Research

I study what hardware faults do to large language models, and how inference can defend itself. Undergraduate researcher at the TRUE (Trustworthy and Responsible) AI Lab, Oregon State University.

• Cross-Scale Model Resilience for LLM Inference Under Hardware Faults

  2025

  Manuscript under review Co-author

  TRUE (Trustworthy and Responsible) AI Lab, Oregon State University

  Hardware faults (bit flips in GPU memory) can silently corrupt LLM inference. This work studies how that corruption behaves across model scales and proposes a speculative-decoding-based defense: a smaller draft model and the target model cross-check each other. The verification step doubles as a fault detector and corrector.

  My contribution: mitigation implementations and the accuracy / fault-tolerance trade-off analysis, with experiments run on an A100 Slurm HPC cluster.

• Defending LLM Inference Against Targeted Bit-Flip Attacks

  2026

  In preparation Initial investigation & results

  TRUE (Trustworthy and Responsible) AI Lab, Oregon State University

  Adversarial bit-flip attacks corrupt a handful of carefully chosen weights to degrade or hijack a model while its output still looks plausible. This direction asks when output-level monitoring can catch these attacks, and when it cannot.

  I ran the initial investigation and early results: I designed a draft-model-gated decoding defense and early-warning detection metrics, and benchmarked 7 fault-tolerance defenses against 2 attack classes (stealth and damage) on GSM8K. The work continues at the lab toward publication.

Code for both directions is lab-owned and private until publication.