LINs Lab

Our research interests lie in the intersection of optimization and generalization for deep learning:

leveraging theoretical/empirical understanding (e.g., loss landscape, and training dynamics)
to design efficient & robust methods (both learning and inference)
for deep learning (centralized) and collaborative deep learning (distributed and/or decentralized),
under imperfect environments (e.g., noisy, heterogeneous, and hardware-constrained).

We are building world-model-grounded native multi-modal agents that keep improving themselves.

News

Oct 20, 2026	Our recent preprints advance how agents learn, reason, and collaborate: ProAS and TPPO improve credit assignment for multi-turn agents and language-model reasoning, while Triage-PiEvo and COEVOLVE study evidence verification and sharing for scientific discovery. MS-Eval benchmarks sub-agent delegation and orchestration in MCP-intensive workflows.
Sep 25, 2026	Four papers from our group were accepted at NeurIPS 2026. Congratulations to all the authors! Manifold Drift in Flow Preference Optimization: A Root Cause of Reward Hacking Representation Preconditioning for Efficient Diffusion Model Training Optimizing Decoding Paths in Masked Diffusion Models by Quantifying Uncertainty Duality Models: An Embarrassingly Simple One-step Generation Paradigm
Sep 4, 2026	We are excited to release LLaDA-Image (arXiv and code), a 6B unified diffusion model family for high-quality image generation and editing. The Base and Turbo checkpoints, inference code, and training recipes are publicly available.
Apr 28, 2026	Together with Inclusion AI, we are pleased to announce the release of LLaDA2.0-Uni.
Feb 21, 2026	Several papers from our group were accepted at this year’s CVPR 2026 conference. Rethinking UMM Visual Generation: Masked Modeling for Efficient Image-Only Pre-training Dual-Granularity Memory for Efficient Video Generation Exploring Spatial Intelligence from a Generative Perspective Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching

Selected publications

LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes

Chen, Chuyan, Chen, Haoxing, Chen, Kun, Cheng, Zhenglin, Cui, Long, Fang, Ruishan, Gu, Zhangxuan, Huang, Zhicheng, Lan, Zhenzhong†, Lei, Yuanting, Li, Haoquan, Li, Jianguo†, Li, Rongchuan, Li, Sidu, Lin, Tao†, Liu, Deyuan, Liu, Jiacheng, Liu, Lin, Lou, Yuxuan, Lu, Zhisheng, Ma, Yuxin, Shen, Shuheng, Sun, Peng, Wang, Chaoyang, Wang, Hongjun, Wang, Xiaomei, Wang, Yongxin, Wu, Chengzhang, Wu, Hongru, and Xie, Jun

(arXiv:2609.03796) preprint

PDF Code
ICML 2026

Principle-Evolvable Scientific Discovery via Uncertainty Minimization

Pu, Yingming, Lin, Tao†, and Chen, Hongyu

In International Conference on Machine Learning (ICML), 2026

PDF Code
ICLR 2026

TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows

Cheng, Zhenglin*, Sun, Peng*, Li, Jianguo, and Lin, Tao

In International Conference on Learning Representations (ICLR), 2026

PDF Code
ICLR 2025

DeFT: Decoding with Flash Tree-Attention for Efficient Tree-structured LLM Inference

Yao, Jinwei*, Chen, Kaiqi*, Zhang, Kexun, You, Jiaxuan, Yuan, Binhang, Wang, Zeke, and Lin, Tao

In International Conference on Learning Representations (ICLR), Spotlight, abridged in ICLR workshop AGI (Oral), 2025

PDF Code
NeurIPS 2024

Efficiency for Free: Ideal Data Are Transportable Representations

Sun, Peng, Jiang, Yi, and Lin, Tao

In Advances in Neural Information Processing Systems (NeurIPS), 2024

PDF Code
ICLR 2020

Don’t Use Large Mini-batches, Use Local SGD

Lin, Tao, Stich, Sebastian U, Patel, Kumar Kshitij, and Jaggi, Martin

In International Conference on Learning Representations (ICLR), 2020

PDF Code