CV

Curriculum vitae and selected research experience.

Contact Information

Name Tianle Gu
Professional Title LLM Safety Researcher
Email gtl23@mails.tsinghua.edu.cn
Website https://github.com/Carol-gutianle

Professional Summary

Focused on LLM safety, alignment, and robust evaluation.

Education

  • 2023 - 2026

    M.Eng.
    Tsinghua University
    Electronic Information
  • 2019 - 2023

    B.Eng.
    Hunan University
    Computer Science and Technology
    • National Scholarship * 2

Publications

  •
    **Tianle Gu**, Zeyang Zhou, Kexin Huang, et al. MLLMGuard: A Multi-dimensional Safety Evaluation Suite for Multimodal Large Language Models
    Accepted by NeurIPS 2024

    We designed a multi-dimensional safety benchmark for MLLMs, including bilingual data, an inference toolkit, and a lightweight automatic evaluator.

  • **Tianle Gu, Zongqi Wang**, Kexin Huang, et al. Invisible Entropy: A Safe and Efficient Paradigm for Low-entropy Watermarking
    ArXiv; Accepted by EMNLP Main Conference (Oral)

    We introduced IE, a watermarking framework for low-entropy text without origin LLMs.

  • **Tianle Gu**, Kexin Huang, Zongqi Wang, et al. Probing the Robustness of Large Language Models Safety to Latent Perturbations
    ArXiv; Submitted to ACL 2026

    We proposed an adversarial attack-and-defense framework to systematically uncover and mitigate latent vulnerabilities in LLM safety.

  • **Tianle Gu**, Kexin Huang, Ruilin Luo, et al. From Evasion to Concealment: Stealthy Knowledge Unlearning for LLMs
    ArXiv; Accepted by ACL 2025 Findings

    We proposed a streamlined knowledge-unlearning algorithm that improves forgetting quality while preserving NLU/NLG utility and resilience to MIA.

  • **Tianle Gu**, Kexin Huang, Lingyu Li, et al. From Sparse Decisions to Dense Reasoning: A Multi-attribute Trajectory Paradigm for Multimodal Moderation
    Submitted to ICML 2026

    We proposed UniMod, a shift from sparse binary decisions to dense reasoning trajectories for fine-grained multimodal safety moderation.

Projects

  • ValuePRM - Core Contributor

    Technical Report

    • Led ValuePRM training with response-level data to evaluate and verify value-aligned behavior in MLLMs.
    • Proficient with OpenRLHF for training and fine-tuning both 7B and 70B models.
  • ChatZoo (80+ stars)
    • Developed ChatZoo, an open-source toolkit for local development and evaluation of multiple LLMs.
  • CoLLiE (400+ stars) icon CoLLiE (400+ stars)
    • Accepted by EMNLP Demo 2023.
    • Integrated efficient optimization and soft-prompt adaptation for better performance and resource utilization.
  • OpenRT (200+ stars)
    • Introduced a standard red-teaming framework for quick evaluation.

Internship

Notes

  • Note: Bold indicates first or co-first author.