Research

Papers by Zichen Tian

PhD candidate, Singapore Management University · Google PhD Fellowship 2026 · Parameter-efficient fine-tuning, LoRA, multi-task learning, remote sensing

Why multi-task LoRA breaks at scale: (A) accuracy vs routing entropy under growing regularization, (B) task alignment concentrated in the top-10% singular components, (C) gradient-conflict distributions for component-level vs block-level LoRA

ICLR 2026

Scalable Multi-Task Low-Rank Model Adaptation

mtLoRA

Zichen Tian, Antoine Ledent, Qianru Sun

Scales multi-task LoRA to 15-25 tasks with Spectral-Aware Regularization, Block-Level Adaptation and Fine-Grained Routing: +2.3% over SOTA, 47% fewer parameters.

88.2 → 2.0%DOTA accuracy when plain multi-task LoRA scales from 5 to 15 tasks

+2.3%average accuracy over the previous state of the art

PEFT on remote-sensing images is sensitive to hyperparameters: (a) PEFT methods compared by tail and overall accuracy, (b) head/tail inter-class feature distance, (c) accuracy by intra-block position and block depth, (d) accuracy by intra-block position and scaling factor

CVPR 2025Highlight

Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning

MetaPEFT

Zichen Tian, Yaoyao Liu, Qianru Sun

Meta-learns module insertion, layer selection and module-wise learning rates for PEFT on long-tailed remote-sensing data.

81.63%mean tail-class accuracy with LoRA + MetaPEFT

+2.1tail-accuracy points on SatMAE → SAR over plain LoRA

debLoRA two key steps: (a) biased LoRA feature space, (b) clustering into attribute clusters, (c) calibrating tail features toward the de-biased center, (d) real samples sharing the streamlined-tail attribute

NeurIPS 2024

Learning De-Biased Representations for Remote-Sensing Imagery

debLoRA

Zichen Tian, Zhaozheng Chen, Qianru Sun

Unsupervised, LoRA-agnostic de-biasing of long-tailed remote-sensing features via clustering and calibration; up to +4.7 tail-class points.

+4.7tail-class points, optical RS → SAR

+3.3tail-class points, natural → optical RS