Shaheen Nabi
Research @ Indian Institute of Science | LLM Post-Training, RLVR, Mechanistic Interpretability
Bengaluru, Karnataka, India
💬About Me
- I am a Researcher at the Indian Institute of Science (IISc), working on reinforcement learning, model distillation, and mechanistic interpretability for reasoning and language models.
- My research spans the full stack around reasoning models — post-training methods that build reasoning (RLVR, preference optimization, reasoning distillation), the test-time strategies that make it efficient (compute allocation, deciding when to stop deliberating), and the empirical/evaluation work needed to trust any of it. Right now I'm going deeper into mechanistic interpretability — studying the internal circuits and representations behind multi-step reasoning in large language models, to understand failure modes like reward hacking and unfaithful chain-of-thought well enough to actually fix them, not just detect them. A specific thread I'm pulling on: disentangling memorization from genuine reasoning mechanistically — which circuits are doing lookup versus computation — and using that understanding to work toward guiding reasoning safely, so models can be steered away from unfaithful or unsafe reasoning traces rather than just penalized after the fact.
- I build and release everything in the open — training/eval systems, checkpoints, and datasets — in lieu of formal publications, via GitHub and Substack.
📝Publications
- (in-coming)
- (in-coming)
💻Experiences
- 2026.06 - now: Researcher, Reinforcement Learning & Reasoning, Indian Institute of Science (IISc), Bengaluru.
- 2025.12 - now: Machine Learning Engineer, LLM Post-Training & RL, Freelance, Remote.
- 2024.10 - 2025.10: Artificial Intelligence Consultant, Freelance / Self-Employed, Jammu & Kashmir.
- 2025.01 - 2025.06: Data Science Intern, iNeuron.ai, Bengaluru.
- 2023.02 - 2023.12: Data Scientist, NIELIT, Jammu & Kashmir.
🛠️Selected Projects
- Open Post-Training A single reusable system that takes a base model through the full post-training recipe, stage by stage, rather than one-off scripts per experiment: evaluation, preference tuning, reasoning distillation, and RLVR training all share the same components. I use it as a research testbed — running hypotheses through the pipeline and producing demos/proof-of-output for specific ideas. One such run: a GRPO ablation on a base model (64.0% MATH-500) dropped to 48.0% at a 5-step checkpoint before recovering to 60.0% at 50 steps, with response length shrinking from 303 → ~220 tokens — evidence that sparse binary-correctness reward is a weak signal without more steps and better reward shaping.
- Olmo-3 From Scratch From-scratch implementation of AllenAI's Olmo-3 (7B/32B) — GQA, YaRN-scaled RoPE (8K→65K), sliding-window attention with periodic global refresh, SwiGLU, RMSNorm — plus a custom KV-cache module with unrotated key/value storage for RoPE compatibility, reducing cache memory via GQA while supporting bfloat16 inference.
- Reinforcement Learning: Zero to Hero A maintained RL repository spanning MDPs through PPO and DDPG, with clean implementations and math annotations for practitioners moving into post-training research.
- LeafLogic Dataset 25K annotated crop/plant images, open-sourced on Hugging Face — downloaded by thousands of developers around the world.
📖Education
- Bachelor of Computer Applications — Information Technology, Machine Learning, Deep Learning & AI Systems