Shubham Gandhi

Language Technologies Institute, Carnegie Mellon University. srgandhi@andrew.cmu.edu

myphoto.jpeg

GHC 6414, Gates and Hillman Centers

5000 Forbes Ave

Pittsburgh, PA 15213, USA

I am a first-year Ph.D. student at the Language Technologies Institute of Carnegie Mellon University, where I am advised by Carolyn Rosé and work on efficiency and reliability in LLM agents. I completed my M.S. in Intelligent Information Systems at CMU in December 2025, and my B.E. in Computer Science at BITS Pilani, Goa.

I have spent two summers at IBM Research working on supervision and training for agents. Before CMU, I was a Research Fellow at TCS Research and a Research Intern at Microsoft Research.

My main interest is efficiency in LLM agents: getting agents to solve long-horizon tasks with less compute, fewer wasted steps, and smaller models. I also work on failure mitigation and reliability for coding agents:

  • Efficiency and failure mitigation. I design lightweight critics and strong-weak collaboration schemes that steer large code agents at a fraction of the cost, and I measure wasted effort such as code slop in agent-generated patches.
    Steer, Don’t Solve (EMNLP’26 Oral) · Strong-Weak Collaboration (EMNLP’25) · Code Slop (WAB @ COLM’26)

  • Supervision and training for long-horizon agents. I build process reward models that course-correct agent trajectories at inference time, rubric-based credit assignment for agent training, and I study which training environments help coding agents generalize.
    SWE-PRM (SEA @ NeurIPS’25) · DRACO (preprint) · Hybrid-Gym (ICML’26)

  • Reliability and security of code generation. I work on utility-preserving defenses for code LLMs, red-teaming of code agents, and detection and repair of security vulnerabilities, as part of the CMU team in the Amazon Nova AI Challenge (team member in Trusted AI ‘25, finalist; red team lead in Trusted Software Agents ‘26).
    Secure and Useful Models are Reasonable (Tech Report ‘25) · ABSENTIA (preprint) · Ulysses (preprint) · StaticFixer (preprint)

news

Oct 26, 2026 At EMNLP 2026, where Steer, Don’t Solve is presented as an oral.
Oct 22, 2026 Three new preprints on securing web applications with LLM agents: ABSENTIA, Self-Evolving Static Analysis, and Ulysses.
Sep 05, 2026 New preprint from my IBM Research internship: DRACO, fine-grained credit assignment with dynamic rubrics for long-horizon agent training.
Sep 04, 2026 Steer, Don’t Solve, on training small critic models for large code agents, is accepted to EMNLP 2026 as an oral presentation!
Sep 01, 2026 Evaluating Code Slop in Long-Horizon Coding Agents is accepted at the Workshop on Agent Behavior at COLM 2026.

selected publications

  1. Preprint
    DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training
    Shubham Gandhi*, Saurabh Goyal*, Kiran Kate, and 1 more author
    arXiv preprint arXiv:2609.04094 (under review), Sep 2026
  2. Evaluating Code Slop in Long-Horizon Coding Agents
    Shubham Gandhi*, Shyam Agarwal*, Nachiket Kotalwar*, and 1 more author
    In Workshop on Agent Behavior (WAB) at the Conference on Language Modeling (COLM), Oct 2026
  3. Steer, Don’t Solve: Training Small Critic Models for Large Code Agents
    Shubham Gandhi*, Yiqing Xie*, Atharva Naik, and 2 more authors
    In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), Nov 2026
  4. Hybrid-Gym: Training Coding Agents to Generalize Across Tasks
    Yiqing Xie, Emmy Liu, Gaokai Zhang, and 7 more authors
    In Proceedings of the 43rd International Conference on Machine Learning (ICML), Jul 2026
  5. When Agents go Astray: Course-Correcting SWE Agents with PRMs
    Shubham Gandhi, Jason Tsay, Jatin Ganhotra, and 2 more authors
    In Scaling Environments for Agents (SEA) Workshop at NeurIPS, Dec 2025
  6. An Empirical Study on Strong-Weak Model Collaboration for Repo-level Code Generation
    Shubham Gandhi, Atharva Naik, Yiqing Xie, and 1 more author
    In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), Nov 2025