Large language models (LLMs) frequently generate fluent yet factually inaccurate content, a phenomenon known as hallucination. Recent inference-time approaches aim to improve truthfulness by steering model activations toward semantically meaningful directions. While effective to some extent, these methods typically process activations independently, neglecting the internal coordination structure of multi-head attention (MHA), where attention heads interact to form semantic representations. In this work, we propose CoFact, an adaptive inference-time mechanism that improves factual consistency by dynamically coordinating attention head behaviors. Inspired by cooperative game theory, CoFact conceptualizes attention heads as collaborative agents. It models the semantic utility and redundancy of each head and adaptively modulates their contributions to the final attention output. Notably, rather than directly altering intermediate representations, CoFact performs token-level coordination to encourage diverse and complementary attention patterns across heads. CoFact is plug-and-play compatible with mainstream LLM architectures and requires no additional supervision or model retraining. Experimental results across multiple standard factuality benchmarks demonstrate that CoFact consistently enhances factual accuracy while maintaining generation fluency.

Ya Su, Hu Zhang, Dan Qiao, YuJie Wang, Yunxiao Zhao, Yue Fan, Shike Li, Ru Li, Hongye Tan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The interconnected nature of the Internet of Things (IoT) significantly enhances the efficiency of industries such as smart manufacturing, but it also raises concerns about data privacy. Federated learning (FL) utilizes an edge-cloud collaborative mode that shares models in the cloud instead of data, effectively mitigating data privacy leakage in edge IoT devices. However, FL suffers from the risk of model leakage, and both the cloud and the edge may illegally copy and sell the model, infringing on model ownership. To prevent such misbehavior, it is essential to design a robust method for verifying the model ownership. In this paper, we propose Fed-RWM, a novel watermarking method for FL models that provides robust ownership verification and avoids both edge and cloud leakage of watermark information. Fed-RWM trains the watermark by sharing parameters with an additional model and incorporates a watermark recovery method during verification to counter complex watermark removal attacks, which enhances the verification robustness. Fed-RWM introduces a new training paradigm that performs continuous watermarking training at the FL task initiator, preventing access to watermark information at both the edge and the cloud. The experimental results demonstrate that Fed-RWM performs well in model ownership verification and fidelity, is robust to different watermark removal attacks, and can provide reliable protection for federated learning models.

In modern urban areas, inefficiency traffic management is one of the main causes of road congestion, leading to reduced fuel efficiency and increased traffic safety hazards. Traditional researches typically focus only on enhancing the throughput of intersections by optimizing traffic signals or individual vehicle trajectories. However, these methods often overlook the dynamic nature of the traffic system and the potential benefits of vehicle platooning, limiting their effectiveness in complex traffic environments. Addressing this challenge, this article presents MARP, a Cooperative Multiagent deep reinforcement learning (DRL) System for connected autonomous vehicle (CAV) Platooning. Utilizing vehicle to vehicle (V2I) and vehicles to infrastructure (V2V) technologies, MARP integrates sensing, computing, and communication to collect and process real-time data on traffic conditions, thereby achieving dynamic synchronization between traffic signal controllers and CAV platoons. By constructing platoons that collaborates with the infrastructure through a multiagent DRL collaboration model, MARP adapts to real-time traffic flow changes, significantly optimizing the fluidity and efficiency of the entire traffic network. Detailed experiments show that MARP effectively reduces traffic congestion, shortens intersection travel times, and cuts fuel consumption and emissions, surpassing the state-of-the-art approach.

Future worldwide 6G research will drive the evolution of emerging intelligent control technologies, such as intelligent speed advisory systems (ISA), to a more advanced generation. As a special type of ISA, consensus-based speed advisory systems (CSAS) can be widely used to recommend a consensus speed for a vehicle platoon, enabling minimizing energy consumption or emissions over a planned route. Recently, speed recommendation services that protect data privacy (i.e., how to obtain an optimal speed in a privacy-preserving way) have drawn tremendous attention. However, current approaches could still encounter service trust issues with central servers and the malicious behavior of vehicles. Furthermore, existing research lacks considering road safety constraints (i.e., safe distance between adjacent vehicles and road speed limits) that are essential for the practical deployment of CSAS. To address the above issues, this paper proposes Eco-CSAS, a safe and eco-friendly consensus speed advisory system using blockchain. We formulate an optimization problem subject to the minimum following distance and maximum road speed limit to minimize the energy consumption of the automatic vehicle platoon. In addition, we introduce a consortium blockchain and cryptographic primitives to ensure service trust and data privacy. We implement the system on the Hyperledger platform, and experimental results show that the system can achieve speed recommendations in a trustworthy and privacy-preserving manner while ensuring a secure platoon.