About Me
I am a third-year undergraduate at the Indian Institute of Technology, Kanpur, pursuing a B.S. in Mathematics and Scientific Computing. I am currently pursuing research with Prof. Sayak Ray Chowdhury from Department of Computer Science and Engineering at IITK. My primary research interests include:
- Alignment in Large Language Models: I have worked on Test-Time Alignment and Optimization in LLMs, I am also working on studying underlying geometry of hidden states in LLMs and its relation to factuality, hallucination, safety and robustness.
- Reinforcement Learning & Robotics: I just love to play with robots. There is no grand philosophical motivation here... I just enjoy treating RL as highly complex toys to tinker with.
Alongside these I have also worked with Computational Neuroscience, where generally I am interested in SNNs and their biological plausibility.
Ongoing Work
Some of my current ongoing work as listed below. Feel free to contact me for any discussion or collaboration.
- AIMO Interpretability Challenge (NeurIPS 2026)
- Early stopping of chain of thought reasoning in LLMs
- Dynamic Latent Steering in LLMs for Safety Alignment
- Context Rot and Overcontextualization in long context LLM conversations
Research Experience
Formulated Anytime-Valid sequential tests to improve the efficiency of test-time alignment within the BoN framework, providing provable error guarantees for early stopping criteria. Paper submitted to ICLR 2027.
Exploited Best-of-N (BoN) strategies for test-time alignment in LLMs as an alternative to standard RLHF pipelines. Invented Reward Weighted Self Consistency improving performance over standard BoN by up to 10% and SPRT-RWSC for early stopping of candidates generation reducing budgets by up to 65%.
Integrated missing NERO robot support in Robosuite for data collection to train baseline imitation learning policy. Implementing the Beyond Action Residuals framework in simulation using Robosuite for AgileX NERO arm. Handling the Sim-to-Real transition from trained simulation-based policy to the real NERO hardware.
Management
Leading a team of 30 secretaries for participation in competitions and handling events. Leading 1 competition (AIMO-Interpretability Challenge, NeurIPS 2026) with a total of 10 team members and managing a research project with 8 secretaries on the team.
Industrial Experience
Preprints
For mitigating catastrophic forgetting in SNNs, we introduce a Potentiation-Factor (P-Factor) to identify neural patterns responsible for predictions, enabling selective memory preservation during new task learning.
Technical Blogs and Articles
Key Projects
- Performed causal analysis from data of 20,000 call-center transcripts on Business Events retrieved and analysed from base data based on natural language queries.
- Designed a causal-graph-driven, multi-agentic RAG framework for proper causal analysis as requested by the prompt, returning, analysis, along with statistics and example conversation citations
- Developed multi-agent query generation and evaluation pipeline with human-feedback-based filtering.
- Integrated MCP powered SQL database for statistical backing of the claims and further automated corrections and further analysis if there is significant difference between analysis and statistics.
- Implemented seminal research papers related to key Deep-RL algorithms (DQN, Double DQN, Prioritized Experience Replay, N-Step Learning) to train an agent for Atari Breakout.
- Achieved a mean score of 169 in final 50 training episodes and a testing score of 290, surpassing human mean score of 30.
- Implemented a research paper based on our goal and demonstrated a game-play using Spiking Neural Networks using weight transfer from a simpler DQN, which exceeded the original ANN performance in the test episodes.
- ANN Performance: Mean Score=0.63, Max Score= 4 | SNN Performance: Mean Score=1.13, Max Score= 11
Designing an agentic system for complete task automation, and complete system control for my system featuring deep personalization and voice-based controls.
Mentored Projects
- Mentored a group of 30 students to demonstrate continual learning in Reinforcement Learning via shallow world models on ContinualBench.
- Guided the implementation of Proximal Policy Optimization (PPO) for robotic arm manipulation across various sequential tasks.
- Investigated catastrophic forgetting in PPO by training agents sequentially across tasks and analyzing other continual learning mitigation methods.
- Led the integration and performance comparison of Model Predictive Control (MPC) based shallow world models.
- Mentored a group of students to train a multi-agent self-improving Reinforcement Learning algorithm for multiplayer games like Pong.
- Guided the implementation of iterative self-play using Deep Q-Networks (DQN) where the agent learns by competing against a clone of itself to create an auto-curriculum.
- Demonstrated robust learning without human data ingestion, showcasing successful convergence compared to training solely against a static, ideal opponent.
Achievements
Get in Touch
I'm currently open to new opportunities, research collaborations, and interesting conversations. Whether you have a question or just want to say hi, feel free to drop a message!