01
Training dynamics
When mechanisms form: whether behavioural properties develop gradually or through qualitative transitions, and what that trajectory reveals about how models acquire, organise, and retain behavioural structure.
shreyans_jain · @py_parrot · independent researcher
mechanistic interpretability · training dynamics · behavioural decomposition
Shreyans Jain▍
i'm shreyans. i studied civil engineering in undergrad, and learned ML and programming afterwards alongside my job solely due to my interest in maths and problem solving.
my aim is to work full time on ML interpretability research because personally, i would like to understand more about whats going under the hood of a complex AI system responsible for deciding whether its a dog or a cat or what word i should be typing next. in addition, my interest lies in reinforcement learning, ai ethics as well so i try to keep myself up to date on these topics too.
two threads pull at me the most right now — training dynamics and behavioural decomposition. training dynamics because i think when a behaviour or capability forms during training often explains more than staring at the final model ever could. and behavioural decomposition because messy, named behaviours like sycophancy feel like they're built out of smaller, measurable traits — and if that's true, we can actually measure and steer them instead of just describing them.
§01·research
Two directions carry most of my current work.
02
Behavioural decomposition
Breaking complex behaviours — sycophancy, deception — into their constituent mechanisms, testing whether they recompose, and identifying the distinct modes a single named behaviour can take.
Decomposition, concretely
sycophancy = Σ atomic traits
- agreeableness
- deference
- praise-seeking
- answer-conformity
Additional research interests
- [Geometry-aware steering]
- [Manifold geometry]
- [Multilingual interpretability]
career trajectory vs anxiety levels
§02·now
I'm a Research Fellow at LASR Labs, mentored by Satvik Golecha, working with the UK AI Security Institute's Model Transparency team on model organisms and the training dynamics of emergent misalignment. Alongside, I'm exploring behaviour compositions and multilingual interpretability.
I'm open to full-time, part-time, or collaboration opportunities in interpretability research. Reach me at jshrey8@gmail.com.
§03·publications
Reverse-chronological. * denotes equal contribution.
- 2026
Gotta Catch Them All: The Modes of Sycophancy
- 2026
Measure What Matters: Psychometric Evaluation of AI with Situational Judgment Tests
- 2025
Beyond Linear Steering: Unified Multi-Attribute Control for Language Models
- 2025
How to Visualize Training Dynamics in Neural Networks
- 2025
Sycophancy as Compositions of Atomic Psychometric Traits
- 2025
Towards Discovering Linguistic Indicators for Misalignment in Language Models
§04·cv
Independent interpretability researcher working on how coherent, high-level behaviours in language models arise from distributed internal representations, and whether they decompose into identifiable mechanisms. My working view is that behaviours we name at the surface — sycophancy, deception — are compositions of smaller measurable traits, and that tracking when those components form during training reveals how models acquire behavioural structure. Before interpretability research, I spent 8 years building production machine learning systems.
Research experience — most recent first
- Jul 2026 – PresentResearch Fellow, London AI Safety Research (LASR) Labswith Satvik Golecha (UK AI Security Institute — Model Transparency Team)
- Apr 2026 – Jul 2026Research Intern, Thoughtworkswith Amirali Abdullah
- Dec 2025 – Mar 2026Research Engineer Fellow, Martianwith Narmeen Oozeer and Phillip Quirke
- Aug 2025 – Oct 2025Research Fellow, Thoughtworkswith Amirali Abdullah
- Jun 2024 – PresentIndependent Researcher, Independent Researchwith Naomi Saphra, Amirali Abdullah, and Shivam Raval
Prior industry experience
8 years building production ML systems before interpretability. Full detail in the CV.
- Oct 2024 – Aug 2025 Reinforcement Learning Engineer (part-time, contract), Physarum
- Jul 2021 – May 2024 Senior Data Scientist, GEP
- Nov 2019 – Jul 2021 Data Scientist II, BookMyShow
- Jan 2019 – Oct 2019 Data Scientist, Hotstar
- Mar 2016 – Jan 2019 Consultant, Fractal Analytics
Education
- 2011 – 2015 B.Tech in Civil Engineering, Malaviya National Institute of Technology, Jaipur
- Nov 2024 AI Safety Fundamentals (Alignment), BlueDot Impact
§05·open-source
Research tools I've built and released.
§06·writing
Research notes and the occasional essay. all posts →
- [Jan 2026] Paper Notes: Mechanistic?
My notes on the "Mechanistic?" paper — the origins of mechanistic interpretability as a term, field, and community.
- [Jan 2026] Lazy Ambitious, or Overthinking?
- [Jan 2025] Feature Geometry in Toy Models: Pentagon vs Hexagon
Why ReLU toy models cap at pentagon geometry, and what pushes them toward a hexagon.
- [Nov 2024] Effects of Non-Uniform Sparsity on Superposition in Toy Models
How non-uniform feature sparsity reshapes superposition in ReLU toy models.
§07·outside work
when i'm not working i'm probably thinking about how i can improve my strength & form in pull-ups and dreaming about muscle ups 🏋️, or learning about the Xs and Os of basketball. i occasionally pick up a book (mainly it's on my kindle :P) too, trying to learn more about indian history, world history — but won't shy away from a good fiction (sci-fi preferred) or some new topic that looks interesting to me.
Repeat reads
- > I Play to Play — Amit Varma
- > On Burnout, Mental Health, And Not Being Okay — Ludicity