📡 Signal & Communications

🤖 AAAI2026 · 3 paper notes

Task Aware Modulation Using Representation Learning for Upscaling of Terrestrial Carbon Fluxes

This paper proposes TAM-RL, a framework that formulates terrestrial carbon flux upscaling as a zero-shot regression transfer learning problem. By combining a BiLSTM task encoder with FiLM modulation and a knowledge-guided loss derived from the carbon balance equation, the method achieves a 9.6% reduction in GPP RMSE and a 43.8% improvement in NEE R² over FLUXCOM-X-BASE across 150+ flux tower sites.

Text-Guided Channel Perturbation and Pretrained Knowledge Integration for Unified Multi-Modality Image Fusion

This paper proposes UP-Fusion, a unified multi-modality image fusion framework comprising three modules — Semantic-aware Channel Pruning Module (SCPM), Geometric Affine Modulation (GAM), and CLIP Text-guided Channel Perturbation Module (TCPM) — that employs a single set of weights (trained solely on infrared-visible data) to simultaneously handle both IVIF and medical image fusion tasks, achieving state-of-the-art performance on both.

Toward Gaze Target Detection in Young Autistic Children

To address the severe class imbalance in gaze target detection for autistic children—where face-directed gaze accounts for only 6.6% of samples—this paper proposes the Socially Aware Coarse-to-Fine (SACF) framework. A fine-tuned Qwen2.5-VL serves as a social-context-aware gate that routes inputs to either a socially aware or a socially agnostic expert model. Evaluated on the newly introduced AGT dataset, the framework substantially improves face gaze detection performance (Face L2 reduced by 13.9% on Sharingan; F1 improved from 0.753 to 0.761).