Changhao He / 何长浩

I am Changhao He (何长浩), a Ph.D. student in the College of Computer Science at Sichuan University, where I am fortunate to be advised by Prof. Peng Hu (胡鹏) and Prof. Xi Peng (彭玺).

Prior to my Ph.D. study, I pursued my Master's degree (2023-2025) in the same college under their supervision. Earlier, I obtained my Bachelor's degree (2019-2023) from the College of Mechanical Engineering at Sichuan University.

Email  /  Google Scholar  /  Github  /  WeChat WeChat QR Code

profile photo

Research

My research focuses on Multi-modal Learning and Noisy Correspondence. Our group maintains the Awesome Noisy Correspondence repository GitHub stars, a resource hub for research in noisy correspondence. If you have any suggestions or ideas, we warmly invite you to get in touch with us.

News

[2026-04] 🎉🎉 Two papers were accepted by ICML 2026! One was accepted as Spotlight (accept rate=2.2%)! Thanks to all coauthors.

[2026-04] Gave a talk @ SouthWest Petroleum University. Slides available.

[2026-02] 🎉🎉 One paper was accepted by CVPR 2026! Thanks to all coauthors.

[2026-01] 🥳 First-time reviewer for ICML 2026, ECCV 2026.

[2025-10] 🥳 First-time reviewer for CVPR 2026. Looking forward to serving the community!

[2025-04] Gave a talk @ Chongqing University of Posts and Telecommunications. Slides available.

[2025-02] 🎉🎉 One paper was accepted by CVPR 2025! Thanks to all coauthors.

[2024-07] 🎉🎉 One paper was accepted by ACM MM 2024! Thanks to all coauthors.

Publications (#: Equal contribution, †: Corresponding author)
RLSF-V RLSF-V: Mitigating Hallucinations in MLLMs via Fuzzy Semantic Self-Feedback

Changhao He, Shuhao Yan, Shuxian Li, Xi Peng, Peng Hu†

ICML, 2026
paper(Coming soon) / code(Coming soon) / Dataset(Coming soon)

Propose a self-feedback framework for constructing preference data that eliminates the need for external large-model evaluators or human annotations.

Present a novel hallucination assessment method that derives local fuzzy semantics directly from internal logits and seamlessly integrates into a preference optimization pipeline.
DOUBT DOUBT: Decoupled Object-level Understanding and Bridging via vMF-based Trustworthiness for Hallucination Detection in MLLMs

Kaiqi Chen, Yang Qin, Changhao He, Xi Peng, Peng Hu†

ICML, 2026, Spotlight (accept rate=2.2%)
paper(Coming soon) / code(Coming soon)

Decouple simple recognition from hard reasoning and guide the model to elicit richer and object-aware responses.

Propose a model-agnostic uncertainty metric based on the vMF distribution to alleviate the instability of traditional entropy measurement.
BML Bootstrapping Multi-view Learning for Test-time Noisy Correspondence

Changhao He, Di Xue, Shuxian Li, Yanji Hao, Xi Peng, Peng Hu†

CVPR, 2026
paper / code / Dataset

Identify and formalize Test-time Noisy Correspondence (TNC) in multi-view/modal learning.

Propose a three-view/modal (RGB-Depth-Text) dataset for multi-view/modal learning community.
TME Learning with Noisy Triplet Correspondence for Composed Image Retrieval

Shuxian Li#, Changhao He#, Xiting Liu, Joey Tianyi Zhou, Xi Peng, Peng Hu†

CVPR, 2025
paper / code

Introducing a novel setting (learning with noisy triplet correspondence) in CIR, thus offering a new design perspective for existing supervised methods.
VITAL Robust Variational Contrastive Learning for Partially View-unaligned Clustering

Changhao He, Hongyuan Zhu, Peng Hu†, Xi Peng

ACM MM, 2024
paper / code

A novel Variational Contrastive Learning paradigm to solve PVP (Partially View-aligned Problem).

Honors and Awards

• First-Class Academic Scholarship, Sichuan University, 2025.07

• National Scholarship, 2024.11

• Outstanding Student of Sichuan University, 2024.10

Services

• Conference Reviewer: CVPR, ICML, NeurIPS, ECCV, PRCV, ICIG, etc.

• Journal Reviewer: IEEE TCSVT, etc.
Hit Counters
Last updated: April 2026
Design and source code from Jon Barron