MSc Dissertation · 2026
Cross-Lingual Activation Steering in LLMs
Evaluation languages: English, Italian, Chinese, Korean and Kazakh
Why: Activation steering can modify LLM behaviour without retraining, but it is unclear whether English-derived interventions generalise reliably across languages and tasks.
How: Evaluated Contrastive Activation Addition (CAA) across three model families and five languages on HPC, using layer-wise interventions, held-out evaluations, bootstrap confidence intervals, and Gaussian and permutation random-vector baselines.
KEY FINDING
Representation similarity ≠ behavioural reliability.
High representational similarity did not guarantee comparable behavioural effects. A descriptive cross-task comparison suggested different patterns for Qwen and OLMo: Qwen transferred sycophancy but not truthfulness, while OLMo showed the reverse. Representational similarity was not sufficient for comparable behavioural transfer; whether it is necessary remains unresolved.
Activation SteeringSycophancyTruthfulQAQwenGemmaOLMo
Collaborative Research · Ongoing
Cross-Lingual & Cross-Task Generalisation of Alignment Interventions
Why: Similar steering representations do not consistently predict behavioural transfer, leaving open whether differences arise from steering-vector strength, direction, evaluation metrics, or target-language behaviour.
CURRENT WORK
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).
How / my role: Extended cross-lingual transfer evaluations and designed source–target comparisons with magnitude controls.
Source–target evaluationNorm-matched controlsCross-lingual transferRepresentation analysis
Research Project · Ongoing
Multilingual Safety Across OLMo 3 Alignment Stages
Why: Safety improvements measured in English may not generalise to other languages, and automated safety judges can introduce language-dependent measurement errors.
KEY FINDINGS
For matched English and Korean harmful requests, scorer-measured refusal rates were approximately 80% versus 10%, with a large gap persisting under human validation. Human annotation also revealed safety-judge errors in opposite directions across the two languages.
How: Built a reproducible framework covering OLMo 3 post-training checkpoints (SFT, DPO and RLVR), evaluated 200 matched prompts per language, and performed stratified human validation of 70 responses with bootstrap and Bayesian uncertainty analysis.
OLMo 3SFTDPORLVRMultilingual Safety