CV/Resume: Hikaru Tsujimura, PhD
Last updated on 19 November, 2026

Profile
Research Interests
Building on my research in cognitive neuroscience (2006-2024)—where I studied how human values, decisions, and actions are internally represented—I now focus on AI safety (2024-present), particularly on understanding how the internal structure of AI systems leads to unintended values, reasoning and behaviors.
This challenge is likely to persist as:
- AI systems are evolving rapidly in capability, complexity, and diversity.
- Digital architectures are not inherently designed to align with human values and thinking processes.
- Forced alignment often induces workaround or deceptive behaviors (e.g. “scheming” AIs).
- We lack standardized methods to reliably interpret internal structures across models, tasks and research methods.
»»» Goals in 2025
Accordingly, my first goal is to develop principled approaches for addressing representational misalignment in AI systems.
In 2025, frontier models began demonstrating new roles in enhancing human creativity and productivity (e.g. New Chess Move Discoveries by DeepMind/Goodfire; Scientific Data Interpretations by Leap Lab). That same year, my study also uncovered hidden emotional and logical structures underlying assertive and persuasive model outputs, offering insights into effective human communication. Building on this, my second goal is to build tools that leverage frontier models to amplify human cognitive capacities.
»»» Goals in 2026 and Beyond
Kulveit et al.’s gradual disempowerment (2025) and the Hugging Face swarm incident in 2026 (OpenAI report, METR investigation) convinced me that we are entering an era of transformative AI, where AI-augmented work becomes the norm and outperforms human work alone. Rather than letting AI replace humans, we should design and protect cooperative societies of human and AI agents that proactively sustain meaningful human roles.
This also means that AI’s influence on people is growing at a global scale. Traditional media, social media, intellectual work, and many other high-stakes fields are increasingly shaped by AI’s voice. AI’s overconfident and excessively persuasive language, sometimes carrying non-factual claims, has therefore become one of the most urgent problems to solve. What we need instead is trustworthy AI systems that communicate transparently and neutrally, and that help expand human knowledge, intellectual abilities, and views. With such systems, people can act in ways that help society flourish, and they can monitor and guard against malicious AI and human actors.
Accordingly, my goals for 2026 and beyond are:
- To continue my first goal, using interpretability to detect unintended model behaviors from internal signals that would otherwise stay hidden.
- My 2026 study uses mechanistic interpretability to reveal a two-stage verification process in Qwen2.5-1.5B, showing how the model responds to an injected false answer in arithmetic problems. Steering each stage at its own position resists the injection much better than steering both stages at once.
- To map large populations of AI agents and human users together, observing their overall behaviors (as in the HF swarm incident) and how their beliefs and goals change, starting with monitoring AI’s influence on humans (e.g. how large populations of AI agents shape human views).
- To mitigate the risk of excessive AI persuasion of humans, I aim to build a population behavior map that monitors how a population behaves as a whole, traces where problems come from, and helps find protections against those causes.
