Advait Gosai, Tyler Vuong, Utkarsh Tyagi, Steven Li, Wenjia You, Miheer Bavare, Arda Uçar, Zhongwang Fang, Brian Jang, Bing Liu, Yunzhong He. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large Language Models (LLMs), while demonstrating remarkable capabilities across various applications, present significant challenges during inference due to their substantial model size, especially when deployed on edge devices. Activation sparsity offers a promising solution to reduce computation and memory movement, enabling more efficient inference, particularly for small-batch on-device applications. However, current approaches face limitations with non-ReLU activation function, which are foundational to most advanced LLMs, or require heavy continual training. Additionally, the difficulty in predicting active channels and limited achievable sparsity ratios constrain the effectiveness of activation sparsity-based methods. In this paper, we introduce R-Sparse, a training-free activation sparsity approach capable of achieving high sparsity levels in advanced LLMs. We conducted two preliminary investigations into how different components contribute to the output within a single linear layer and found two key observations: (i) the non-sparse components of the input function can be regarded as a few bias terms, and (ii) The full computation can be effectively approximated by an appropriate combination of input channels and weight singular values. Building on this, we replace the linear layers in LLMs with a rank-aware sparse inference method that leverages the sparsity of input channels and singular value components, eliminating the need for active channel prediction like the output sparsity based approaches. Experiments on Llama-2/3 and Mistral models across ten diverse tasks demonstrate that R-Sparse achieves comparable performance at 50\% model-level sparsity, resulting in a significant 43\% end-to-end efficient improvements with customized kernels.

Deep reinforcement learning (RL) algorithms typically parameterize the policy as a deep network that outputs either a deterministic action or a stochastic one modeled as a Gaussian distribution, hence restricting learning to a single behavioral mode. Meanwhile, diffusion models emerged as a powerful framework for multimodal learning. However, the use of diffusion policies in online RL is hindered by the intractability of policy likelihood approximation, as well as the greedy objective of RL methods that can easily skew the policy to a single mode. This paper presents Deep Diffusion Policy Gradient (DDiffPG), a novel actor-critic algorithm that learns from scratch multimodal policies parameterized as diffusion models while discovering and maintaining versatile behaviors. DDiffPG explores and discovers multiple modes through off-the-shelf unsupervised clustering combined with novelty-based intrinsic motivation. DDiffPG forms a multimodal training batch and utilizes mode-specific Q-learning to mitigate the inherent greediness of the RL objective, ensuring the improvement of the diffusion policy across all modes. Our approach further allows the policy to be conditioned on mode-specific embeddings to explicitly control the learned modes. Empirical studies validate DDiffPG's capability to master multimodal behaviors in complex, high-dimensional continuous control tasks with sparse rewards, also showcasing proof-of-concept dynamic online replanning when navigating mazes with unseen obstacles. Our project page is available at https://supersglzc.github.io/projects/ddiffpg/.

As an excellent optimization algorithm widely used to solve various practical problems, differential evolution (DE) algorithm has few parameters, yet its performance is significantly affected by these parameters. To address this issue, this paper presents a novel variant of DE called DACDE, which utilizes data fusion-based parameter adaptation and a complementary mutation strategy. Most parameter adaptation methods based on successful history only analyze the mean of parameters and ignore their degree of dispersion. In DACDE, the successful parameter distribution is recorded and described by both the mean and variance. Data fusion is then used to combine records and generate an estimated distribution, which is applied to a Gaussian distribution to generate new parameters. Inspired by opposition-based learning, we introduce a complementary mutation strategy. This strategy employs a symmetric selection mechanism to adapt to the varying search abilities required by the algorithm at different stages. The new variant is verified on 32 single-objective functions from CEC 2011 and 2014 benchmark suites, and the results show that DACDE is competitive compared to other 29 evolutionary algorithms.

In recent years, extensive research in the HCI literature has explored interactive techniques for digital fabrication. However, little attention in this body of work has examined how to involve and guide human workers in fabricating larger-scale structures. We propose a novel model of crowdsourced fabrication, in which a large number of workers and volunteers are guided through the process of building a pre-designed structure. The process is facilitated by an intelligent construction space capable of guiding individual workers and coordinating the overall build process. More specifically, we explore the use of smartwatches, indoor location sensing, and instrumented construction materials to provide real-time guidance to workers, coordinated by a foreman engine that manages the overall build process. We report on a three day deployment of our system to construct a 12-tall bamboo pavilion with assistance from more than one hundred volunteer workers, and reflect on observations and feedback collected during the exhibit.

This paper presents a ranked differential evolution (RDE) algorithm for solving the identification problem of non-linear discrete-time systems based on a Volterra filter model. In the improved method, a scale factor, generated by combining a sine function and randomness, effectively keeps a balance between the global search and the local search. Also, the mutation operation is modified after ranking all candidate solutions of the population to help avoid the occurrence of premature convergence. Finally, two examples including a highly nonlinear discrete-time rational system and a real heat exchanger are used to evaluate the performance of the RDE algorithm and five other approaches. Numerical experiments and comparisons demonstrate that the RDE algorithm performs better than the other approaches in most cases.

The tasks of entrapping/escorting and patrolling around an autonomous target are presented making use of the multi-robot cluster space control approach. The cluster space control technique promotes simplified specification and monitoring of the motion of mobile multi-robot systems of limited size. Previous work has established the conceptual foundation of this approach and has experimentally verified and validated its use for 2-robot, 3-robot and 4-robot systems, with varying implementations ranging from automated trajectory control to human-in-the-loop piloting. In this publication, we show that the problem of entrapping/escorting/patrolling is trivial to define and manage from a cluster space perspective. Using a 3-robot experimental testbed, results are shown for the given tasks. We also revise the definition of the cluster space framework for a three-robot formation and incorporate a robot-level obstacle avoidance functionality.