Alexander F. Spies

Research Scientist at FAR.AI (PhD, Imperial College London) developing methods to detect and mitigate misaligned behaviour, notably deception, in frontier models at scale. Background in mechanistic interpretability and representation learning: reverse-engineered causal world models in maze-solving transformers using SAEs, activation patching, and causal interventions. Previously shipped production LLM services (10k+ weekly queries) at Epic Games; implemented scalable finetuning and evaluation infrastructure. Led collaborative research teams (9 researchers), published at NeurIPS/ICLR workshops, and open-sourced interpretability tooling. Deeply motivated by long-term AI safety.

Thesis: Interpretable Representations in Artificial Neural Networks

• Improved representations in Object-Centric Learning and reasoning
• Investigated failures and pathologies of multimodal foundation models
• Reverse-engineered world models in transformers using mechanistic interpretability techniques

• Thesis: Unsupervised World Models in the Animal-AI Environment
• Independent project: Neurosymbolic Learning & Neurally Weighted dSILP

• Completed graduate-level courses as an undergrad, alongside research

• Thesis: AI for the Automated Diagnosis of Atrial Fibrillation

Developing methods for detecting and mitigating misaligned behaviour, particularly deception, in frontier models at scale.

• Designing scalable evaluations to surface deceptive and otherwise misaligned behaviour in frontier LLMs.
• Building mitigation techniques that translate from controlled experiments to production-scale deployments.

Built infrastructure for production LLM systems at scale. (Promoted from intern to full-time in June 2025.)

• Built scalable finetuning and evaluation infrastructure (UnSloth, SageMaker, vLLM, W&B); enabling systematic ablations and reproducible experiments across model variants.
• Shipped production LLM services (10k+ weekly queries); designed robust data pipelines and evaluation harnesses to monitor model behavior across diverse inputs.
• Developed agentic pipelines for multi-turn code generation with retrieval and tool-use.

Led independent research groups on mechanistic interpretability of "model organisms" of planning.

• Defined agenda on mechanistic interpretability for maze-solving LMs; trained transformers and Sparse Autoencoders, managed 9 researchers across 2 projects. Yielded 2 workshop papers and a best-poster award.
• Built SAE-based circuit analysis pipelines using activation patching, causal interventions, and feature visualization to reverse-engineer transformer internals; found interpretable causal world models.
• Research tooling and datasets made public and published in Journal of Open Source Software.

Carried out research on mechanistic interpretability of transformer internals and multimodal reasoning.

Provided technical guidance on AI Safety Research to 8+ Research Fellows

NeurIPS, ICLR, ICML, AAAI, UAI, AIJ; MATS Research Proposals (2024)