While fusing the capacities and advantages of various large language models offers a pathway to construct more powerful and versatile models, a fundamental challenge is to properly select advantageous model during training. Existing fusion methods primarily focus on the training mode that uses cross entropy on ground truth in a teacher-forcing setup to measure a model's advantage, which may provide limited insight towards model advantage. In this paper, we introduce a novel approach that enhances the fusion process by incorporating both the training and inference modes. Our method evaluates model advantage not only through cross entropy during training but also by considering inference outputs, providing a more comprehensive assessment. To combine the two modes effectively, we introduce ProFuser to progressively transition from inference mode to training mode. To validate ProFuser's effectiveness, we fused three models, including Vicuna-7B-v1.5, Llama-2-7B-Chat, and MPT-7B-8K-Chat, and demonstrated the improved performance in knowledge, reasoning, and safety compared to baseline methods.

Recent advances in diffusion models have enabled the creation of deceptively real images, posing significant security risks when misused. In this study, we empirically show that different timesteps of DDIM inversion reveal varying subtle distinctions between synthetic and real images that are extractable for detection, taking the forms of such as Fourier power spectrum high-frequency discrepancies and inter-pixel variance distributions. Based on these observations, we propose a novel detection method named ESIDE that directly utilizes features of intermediately noised images by training an ensemble on multiple noised timesteps, circumventing the overtime of conventional reconstruction-based strategies. To enhance human comprehension, we introduce a metric-grounded explanation refinement module to identify and explain AI-generated flaws. Additionally, we present the benchmarks GenHard and GenExplain, offering detection samples of greater difficulty and high-quality rationales for fake images. Extensive experiments show that ESIDE achieves state-of-the-art performance with 98.91% and 95.89% detection accuracy on regular and challenging samples respectively, and demonstrates generalizability and robustness.

Tianyuan Shi, Canbin Huang, Fanqi Wan, Longguang Zhong, Ziyi Yang, Weizhou Shen, Xiaojun Quan, Ming Yan. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

While fusing heterogeneous open-source LLMs with varying architectures and sizes can potentially integrate the strengths of different models, existing fusion methods face significant challenges, such as vocabulary alignment and merging distribution matrices. These procedures are not only complex but also prone to introducing noise and errors. In this paper, we propose an implicit fusion method, Weighted-Reward Preference Optimization (WRPO), which leverages preference optimization between the source LLMs and the target LLM to transfer their capabilities effectively. WRPO eliminates the need for vocabulary alignment and matrix fusion and can be efficiently scaled to accommodate various LLMs. To address distributional deviations between the source and target LLMs, WRPO introduces a progressive adaptation strategy that gradually shifts reliance on preferred examples from the target LLM to the source LLMs. Extensive experiments on the MT-Bench, AlpacaEval-2, and Arena-Hard benchmarks demonstrate that WRPO consistently outperforms existing knowledge fusion methods and various fine-tuning baselines. When applied to LLaMA3-8B-Instruct as the target model, WRPO achieves a length-controlled win rate of 55.9\% against GPT-4-Preview-1106 on AlpacaEval-2 and a win rate of 46.2\% against GPT-4-0314 on Arena-Hard. Our code is available at https://github.com/SLIT-AI/WRPO.

Large language model (LLM) routers improve the efficiency of multi-model systems by directing each query to the most appropriate model while leveraging the diverse strengths of heterogeneous LLMs. Most existing approaches frame routing as a classification problem based solely on the input query. While this reduces overhead by avoiding inference across all models, it overlooks valuable information that could be gleaned from potential outputs and fails to capture implicit intent or contextual nuances that often emerge only during response generation. These limitations can result in suboptimal routing decisions, particularly for complex or ambiguous queries that require deeper semantic understanding. To address this challenge, we propose Lookahead, a routing framework that "foresees" potential model outputs by predicting their latent representations and uses these predictions to guide model selection, thus enabling more informed routing without full inference. Within this framework, we implement two approaches based on causal and masked language models. Empirical evaluations across seven public benchmarks—spanning instruction following, mathematical reasoning, and code generation—show that Lookahead consistently outperforms existing routing baselines, achieving an average performance gain of 7.7\% over the state-of-the-art. Our code is available at https://github.com/huangcb01/lookahead-routing.

Xiaohua Wang, Zhenghua Wang, Xuan Gao, Feiran Zhang, Yixin Wu, Zhibo Xu, Tianyuan Shi, Zhengyuan Wang, Shizheng Li, Qi Qian, Ruicheng Yin, Changze Lv, Xiaoqing Zheng, Xuanjing Huang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Efficient knowledge retrieval plays a pivotal role in ensuring the success of end-to-end taskoriented dialogue systems by facilitating the selection of relevant information necessary to fulfill user requests.However, current approaches generally integrate knowledge retrieval and response generation, which poses scalability challenges when dealing with extensive knowledge bases.Taking inspiration from open-domain question answering, we propose a retrievergenerator architecture that harnesses a retriever to retrieve pertinent knowledge and a generator to generate system responses.Due to the lack of retriever training labels, we propose relying on feedback from the generator as pseudo-labels to train the retriever.To achieve this, we introduce a dual-feedback mechanism that generates both positive and negative feedback based on the output of the generator.Our method demonstrates superior performance in task-oriented dialogue tasks, as evidenced by experimental results on three benchmark datasets.

Although quite natural for human beings to communicate based on their own personality in daily life, it is rather challenging for neural dialog systems to do the same. This is because the general dialog systems are difficult to generate diverse responses while at the same time maintaining consistent persona information. Existing methods basically focus on merely one of them, ignoring either of them will reduce the quality of dialog. In this work, we propose a two-stage generation framework to promote the persona-consistency and diversity of responses. In the first stage, we propose a persona-guided conditional variational autoencoder (persona-guided CVAE) to generate diverse responses, and the main difference when compared with general CVAE-based model is that we use additional dialog attribute to assist the latent variables to encode the effective information in the response and further use it as a guiding vector for response generation. In the second stage, we employ persona-consistency checking module and the response rewriting module to mask the inconsistent word in the generated response prototype and rewrite it to more consistent. Automatic evaluation results demonstrate that the proposed model is able to generate diverse and persona-consistent responses.