light_mode

Som Sagar — Curriculum Vitae

PhD Student in Computer Science, Arizona State University
ssagar6@asu.edu · Tempe, Arizona, United States

picture_as_pdf CV (PDF) library_books Publications school Google Scholar code GitHub business LinkedIn

Research interests

Reinforcement learning for post-training (RLHF, preference optimization, supervised fine-tuning); agentic and multi-agent LLM systems; failure diagnosis, red teaming, and safety evaluation of foundation models; interpretability and uncertainty quantification for vision-language models and robot manipulation policies.

Education

Arizona State University, Tempe, Arizona — PhD in Computer Science, Aug. 2023 – present.
Advisor: Ransalu Senanayake, LENS Lab. GPA 3.9/4.0.
Coursework: natural language processing, data mining, planning, learning methods in AI, statistical machine learning, knowledge representation, reinforcement learning.

Indian Institute of Information Technology (IIIT) Kottayam, Kerala, India — B.Tech (Honors) in Computer Science, Aug. 2019 – May 2023. CGPA 8.72/10.0.

Experience

Student Researcher, Google LLC, Sunnyvale, California — May 2026 – Aug. 2026.
Built a per-query agentic workflow router that represents multi-agent systems as a directed acyclic graph and compiles a custom topology via Monte Carlo Tree Search, lifting GAIA from 14% to 69% and SpreadsheetBench from 46% to 81% over zero-shot, and outperforming LATS, TRACE, GPTSwarm, and DyLAN. Adapted a frozen Gemma 2B backbone by replacing its vocabulary head with 10-action policy and scalar value heads (6.8M LoRA parameters), warm-started from pass-fail embedding arithmetic. Trained via policy-value distillation on MCTS visit-count targets under a token- and latency-aware composite reward, converging to compact four-node workflows rather than agent sprawl.

Machine Learning Research Intern, LinkedIn Corporation, Agents Platform team, Sunnyvale, California — May 2025 – Aug. 2025.
Designed a multi-head preference RL framework in which each head learns a distinct reasoning pattern and user preference signal rather than collapsing them into a single reward, lifting decision quality by 20%. Engineered structured reward pipelines driven by human-in-the-loop feedback, curating 1,200+ annotations to train and validate the reward model and raising agent alignment success rates by 40%. Integrated HydraLoRA adapters for parameter-efficient fine-tuning across task domains without retraining the base model, cutting compute costs by 25%.

Research Assistant, LENS Lab, Arizona State University — Aug. 2023 – present.
Research at the intersection of reinforcement learning, foundation models, and robotics, focused on failure detection, robustness, interpretability, and trustworthiness of deployed models.

Publications

* denotes equal contribution.

Conference papers

  1. Som Sagar, Jiafei Duan, Sreevishakh Vasudevan, Yifan Zhou, Heni Ben Amor, Dieter Fox, Ransalu Senanayake. RoboMD: Uncovering Robot Vulnerabilities through Semantic Potential Fields. International Conference on Learning Representations (ICLR), 2026.
  2. Joshua Tint, Som Sagar, Aditya Taparia, Caleb Liu, Kelly Raines, Bimsara Pathiraja, Ransalu Senanayake. ExpressivityArena: Can LLMs Express Information Implicitly? Findings of the Association for Computational Linguistics: EACL, 2026.
  3. Sreevishakh Vasudevan, Som Sagar, Ransalu Senanayake. Strategic Vantage Selection for Learning Viewpoint-Agnostic Manipulation Policies. IEEE International Conference on Robotics and Automation (ICRA), 2026.
  4. Atharva Gundawar*, Som Sagar*, Ransalu Senanayake. PAC Bench: Do Foundation Models Understand Prerequisites for Executing Manipulation Policies? Conference on Neural Information Processing Systems (NeurIPS), 2025.
  5. Som Sagar*, Aditya Taparia*, Harsh Mankodiya, Pranav Bidare, Yifan Zhou, Ransalu Senanayake. Trustworthy Explanations for Robot Behaviors. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025.
  6. Aditya Taparia, Som Sagar, Ransalu Senanayake. Explainable Concept Generation through Vision-Language Preference Learning for Understanding Neural Networks' Internal Representations. International Conference on Machine Learning (ICML), 2025.
  7. Som Sagar, Aditya Taparia, Ransalu Senanayake. Failures Are Fated, But Can Be Faded: Characterizing and Mitigating Unwanted Behaviors in Large-Scale Vision and Language Models. International Conference on Machine Learning (ICML), 2024 — Spotlight (top 3.5%).

Workshop papers

  1. Som Sagar, Jiafei Duan, Sreevishakh Vasudevan, Yifan Zhou, Heni Ben Amor, Dieter Fox, Ransalu Senanayake. From Mystery to Mastery: Failure Diagnosis for Improving Manipulation Policies. Robotics: Science and Systems (RSS) Workshop on Out-of-Distribution Generalization in Robotics, 2025.
  2. Som Sagar, Aditya Taparia, Ransalu Senanayake. LLM-Assisted Red Teaming of Diffusion Models through “Failures Are Fated, But Can Be Faded”. NeurIPS Workshop on Red Teaming GenAI: What Can We Learn from Adversaries?, 2024.

Other

  1. Som Sagar, Swani Sundara Didde, Cinu C. Kiliroor. Sentiment-Based Simplification of Legal Text. 4th International Conference on Communication, Devices and Computing (ICCDC), Lecture Notes in Electrical Engineering, vol. 1046, Springer, 2023.

Awards and honors

  • Spotlight (top 3.5%), International Conference on Machine Learning — 2024
  • SCAI Conference Award, School of Computing and Augmented Intelligence — 2025
  • Travel awards: ASU GPSA, ASU Graduate College, and NeurIPS Interpretable AI — 2024
  • Prime Minister Scholarship, Government of India — 2019 – 2023
  • Inter IIIT Hackathon Winner — 2022

Service

Reviewer: ICLR 2025 and 2026, CoRL 2025 and 2026, NeurIPS 2024 and 2026, IROS 2024 and 2025.

Teaching

  • Instructor, FSE 100: Introduction to Engineering, ASU — Fall 2023, 2024, 2025
  • Teaching Assistant, CSE 598: Operationalizing Deep Learning, ASU — Spring 2024
  • Teaching Assistant, CSE 100: Principles of Programming with C++, ASU — Spring 2024

Skills

Languages: Python, C, C++, Dart, JavaScript.
RL and post-training: PyTorch, Transformers, TRL, PEFT/LoRA, verl, Stable Baselines, Gymnasium, Diffusers, Captum.
Simulation: MuJoCo, Isaac Gym / Sim / Lab, RLBench, CARLA.
Also: TensorFlow, scikit-learn, NumPy, Pandas, Gradio.
Infrastructure: Google Cloud Platform, AWS, MySQL, Firebase, Git, Docker.