Jiaqi Li, Guangming Wang, Shuntian Zheng, Minzhe Ni, Xiaoman Lu, Guanghui Ye, Yu Guan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Traditional feedback learning for hallucination reduction relies on labor-intensive manual labeling or expensive proprietary models. This leaves the community without foundational knowledge about how to build high-quality feedback with open-source MLLMs. In this work, we introduce RLAIF-V, a novel framework that aligns MLLMs in a fully open-source paradigm. RLAIF-V maximally explores open-source MLLMs from two perspectives, including high-quality feedback data generation for preference learning and self-feedback guidance for inference-time scaling. Extensive experiments on six benchmarks in both automatic and human evaluation show that RLAIF-V substantially enhances the trustworthiness of models at both preference learning and inference time. RLAIF-V 7B reduces object hallucination by 80.7% and overall hallucination by 33.7%. Remarkably, RLAIF-V 12B further reveals the self-alignment potential of open-source MLLMs, where the model can learn from feedback of itself to achieve super GPT-4V trustworthiness.

In recommender systems, learning high-quality user and item representations is crucial for predicting user preferences. However, there are various confounding factors in observational data, resulting in data bias, which hinders the learning of user and item representations. Recent work proposed to use uniform data to alleviate bias problem. However, these methods fail to learn pure representations for unbiased prediction, which are not affected by confounding factors. This paper introduces a novel disentangled framework, named CDLRec, for learning unbiased representations, leveraging uniform data as supervisory signal for disentangling. Furthermore, to address the scarcity problem of uniform data, the contrastive learning is utilized to implement disentanglement by providing augmented samples. Specifically, two contrastive strategies are designed based on different sampling ways for positives and negatives. Extensive experiments are conducted over two real-world datasets and the results demonstrate the superior performance of our proposed method.

The past decade has witnessed the disaggregation of remotely sensed land surface temperature (DLST), which aims for the generation of high temporal and spatial resolution land surface temperature (LST) and which has steadily evolved into a relatively independent subfield of thermal remote sensing. Limited by Tobler's first law of geography, DLST methods require a regression between LSTs and scaling factors using image pixels within a globalized or a localized regression window. Recommendations regarding the selection of the regression window have been provided, but they are mainly subjective and based on highly specific examples. In this context, 100 DLST samples with diversified land cover types and climates were employed to assess the global window strategy (GWS) and the local window strategy (LWS). To optimize disaggregation accuracy and computational complexity, the assessments show that the optimal moving-window size (MWS) for the LWS can be estimated by the resolution ratio between pre- and postdisaggregated LSTs. To identify the better strategy between the GWS and the LWS, an indirect criterion based on aggregation-disaggregation (ICAD) was formulated, which determines the better strategy from medium to high resolution according to the associated performances from low to medium resolution. Validations demonstrate that the accuracy predicted by the ICAD achieves 72%, and in cases in which predictions are incorrect, the performances of the GWS and the LWS are similar. Further evidences indicate that the use of historical high-resolution LSTs improves the LWS by using a locally varying MWS. These findings are able to guide researchers in choosing the most suitable regression window for any particular DLST.

The forest canopy vertical structure, understory spatial distribution as well as bare soil ground under a forest canopy strongly affect the accuracy of retrieving leaf area index (LAI) using optical remotely sensed data. Combining the reflectance from two different view directions and geometric optical (GO) radiative model, we investigated the effects of background information on retrieving forest canopy LAI using moderate resolution data for temperate forest in Northeastern China. Firstly, the forest background reflectance (BR) was computed using the bidirectional reflectance distribution function (BRDF) product from MODIS data. Then the forest canopy effective leaf area index (LAIe) was retrieved using the four scale geometric optical radiative model after removing the effects of forest background. It was found that the LAIe estimates without considering and with considering forest background contribution can explain 46 % and 36 % variations of field measurement-based LAIe (p<; 0.01, N=28), respectively. The forest canopy LAIe estimation without removing background influence can result in the LAIe overestimation ranging from 20 % to 46 %, thus it is necessary to separate the contribution of forest background to total forest canopy reflectance.