Jalend Bantupalli
Research Scientist – LLMs, Program Synthesis, Embodied Agents
jbantupalli [at] ucsd [dot] edu

I'm a Member of Technical Staff at MyoLab, working on LLM-based reasoning systems. I completed my Master’s in Computer Science at UC San Diego, where I focused on LLM Reasoning and Program Synthesis. Previously, I worked at Google and Microsoft on applied AI systems.

At Google, I worked on Gemini Networking, and at Microsoft, I contributed to AI-powered Dynamics 365. I earned my B.Tech from IIT Kharagpur, where I was advised by Professor Animesh Mukherjee on fairness in NLP, culminating in our paper on decoding demographic bias.

Email  /  CV  /  LinkedIn  /  Scholar  /  Twitter  /  Github

Research

My research focuses on Large Language Model (LLM) reasoning, program synthesis, and multimodal understanding. I’m particularly interested in enabling LLMs to perform structured, step-by-step reasoning in complex environments where inputs span across natural language, formal representations (e.g., code, logic), and physical modalities such as motion or visual context. This includes developing models that can solve problems by abstracting patterns, composing subroutines, and leveraging prior examples — rather than relying on brute-force memorization. I aim to design systems that are not only general-purpose and data-efficient, but also interpretable and aligned with human-like reasoning capabilities.

📝 I recently co-authored ImplexConv, a large-scale dataset and retrieval framework for implicit reasoning in personalized multi-session conversations. Our proposed method, TaciTree, introduces hierarchical multi-level summarization to support efficient long-context reasoning. (Preprint on arXiv)  · See more publications ↓

At MyoLab, I work on developing LLM-based agents that interact with embodied data (e.g., body movements, sensor inputs) to solve complex reasoning tasks. My broader goal is to create learning frameworks that combine LLMs with reinforcement learning and vision to enable grounded, goal-directed intelligence.

Publications

• Toward Multi-Session Personalized Conversation: ImplexConv
  Xintong Li, Jalend Bantupalli, Ria Dharmani, Yuwei Zhang, Jingbo Shang
  EMNLP Main Conference, 2025
  arXiv
  Introduced 2500-example long-term dialogue dataset and TaciTree framework for hierarchical summarization.


• Flow of Reasoning: Training LLMs for Divergent Problem Solving with Minimal Examples
  Contribution: Dataset, Training, Evaluation
  ICML, 2025
  PDF / Project page
  Achieved SOTA 50.37% on 1D ARC; proposed curriculum-based training for LLM generalization.


• Decoding Demographic Unfairness from Indian Names
  Vahini Medidoddi, Jalend Bantupalli, Souvic Chakraborty, Animesh Mukherjee
  SocInfo, 2022
  arXiv / Code
  Created large-scale Indian name datasets; trained transformers for gender and caste inference.

Media & Mentions

• Motion-Text Demo shared publicly on LinkedIn. Engaged the AI/ML community with discussions around LLMs for embodied agents.