Estimating choroidal thickness (CT) from fundus images offers a scalable alternative to optical coherence tomography (OCT), but remains challenging under limited supervision. A key difficulty lies in modality pretraining imbalance, where auxiliary modalities lack domain-aligned representations, leading to unstable multimodal fusion. We propose ORCA (Orthogonal Residual Cross-modal Adaptation), an asymmetric dual-stream framework that treats color fundus photography (CFP) as a pretrained anchor and integrates complementary fundus autofluorescence (FAF) information via an Orthogonal Residual Adapter (ORA). By constraining cross-modal updates to orthogonal directions and modulating their magnitude through a norm-driven adaptive gate, ORCA enables stable and sample-adaptive fusion without disrupting pretrained representations. Experiments on a clinical dataset of 390 paired CFP–FAF images demonstrate that ORCA achieves state-of-the-art performance (MAE 46.64 μm), outperforming both unimodal and multimodal baselines. These results highlight the importance of explicitly modeling modality asymmetry and demonstrate that representation-level constraints provide an effective solution for robust multimodal regression in data-scarce settings.
In online medical recommendation systems, modeling heterogeneous interactions between patient needs and medical resources is still challenging under sparse and few-shot scenarios. An Iterative Intra–Inter-Node Contrastive Learning (I3CL) framework is proposed, which alternately updates an intra-node contrastive module for attribute-level discrimination and an inter-node contrastive module for cross-graph alignment, forming a closed-loop optimization process. The BioBERT encoder is frozen to retain medical knowledge and mitigate overfitting. Experiments on two datasets show that I3CL improves Recall, AUC, and F1-score by 5.64
Multi-modal document question-answering aims to search multi-modal evidences distributed across multiple pages for answer generation. Existing methods can be categorized into Large Vision-Language Model (LVLM)-based and Retrieval-Augmented Generation (RAG)-based methods. However, the former were susceptible to hallucinations, while the latter struggled for inter-modal disconnection and cross-page fragmentation. To address these challenges, a novel multi-modal RAG model, named MHier-RAG, was proposed, leveraging multi-modal cues across multi-pages to reason accurate answers. A hierarchical indexing method with the integration of flattened in-page and topological cross-page chunks was designed to jointly establish in-page associations and cross-page dependencies. By means of joint similarity evaluation and large language model (LLM)-based re-ranking, a multi-granularity retriever, including page-level parent page retrieval and document-level summary retrieval, was proposed to encourage multi-modal connection and long-distance reasoning. Experimental results performed on MMLongBench-Doc and LongDocURL demonstrated the superiority of MHier-RAG in understanding modality-rich and multi-page documents. Our code is available at https://github.com/Gzy1112/MHier-RAG .
LLM-based Text-to-SQL progress is reported across heterogeneous benchmarks, backbones, and inference protocols, making cross-system comparison fragile. We reframe the field as a leaderboard aggregation: we collect the metrics authors themselves report and organize them along an inference-autonomy axis spanning constrained, in-context, iterative, agentic, and reasoning-internalized generation, with traceable provenance for every cell. To anchor the aggregation empirically, we run a focused case study on Spider, comparing 8B open-source backbones with and without chain-of-thought (CoT) supervision against few-shot DeepSeek V3 and GLM-4 baselines. Four patterns emerge: Spider gains transfer unevenly to BIRD and Spider 2.0; autonomy buys robustness at non-trivial cost; reasoning internalization sits between answer-only decoding and externally orchestrated agents; and CoT gains concentrate on Hard and Extra-Hard queries. We release a Python harness mirroring the autonomy axis so that future methods can be added directly to the leaderboard.
Existing deep learning-based infrared small-target detection (IRSTD) methods, largely dominated by spatial convolutions, still face several limitations: continuous downsampling weakens tiny target responses, cross-scale fusion introduces spatial misalignment and shallow-feature noise leakage, and background false responses lacking frequency evidence remain insufficiently constrained during optimization. Motivated by the aforementioned limitations, this study introduces SFGA-Net, a novel architecture for IRSTD driven by spatial-frequency joint representation and gated local alignment. Specifically, we design a spatial-frequency joint encoding module that explicitly incorporates the Discrete Cosine Transform into the shallow layers of the network to enhance the representation of high-frequency structures associated with weak targets and alleviate feature dissipation. We further design a gated cross-scale patch alignment module, which improves cross-layer feature alignment and suppresses the propagation of shallow-layer noise to the decoder through local patch-level attention interactions and an asymmetric soft-gating mechanism. In addition, we construct a frequency-evidence-guided false alarm suppression loss, which leverages a Laplacian prior to constrain background responses lacking supporting frequency evidence, thereby reducing unreliable false alarms. Experimental results show that the proposed method achieves competitive detection performance on both the NUAA-SIRST and IRSTD-1 k datasets, with particularly clear advantages in false alarm suppression.