Technical Skills
- Python: PyTorch, Transformers, Peft, Accelerate, TransformerLens, NNsight, Scikit-learn, Numpy, Pandas, Matplotlib, Wandb, and more
- LLMs: Llama2-3.2, Qwen2.5/3/3.5, DeepSeek-R1-Distill, Gemma2, GPT-OSS, DeepCogito, Magistral, OpenAI and Anthropic APIs
- AI Safety:
- Chain-of-Thought (CoT) Reasoning in LLMs
- Mechanistic Interpretability (SAE, TransformerLens)
- Overconfidence Analysis in LLMs
- LLM Persona Engineering
- AI Security & Red Teaming (Frontier models)
- Safe Reinforcement Learning (PPO)
- Model Evaluation (incl. LLM-as-a-Judge)
- Robustness to Misleading CoT / Injected Answers
- Interpretability Methods: Linear Probes, Logit Lens, J-Lens, Attribution / Logit-Difference Analysis, Activation Steering (Mean-Difference, Position-Specific Multi-Layer), Causal & Head Ablation, Residual-Stream Decomposition, Random-Direction Controls
- AI Research Infrastructure
- Quantized Fine-Tuning (QLoRA), Unsloth, Vast.AI, Google Colab, Remote Server
- LoRA Fine-Tuning (up to 11B), Multi-GPU Servers, AI Coding Tools (Claude Code)
- Programming: Bash, JavaScript, R, Git, SLURM/Job Scheduling, Matlab, SQL (PostgreSQL), Docker, FastAPI, Streamlit
- Analysis:
- Statistical: Regression, ANOVA, Bayesian methods, PCA, Factor Analysis, Mixed Models, Multilevel / Hierarchical GLMs
- Machine Learning: Supervised & Unsupervised Learning, Variational Inference, Neural ODE, RNN (LSTM, GRU), Knowledge-Graph Embedding
- Behavioral: Human-subject Experiments, Large-scale Online Studies, Survey Design (GDPR-compliant)
- Neuroimaging: Sleep-EEG, fMRI, Brain Activity Pattern Modeling
- Experimentations: Large-scale, Online, Human-subject/GDPR, Behavioral/Survey, Brain-imaging (Sleep-EEG/MRI)
- Platforms: Prolific, Qualtrics, PsychoPy/Pavlovia, Testable, SONA
- Writing: Journal papers, Preprints, GitHub Blog/Codebase
- Communication: Leadership to Masters/PhDs, Teaching, Conference presentations, Collaboration with scholars/industrial partner
- Research Team Lead (4 PhD students; IRG 2025 award)
Recent Experience
Visiting Researcher (Cardiff / UCL) and Independent Researcher, 2023-Present
⇨ Machine Learning (ML) / AI Safety Projects
# 1. Independent Mechanistic Interpretability Work II (September, 2026 - Present)
- Independently investigating internal mechanisms of LLM reactions to misleading CoT reasoning.
- Applied J-Lens, per-submodule residual-stream decomposition, causal ablation and random-direction control tests.
- Localized component-level mechanisms (MLP-18 for stated-claim tracking, attention L23/L24 for correctness tracking).
# 2. Founding Member @ Stealth AI Startup (March, 2026 - August, 2026)
- Developed automated sports coaching tools with Mechanistic Interpretability and Reinforcement Learning methods.
# 3. Independent Mechanistic Interpretability Work I (January, 2026 - Present)
- Independently investigated LLM vulnerabilities against misleading CoT reasoning (started as the MATS 10.0 application project, December 2025).
- Applied linear probes, logit lens and attribution analysis to uncover a two-stage verification architecture.
- Developed a position-specific multi-layer steering method that corrected 51% of failures (accuracy 52% → 76.5%) (Google Doc).
# 4. Independent Persona Stability Work (November, 2025 - December, 2025)
- Independently investigated how system prompt phrasing affects persona elicitation stability in Qwen3.
- Extracted activations under contrastive evil/non-evil prompts and used Claude API as an automated trait judge.
- Discovered that strongly trait-eliciting prompts produce stable, clustered activation patterns.
# 5. MATS 2025 Application Project (August, 2025)
- Conducted an 18/20 hour Mechanistic Interpretability project (two experiments).
- Employed PyTorch, nnsight, as well as providing project codebase.
- Probed hidden representational structures in Qwen3 to identify implicit mathematical reasoning knowledge (Google Doc).
# 6. Impact Research Groups (IRG) 2025 @ Arcadia (July, 2025 - October, 2025)
- Conducted research on Chain-of-Thought (CoT) reasoning in LLMs (DeepSeek, Qwen3, Gemma2, GPT-OSS).
- Replicated Emmons et al., (2025)’s work, showing CoT faithfulness under misleading prompts, when think-mode is off.
- Discovered that such faithful behaviors vanished with think-mode, raising a need for a new intervention (Google Slide).
# 7. Machine Learning Institute 2025 (May, 2025 - July, 2025)
- Implemented and deployed hackathon-style weekly ML/AI projects for 6 weeks.
- Utilized Streamlit, Docker, FastAPI, PostgreSQL, LLMs (e.g. OpenAI’s Whisper model) and PPO algorithms.
