Shaheen Nabi
Hi! 👋 I'm Shaheen, a Research Engineer based in India, working on reasoning and thinking models — understanding how large language models perform multi-step reasoning and how training and post-training methods can improve their reliability, efficiency, and scalability.
I study reasoning across the full model development pipeline: from pre-training data composition and architectural choices to the post-training systems that shape model behavior. My primary focus is the post-training stack — supervised fine-tuning (SFT), preference optimization, RL-based training (including RLVR and related approaches), and test-time compute strategies such as agentic scaffolding that enable models to reason more effectively without requiring larger models.
Alongside training methods, I'm interested in the interpretability of reasoning models — studying the internal mechanisms and representations that support multi-step reasoning, and using mechanistic analysis to diagnose failures such as shortcut reasoning, reward hacking, or unfaithful chain-of-thought.
My goal is to build and study systems that make reasoning models more reliable, more efficient, and easier to scale — combining post-training, inference-time reasoning strategies, rigorous evaluation and benchmarking, and mechanistic analysis to better understand how reasoning emerges and how it can be improved in practice.
I contribute to open source through code and writing, and am actively working toward contributions to large-scale LLM infrastructure and post-training frameworks, with a focus on end-to-end implementations of reasoning-focused training pipelines.
Research Interests
- Scaling Foundation Models Data composition, quality, and curation at scale. How pre-training decisions — compute allocation, data mixtures, and training dynamics — shape the capabilities available to post-training.
- Continual & Mid-Training How continued pre-training and mid-training interventions affect the reasoning qualities that emerge in later stages. The relationship between training phase, data distribution, and downstream reasoning behavior.
- Architectures for Reasoning How architectural choices — hybrid MoE, state space models, and dense-sparse combinations — affect reasoning capabilities. Studying new open model families to understand what architectural decisions make models more or less amenable to post-training reasoning improvements.
- Post-Training Pipelines RLVR, SFT, and preference optimization. How reward signals, data quality, and training dynamics in post-training interact with the capabilities laid down during pre-training.
- Test-Time Compute & Efficient Reasoning Making models think better at lower cost. Process reward models, adaptive compute budgeting, and search-guided generation. The core question: how do we build models that know when to stop thinking, not just how to think longer.
- Interpretability of Reasoning Studying the internal mechanisms and representations that support multi-step reasoning. Using mechanistic analysis to diagnose failure modes — shortcut reasoning, reward hacking, and unfaithful chain-of-thought — and understand how reasoning emerges in large models.
- Emerging Reasoning Architectures Actively following new directions: small reasoning models, hierarchical reasoning architectures, and novel thinking paradigms. Interested in how these approaches can inform the next generation of frontier reasoning models and open new possibilities on hard benchmarks.
Outputs
Open-source repositories, writing, and models — in lieu of publications.