Paper Reading Notes
Read arXiv · Take notes · Summarize · Link to daily digests
Topics I'm Reading
- Recently reading: foundation models and multimodal
- Recently reading: computer vision
- Recently reading: reinforcement learning and agents
Reading Highlights
- Textual Supervision Enhances Geospatial Representations in Vision-Language Models
Experiments show that textual supervision significantly improves geospatial representation accuracy; vision-language models possess good geospatial representations in early language block layers, whereas pure vision models tend to do so in final layers; prompting helps propagate geospatial information across layers.
- Illusions of the Gold Standard: A Large-scale Analysis of Human Evaluation Protocols for Long-form Text Generation
The analysis indicates that the current community generally underreports details regarding human evaluation design and shows a trend towards shifting to LLM evaluators; the authors suggest adopting standardized reporting norms to enhance transparency.
- IEA: Amateur-Friendly Conversational Image Editing Agent via Three Stages of Multitask Alignment
Experiments show that a tool-centric VLM combined with an interpretable action space effectively improves instruction-following capabilities and perceptual quality. The three-stage alignment strategy significantly outperforms strong baselines, and the editing trajectories are transparent and debuggable.
- Robust-U1: Can MLLMs Self-Recover Corrupted Visual Content for Robust Understanding?
Experiments show that high-quality visual recovery directly enhances reasoning performance, establishing self-recovery as a key mechanism for robust visual understanding.
- From Player to Master: Enhancing Test-Time Learning of LLM Agents via Reinforcement Learning over Memory
Experiments show that treating memory updates as a trainable process rather than static prompt engineering significantly improves performance; per-turn reward signals and context-independent advantage estimation aid credit assignment and training stability in multi-turn settings.
Topic Index Preview
Related Reading Roundups
2026-05-06Getting Started with PaddleOCR: Conda Env + GPU Inference End-to-End