Li Shang

Large language models (LLMs) exhibit strong performance across applications, but their inference is computationally intensive, posing significant challenges for edge deployment. Quantization is among the most effective and widely used optimizations. In particular, ternary-weight quantization further lowers compute cost by replacing many multiplications with lightweight additions, thereby reducing complexity and energy. However, existing GPU/ASIC solutions often exhibit low utilization for ternary-weight LLMs and fail to exploit their inherent sparsity, leading to inefficient execution. To address this, we propose SATLLM, a lightweight accelerator optimized for ternary weights and sparsity. At the algorithm level, we propose a co-optimization method based on unstructured pruning and ternary-weight clustering to effectively enhance the sparsity. To further exploit this sparsity, we design a hardware-friendly sparsity-aware merging scheme and develop a customized BitLinear engine to support efficient sparse computation. In addition, we optimize the dataflow between the BitLinear engine and the GEMM unit, effectively alleviating performance bottlenecks caused by workload imbalance. Together, these optimizations significantly improve the throughput and energy efficiency of ternary-weight LLM inference. The experiments show that our proposed accelerator achieves an energy efficiency of 9.42 TOPS/W, with improvements of up to 42.41 ×, 16.99 ×, 4.42 ×, 2.49 ×, and 1.90 × compared to NVIDIA V100, A100, ANT, MECLA, and RMFA, respectively.

Mixture-of-Experts (MoE) architectures scale Large Language Models via expert specialization induced by conditional computation. In practice, however, expert specialization often fails: some experts become functionally similar, while others functioning as de facto shared experts, limiting the effective capacity and model performance. In this work, we analysis from a spectral perspective on parameter and gradient spaces, uncover that (1) experts share highly overlapping dominant spectral components in their parameters, (2) dominant gradient subspaces are strongly aligned across experts, driven by ubiquitous low-rank structure in human corpus, and (3) gating mechanisms preferentially route inputs along these dominant directions, further limiting specialization. To address this, we propose Spectral-Decoupled MoE (SD-MoE), which decomposes both parameter and gradient in the spectral space. SD-MoE improves performance across downstream tasks, enables effective expert specialization, incurring minimal additional computation, and can be seamlessly integrated into a wide range of existing MoE architectures, including Qwen and DeepSeek.

Processors serve as the cornerstone of modern computing systems. Although processor design encompasses multiple VLSI levels, the architectural design plays a critical role in determining performance, power consumption, and area efficiency. To address the growing pressure to shorten chip time-to-market, there is an increasing demand for rapid iteration methods in processor architecture development. To achieve efficient and effortless architecture design optimization, we develop ChatArch, a knowledge-driven graph-of-thought multi-LLM-agent framework for processor architecture optimization. Based on processor architecture expertise, we decompose the processor architecture design space and construct an LLM agent graph-of-thought framework to characterize and iteratively optimize these subspaces. Also, by systematically consolidating domain-specific knowledge and empirical design principles validated by experts, we establish a comprehensive RISC-V processor design knowledge repository. Moreover, a knowledge-driven multi-agent framework is developed to enable efficient microarchitecture optimization. Finally, the optimized microarchitecture modules aggregate to form system-level designs. This methodology achieves automated iterative optimization of microarchitectures targeting PPA objectives while generating corresponding behavioral models. The experiments demonstrate that our method effectively designs behavioral processor models, with LLM-generated architectures achieving a validation success rate of over 97.39%, surpassing the performance of other LLMs, including GPT-4o. ChatArch consistently meets requirements, delivering up to 9.97% times better PPA and 32–68x efficiency gains compared with traditional black-box optimization methods.

As recommender systems become increasingly important components in online communities, studying their impact on these communities becomes ever more crucial. Facing the high costs and ethical risks of real-world social experiments, researchers construct recommendation simulators to study the interactions between recommender systems and users. However, existing simulators face challenges in providing universal, accurate, and scalable interaction modeling for various types of online communities involving millions of contents and users with diverse action types. To address these challenges, we propose RECOSIM, a simulation framework capable of offering efficient recommendation interaction simulations across a wide range of scenarios. RECOSIM decomposes the user agent into five fundamental modules: Encode Model, Decode Model, Activity Model, Scoring Model, and Generation Model, allowing for accurate and extensible modeling of user behavior and interaction dynamics. The recommender system agent adheres to established industry architectures, implementing three stages and four fundamental strategies, thereby improving generalizability across various platforms and the computational efficiency of simulation. Utilizing two real-world datasets (Weibo and Zhihu), we validate the accuracy and stability of each component and the overall framework of RECOSIM, demonstrating the reliability of RECOSIM as a simulation environment. Subsequently, we delve into analyzing the impact of the four fundamental recommendation strategies on online communities, providing design inspirations for enhancing user engagement and community growth.

Integrating vision-language models (VLMs) with wearable devices offers great potential for continuous and responsive video understanding, a key capability for applications such as smart eyewear-based conversational assistants. However, achieving this on resource-constrained devices is challenging due to the high energy demands of continuous spatial-temporal sampling and transmission. We propose ActiveEye , a VLM designed for energy-efficient and responsive video understanding. ActiveEye separates visual and motion semantic representations and incorporates an active perception-based feedback path to adaptively adjust spatial-temporal sampling and transmission rates. Implemented as a wearable-mobile-cloud system, ActiveEye is evaluated for energy efficiency, real-time semantic change detection, and video understanding in both laboratory and field studies. Using the EgoSchema dataset, ActiveEye reduces the front-end energy consumption by 49.14%, supporting 8.37 hours of continuous operation on a 2.1 Wh battery. It achieves the highest F1 score (0.80) and the lowest average time difference (1.30 s) compared with heuristic-based event detection algorithms, validating its timely semantic detection. Furthermore, ActiveEye achieves a visual question answering (VQA) accuracy of 61.6%, which is comparable to state-of-the-art VLM agents, despite their reliance on larger language decoders and more computationally intensive frame selection strategies. Two rounds of in-field user evaluations further confirm its effectiveness in real-world settings, demonstrating its practical viability as a continuous and responsive video understanding system, conversational assistant, and wearable companion.