Multi-LLM systems harness the complementary strengths of diverse Large Language Models, achieving performance and efficiency gains unattainable by a single model. In existing designs, LLMs communicate through text, forcing internal representations to be transformed into output token sequences. This process both loses rich semantic information and incurs token-by-token generation latency. Motivated by these limitations, we ask: Can LLMs communicate beyond text? Oracle experiments show that enriching the KV-Cache semantics can improve response quality without increasing cache size, supporting KV-Cache as an effective medium for inter-model communication. Thus, we propose Cache-to-Cache (C2C), a new paradigm for direct semantic communication between LLMs. C2C uses a neural network to project and fuse the source model's KV-cache with that of the target model to enable direct semantic transfer. A learnable gating mechanism selects the target layers that benefit from cache communication. Compared with text communication, C2C utilizes the deep, specialized semantics from both models, while avoiding explicit intermediate text generation. Experiments show that C2C achieves 8.5-10.5% higher average accuracy than individual models. It further outperforms the text communication paradigm by approximately 3.0-5.0%, while delivering an average 2.0x speedup in latency. Our code is available at https://github.com/thu-nics/C2C.

Personalized large language models (LLMs) aim to tailor interactions, content, and recommendations to individual user preferences. While parameter-efficient fine-tuning (PEFT) methods excel in performance and generalization, they are costly and limit communal benefits when used individually. To this end, we introduce Personalized Pieces (Per-Pcs), a framework that allows users to safely share and assemble personalized PEFT efficiently with collaborative efforts. Per-Pcs involves selecting sharers, breaking their PEFT into pieces, and training gates for each piece. These pieces are added to a pool, from which target users can select and assemble personalized PEFT using their history data. This approach preserves privacy and enables fine-grained user modeling without excessive storage and computation demands. Experimental results show Per-Pcs outperforms non-personalized and PEFT retrieval baselines, offering performance comparable to OPPU with significantly lower resource use across six tasks. Further analysis highlights Per-Pcs's robustness concerning sharer count and selection strategy, pieces sharing ratio, and scalability in computation time and storage space. Per-Pcs's modularity promotes safe sharing, making LLM personalization more efficient, effective, and widely accessible through collaborative efforts.

Diffusion models have recently gained significant attention in both academia and industry due to their impressive generative performance in terms of both sampling quality and distribution coverage. Accordingly, proposals are made for sharing pre-trained diffusion models across different organizations, as a way of improving data utilization while enhancing privacy protection by avoiding sharing private data directly. However, the potential risks associated with such an approach have not been comprehensively examined. In this paper, we take an adversarial perspective to investigate the potential privacy and fairness risks associated with the sharing of diffusion models. Specifically, we investigate the circumstances in which one party (the sharer) trains a diffusion model using private data and provides another party (the receiver) black-box access to the pre-trained model for downstream tasks. We demonstrate that the sharer can execute fairness poisoning attacks to undermine the receiver's downstream models by manipulating the training data distribution of the diffusion model. Meanwhile, the receiver can perform property inference attacks to reveal the distribution of sensitive features in the sharer's dataset. Our experiments conducted on real-world datasets demonstrate remarkable attack performance on different types of diffusion models, which highlights the critical importance of robust data auditing and privacy protection protocols in pertinent applications.

Misinformation is a global concern and limiting its spread is critical for protecting democracy, public health, and consumers. We propose that consumers' own social media post-histories are an underutilized data source to study what leads them to share links to fake-news. In Study 1, we explore how textual cues extracted from post-histories distinguish fake-news sharers from random social media users and others in the misinformation ecosystem. Among other results, we find across two datasets that fake-news sharers use more words related to anger, religion and power. In Study 2, we show that adding textual cues from post-histories improves the accuracy of models to predict who is likely to share fake-news. In Study 3, we provide a preliminary test of two mitigation strategies deduced from Study 1 - activating religious values and reducing anger - and find that they reduce fake-news sharing and sharing more generally. In Study 4, we combine survey responses with users' verified Twitter post-histories and show that using empowering language in a fact-checking browser extension ad increases download intentions. Our research encourages marketers, misinformation scholars, and practitioners to use post-histories to develop theories and test interventions to reduce the spread of misinformation.

The study of network sizes in digital social networks is a research question of significant interest. Here, we explore the phenomenon of trimming, which is the decrease in the size of one’s network, and analyze if the rules of social exchange theory – namely, status consistency and reciprocity- can affect trimming. To this end, we use a Hidden Markov Model to investigate the relationship between the frequency of interaction and one’s network size, in which we are able to control for the current size of one’s digital social network. We find that there are significant patterns in sharing tendencies in digital social networks. One is that users who do not share enough are the group that is most likely to be trimmed from a network. Another is that users prefer to have moderate sized networks, i.e. networks with 500 – 1000 friends and prefer friends with moderate sharing tendencies (sharing approximately once a week). We also find that one’s sharing preferences over time tend to align with moderate sharing.