MSc): built a full 5×5 source–target transfer matrix with norm-matched controls to distinguish source-vector effects from target-language response differences.

• First-authored LSTM study on container freight forecasting across five universities, reducing RMSE by 71% over the baseline (published ACK 2023).

• Built a reproducible framework to investigate cross-lingual transfer of safety refusal behaviour across OLMo 3 post-training checkpoints.
• Evaluated 200 matched EN/KO prompts per language, finding ~80% vs. ~10% scorer-measured harmful-request refusal rates, with a large gap persisting under human validation.
• Conducted stratified human validation (70 responses) with bootstrap and Bayesian uncertainty analysis, revealing opposing patterns of safety-judge errors across languages.

• Built an end-to-end compliance verification system over 11 corporate ESG reports, mapping regulatory criteria to retrieved evidence and producing traceable covered/partial/missing judgements using hybrid retrieval and LLM-assisted verification.
• Developed and audited a manually annotated retrieval gold set; hybrid retrieval with reranking improved Hit@5 from 83.3% to 100% and MRR from 0.628 to 0.819 over BM25-only retrieval.
• Used retrieval and verifier error analysis to identify annotation gaps and evidence-selection failures, then modularised the system with FastAPI, automated testing, Docker, and a separate Streamlit interface.

• Contributed to 170+ page sustainability reports for major public and financial organisations, conducting TNFD climate scenario analysis and developing a CSDDD-aligned supply-chain risk assessment framework.

• Built workflow automation using the Slack API, Google Sheets, and Looker Studio to analyse 10,000+ service records, reducing unnecessary field visits by 17%.

• Provided individualized Korean/English instruction to 5+ learners aged 6–18, from early literacy to CSAT-level advanced reading.