About

I am a first-year Ph.D. student in Computer Science at UC Santa Cruz, advised by Dr. Ian Lane. My research is on LLM post-training, factuality, and the safety-focused evaluation of language models and agents — where capable models quietly fail, and how to measure it well enough to fix it.

That work currently spans agent benchmarks, reward hacking in agent evaluation infrastructure, and how supportive language models shape a user's reliance on them. Alongside that I have worked at Samsung Research America and Nokia Bell Labs, after six years building production ML at Adani Group and BlackRock at 100M+ user scale.

LLM Agents Alignment & Safety Human-Centered AI

News

  • Oct 2026 RELATE is under review at ICLR 2027 (co-first author) — arXiv
  • Sep 2026 Started my Ph.D. at UC Santa Cruz with Dr. Ian Lane, supported by the UCSC Regents Award and the International Doctoral Recruitment Fellowship.
  • Sep 2026 BenchShield preprint is out, on reward integrity in LLM-agent evaluation infrastructure — arXiv, Twitter
  • Aug 2026 Signal-Driven Observation accepted as a poster at the BayLearn Symposium 2026 (first author).
  • Jul 2026 Joined BenchFlow as an open-source contributor.
  • Jun 2026 Signal-Driven Observation for Long-Horizon Web Agents accepted to the ICML 2026 FAGEN workshop — arXiv
  • Jun 2026 SkillOptimizer and FAULTLOC accepted to the ICML 2026 AIWILD workshop on safety, security, and beyond.
  • 2026 1st place, Cybersecurity Agent track, UC Berkeley AgentBeats Competition.
  • 2026 Winner, NVIDIA GTC PyTorch Helion Hackathon (LLM-guided Triton kernel optimization).
  • 2026 Selected for the BlueDot Impact AI Safety Fellowship.
  • 2025 Received a Thinking Machines Lab Tinker Research Grant with Dr. Chenguang Wang.

Selected Publications

* denotes equal contribution. See all publications →

RELATE pipeline: source situations, interactive dialogue environment, and sentence-level evaluation
RELATE: An Evaluation Framework for Measuring Relational Orientation of Large Language Models Shivam Shukla*, Jihye Kim*, Shubham Gaur*, Mahnaz Roshanaei, Magy Seif El-Nasr
Under review at ICLR 2027
Human-Centered AI Alignment & Safety

Across 1,596 dialogues and seven LLMs, supportive models drift inward: the share of sentences positioning the AI as the user's ongoing source of support is higher at the sixth assistant turn than the first, while language encouraging real-world human connection is substantially lower for hesitant, indirect users than for explicit ones.

arXiv
FAULTLOC: Evaluating AI Coding Agents for Fault Localization from Crash to Cause J. Tu, Shubham Gaur, R. Murtinty, Z. Wang, T. Shi, D. Song, C. Wang
ICML 2026 AIWILD Workshop: Safety, Security, and Beyond
Agent Evaluation

A 900-CVE benchmark tracing a crash back to its root cause: frontier coding agents recover the true root cause in only 40% of cases.

OpenReview Code
Signal-Driven Observation architecture: root LM, browser, signal detector and sub-RLM
Signal-Driven Observation for Long-Horizon Web Agents Shubham Gaur, Ian Lane
ICML 2026 Workshop on Failure Modes in Agentic AI (FAGEN), position paper · First author
LLM Agents

Long-horizon web agents re-ingest tens of thousands of DOM tokens at every action step, eroding reasoning before the task completes. Decoupling observation from action — re-reading only when a URL change, a newly visible element, or an action failure fires — makes observation compression an architectural decision rather than a prompting one.

arXiv

Honors & Awards

Service

  • Peer Reviewer · ICLR 2026, ACL 2025.
  • Open source · MassGen (multi-agent LLM framework) and BenchFlow evaluation infrastructure.