ETH Zurich Postdoc Bingcong Li Joins HKU to Drive Efficient LLM Scaling Research

ETH Zurich Computer Science postdoctoral researcher Bingcong Li (李冰聪) is set to join the University of Hong Kong (HKU) as a tenure-track assistant professor. Li earned his bachelor's degree with Highest Honors from Fudan University's School of Information Science and Engineering, receiving the National Scholarship twice. He completed his Ph.D. in Electrical and Computer Engineering at the University of Minnesota in 2022 and joined ETH Zurich as a postdoc in 2024, securing multiple compute grants from the Swiss AI Initiative.

Questioning Naive Parameter Scaling

In a 2026 blog post titled Scaling Almost Anything? , Li examines whether simply increasing model size guarantees better performance. He argues that parameter count is only a rough proxy for model capacity; what matters is whether the training process can effectively utilize that capacity. Three case studies illustrate the point:

LoRA fine-tuning: Increasing the adapter rank does not necessarily yield better results.

Transformer layers: Some layers may not be fully utilized during training.

Mixture-of-Experts (MoE) with shared experts: Activating more experts does not continuously improve performance.

These examples converge on a single insight: adding parameters without improving capacity utilization merely increases compute cost. Therefore, architectural design and optimization methods that boost utilization efficiency are equally critical.

Width vs. Depth Scaling: An Optimization-Aware Framework

In his 2026 academic report Efficient Scaling of LLMs via Optimization-Aware Architecture Design , Li decomposes model scaling into two dimensions:

Width Scaling – introducing manifold structure to improve the scaling efficiency of linear layers in neural networks.

Depth Scaling – studying how the topology of residual connections affects optimization and convergence.

This work spans large language model pre-training, Diffusion Transformers (DiT), and reinforcement learning tasks. (Figures illustrating the framework are included in the original article.)

Zero-Order Optimization for Private Fine-Tuning

Li's research on training efficiency includes DPZero , published at ICML 2024 . DPZero combines zero-order optimization with differential privacy to enable private fine-tuning of language models without relying on traditional backpropagation.

Parameter-Efficient Fine-Tuning: RefLoRA and PoLAR

In the LoRA ecosystem, Li contributed two papers accepted at NeurIPS 2025 :

RefLoRA – re-parameterizes low-rank adapters for more efficient fine-tuning.

PoLAR – leverages polar decomposition to explore more efficient adapter representations.

Both approaches aim to reduce the parameter and compute overhead of adaptation while maintaining performance.

Unified Research Agenda

From architecture design to optimization algorithms and fine-tuning techniques, Li's work consistently targets efficient scaling : how to lower training costs and improve compute efficiency as model scale continues to grow. The article concludes that competition in large models is no longer just about parameters, data, and compute; unlocking the full potential of scaling through better architecture and optimization remains an open and vital research direction.

As Li begins his faculty position at HKU, his future contributions to AI scaling and efficient optimization will be closely watched.