Neural & Evolutionary

Han Xu, Xuerui Qiu, Baiyu Chen et al.

This work addresses the challenge of building energy-efficient LLMs with billions of parameters by integrating brain-inspired spiking mechanisms, offering a potential architecture for neuromorphic hardware.

Jesse Cool, Benedikt Hartl, Michael Levin et al.

This work provides a foundational understanding of how emergent agents' behavioral propensities interact with informational topography, relevant for researchers studying artificial life and emergent behavior.

Lakshya A Agrawal, Donghyun Lee, Shangyin Tan et al.

Demonstrates that text optimization with LLM-based search is a general-purpose problem-solving paradigm, unifying tasks traditionally requiring domain-specific algorithms.

Han Xu, Zhiyong Qin, Di Shang et al.

This work addresses the computational and energy bottlenecks of MLLMs for deployment in resource-constrained environments by proposing a spike-based approach with algorithm-hardware co-design.

Tao Jiang, Xinmeng Yu, Chenhao Yi et al.

This work addresses the inefficiency of hand-crafted operators in evolutionary model merging for LLMs, offering a more data-efficient and robust optimization method.

Jacob Fein-Ashley, Paria Rashidinejad

For practitioners of large-scale language models and reasoning systems, Attractor Models offer a scalable way to incorporate iterative refinement without the overhead of deep recurrence.

Manuj Malik, Jianan Zhou, Shashank Reddy Chirra et al.

This addresses the problem of reducing manual tuning in combinatorial optimization for researchers and practitioners, though it is incremental as it builds on existing LLM-driven evolutionary methods.

Xiao Wang, Yiyang Wu, Yuntian Wang et al.

This addresses a bottleneck in computational imaging for applications like biomedical imaging and remote sensing, but it is incremental as it builds on existing SPI methods.

Joel Z. Leibo, Alexander Sasha Vezhnevets, Manfred Diaz et al.

This offers a novel theoretical framework for cognitive science and social norms, potentially redefining rationality, but it is incremental as it builds on existing models like LLMs.

Ahmadreza Jeddi, Minh Ngoc Le, Hakki C. Karaimer et al.

For developers of autonomous research agents, GEAR demonstrates that maintaining multiple promising directions and adapting search strategy over time improves effectiveness.

Ryan Bahlous-Boldi, Isha Puri, Idan Shenfeld et al.

For LLM post-training, VPO addresses the bottleneck of low-entropy responses that hinder test-time search, offering a drop-in replacement for GRPO that improves search performance.

Ye Lu, Bingdong Li, Aimin Zhou et al.

For researchers and practitioners in expensive black-box optimization, this work provides a scalable, retraining-free surrogate paradigm that outperforms strong baselines.

Xudong Wang, Chaoning Zhang, Chenghao Li et al.

For practitioners using LLMs for complex reasoning, Agent-GWO provides an automatic, robust prompt optimization method that jointly optimizes prompts and decoding settings, reducing performance fluctuations and improving transferability.

Nico Pelleriti, Sree Harsha Nelaturu, Zhanke Zhou et al.

For researchers developing and evaluating evolutionary coding agents, this work provides a diagnostic methodology to distinguish between different mechanisms behind performance gains, revealing that final benchmark scores can be misleading.

Zecheng Hao, Shenghao Xie, Kang Chen et al.

This work addresses the triad of memory overhead, learning capability, and energy budget in S-ViTs, which is a domain-specific problem for energy-efficient AI applications, and it is novel as the first to systematically establish multi-dimensional grouped computation for this purpose.

Vugar Ismailov

For researchers in approximation theory and geometric deep learning, it provides a theoretical foundation for using neural networks on non-Euclidean data, though the results are largely theoretical and incremental.

Dat Thanh Tran, Van Khu Vu, Yining Ma

For combinatorial optimization practitioners, DyNACO offers a scalable, dynamic neural guidance framework that improves upon static priors, with demonstrated generalization and efficiency gains.

Ethan Caballero, Priyank Jaini, David Krueger et al.

Provides a more accurate tool for predicting performance of large-scale neural networks across diverse domains, aiding resource allocation and architecture design.

Mingyang Yu, Jiaqi Zhang, Haorui Yang et al.

This addresses computational challenges in finance for portfolio managers, though it appears incremental as it builds on existing quantum-inspired and evolutionary methods.

Rui Tang, Kaiyu Xu, Pengsen Cheng et al.

For LLM safety researchers, EvoJail provides a more adaptive and diverse automated jailbreak generation method to uncover safety weaknesses across evolving models.