Despite its success in enriching LLMs with external knowledge, RAG remains plagued by faithfulness hallucinations, where generated text contradicts the retrieved source information. Previous research on faithfulness hallucination in LLMs is frequently hindered by prohibitive manual annotation costs and a dependency on static datasets, which caps their performance and adaptability. Furthermore, these models lack a clear training mechanism to explicitly promote contextual focus. In this work, we propose a novel iterative self-evolution framework to enhance model faithfulness. This framework autonomously generates high-quality data and leverages it for the continuous self-optimization of the model, leading to significant improvements in faithfulness. Our experimental analysis reveals that improving model faithfulness encourages a closer alignment of the attention distribution with the given context. Based on this finding, we design an attention-based loss function to further promote this process. Experimental results show that our model achieves state-of-the-art faithfulness on a range of context-based question-answering datasets, marking a significant advancement over previous approaches.

Faithfulness hallucinations are claims generated by a Large Language Model (LLM) not supported by contexts provided to the LLM. Lacking assessment standards, existing benchmarks focus on “factual statements” that rephrase source materials while overlooking “cognitive statements” that involve making inferences from the given context. Consequently, evaluating and detecting the hallucination of cognitive statements remains challenging. Inspired by how evidence is assessed in the legal domain, we design a rigorous framework to assess different levels of faithfulness of cognitive statements and introduce the CogniBench dataset where we reveal insightful statistics. To keep pace with rapidly evolving LLMs, we further develop an automatic annotation pipeline that scales easily across different models. This results in a large-scale CogniBench-L dataset, which facilitates training accurate detectors for both factual and cognitive hallucinations. We release our model and datasets at: https://github.com/FUTUREEEEEE/CogniBench

Traditional sound source localization methods encounter significant challenges in simultaneously locating rotating and static sources. These challenges arise from the different motion patterns of these two types of sound sources, and they are typically not situated on the same plane. To address this issue, a method based on Modal Composition Beamforming (MCB) and the equivalent source method is proposed for separating rotating and static sound sources, which fully utilizes the prior knowledge of the spatiotemporal properties of these sources. The proposed approach involves establishing a Rotating-Static Sources Power Propagation (R-S2P) model, utilizing the relationship between the equivalent source strength and the actual beamforming output. By employing this forward model and applying an appropriate inversion method, it is possible to separate the components of rotating and static sources. Simulations for three cases with different source strengths are presented, and the R-S2P inversion problem is resolved using the Least Absolute Shrinkage and Selection Operator (LASSO) method. We showed that this method enables accurate separation and localization of rotating and static sources on different planes with varying relative intensities, even if the background noise is strong.

Rotating source beamforming techniques have been effective means of noise localization on rotary machines. In this letter, we derive an alternative expression for modal composition beamforming (MCB) and subsequently consider the equivalent source assumption and cyclostationarity of the constant angular-speed rotating sound source so that a rotating sound source power (RSP) propagation model is derived. By estimating a suitable solution for the RSP model using the subspace variational Bayesian (SVB) technique with sparsity and total variation (TV) priors, the validity of the RSP model was established. According to the simulation results, the proposed RSP-SVB method leads to a significantly higher resolution than the MCB method. It can localize multiple fast-rotating sound sources accurately, rapidly, and effectively in environments with strong background noise interference. Therefore, our proposed RSP-SVB can offer a reliable solution for identifying fast-rotating blade noise.