-
| Selected Projects: Digit recognizer (Demo) |
Own GPT model from scratch (GitHub Repo) |
# 8. Supervised Program for Alignment Research (SPAR) 2025 @ Kairos (February, 2025 - May, 2025)
- Mentored by Kellin Pelrine and Puelma Touzel (Mila) on building more trustworthy, reliable AI systems.
- Fine-tuned Llama 3.2 models (1B/3B/11B) to detect overconfident behaviors in outputs. (Group/Preprint)
- Visualized internal confidence signals with TransformerLens to spot model overconfidence. (Solo/Preprint)
# 9. AI Safety Fundamentals - Alignment Course 2024 @ BlueDot (October, 2024 - February, 2025)
- Gained core knowledge on AI safety, such as Scalable Oversight, Mechanistic Interpretability, and AI Controls.
- Led an individual project developing a novel interpretability method with SAE and generative models (GitHub).
# 10. Group Project 2024 @ Neuromatch Academy Workshop in NeuroAI course (July, 2024)
- Led a team of 2 PhD students to complete a 2-week NeuroAI project.
- Used RNN-based models (Neural ODE, LSTM, GRU) for motor tasks (3-Bit Flip Flop and Random Target task).
- Achieved strong alignment between model outputs and muscle activity data across varied tasks (GitHub).
# 11. Collaborative Project 2024-2025 @ University College London (May, 2024 - Present)
- Visiting postdoc in Prof. Bradley Love’s lab, extending the previous independent work (#12).
- Utilizing Tensorflow, Variational Inference, ML models (e.g. Deep Support Vector Machine).
- Training ML models to predict and generate human behaviors and brain activity patterns (fMRI).
# 12. Independent Projects (August, 2023 - August, 2024)
- Optimized model alignment with individual participants using the data from #13.
- Utilized pre-trained networks, efficient quantized fine-tuning (QLoRA), LLMs (e.g. Llama 2).
- Discovered participant-specific internal representations (GitHub).
# 13. Group Project 2023 @ Neuromatch Academy Workshop in Deep Learning course (July, 2023)
- Led a team of 7 PhD students and researchers to complete a 3-week ML project.
- Utilized PyTorch, pre-trained networks (e.g. CLIP) and 4.6M human choice data from 14K participants.
- Aligned models with the above data to visualize internal representations of everyday objects in humans (GitHub).
Post-Doctoral Researcher, 2019-2023 - Cardiff University
⇨ Cognitive Science Projects
# 14. Decision Making Projects @ Cognition and Computational Brain Lab (2022-2023)
- Conducted 28 online behavioral studies (30-45 min long) with 700 subjects.
- Reduced online study duration by 99% using Python/R for agile implementation, analysis and visualization.
- Uncovered impacts of initial practices for later choice preferences, enhancing user behavior prediction. (Preprint)
# 15. Memory & Curiosity Projects @ Motivation and Memory Lab (2019-2022)
- Collaborated with BBC on online studies (behavioral/survey) with 500 subjects.
- Utilized Python, JavaScript, SQL, Regression, PCA, Factor analysis, K-means Clustering, Linear Mixed Model.
- Identified three motivational factors from survey data that predicted BBC users’ news consumption behavior. (Preprint)
Prior Experience
PhD Student, 2013-2019 - University of Manchester
# 16. Sleep & Memory Projects @ Sleep Lab (2013-2019)
- Supervised by Prof. Penny Lewis
- Investigated impacts of sleep on memory and brain activity changes over time (Paper).
- Gained hands-on experience in sleep (EEG) and brain-imaging (fMRI) data analysis, teaching and supervision.
- Learned Python/R and analysis of A/B tests, t-tests, ANOVA, Bayesian statistics, linear/non-linear model fitting.
Voluntary Research Assistant, 2011-2013 - Goldsmiths, University of London
# 17. Face Processing Projects @ Banissy Lab (2011-2013)
- Supervised by Prof. Michael Banissy
- Investigated on various characteristics of face processing (e.g. perceiving physical fitness from static facial images)
- (Paper1, Paper2)
Education
- PhD in Psychology, University of Manchester, Manchester, UK, 2013-2018
- MSc in Cognitive Neuroscience, University College London, London, UK, 2010-2011
- BA in Psychology, Saint Louis University, St. Louis, MO, USA, 2006-2010