Boosting Fine-Grained Visual Perception via Multi Granularity Data Synthesis and Local-to-Global Curriculum Learning

W. Yan^#, Y. Liu^#, D. Zou et al.

Vision-language models can describe a scene while still missing object attributes or spatial relationships. This work addresses that gap with a multi-granularity data pipeline and Local-to-Global Curriculum Reinforcement Learning (LoGo-CRL).

The pipeline combines object tags and grounded regions with teacher-generated questions and answers, then filters examples through agreement between two teacher models and the student model's current capabilities. Training progresses from object grounding and attribute recognition to relationships between objects, then to counting and detailed scene understanding. Each stage uses task-specific rewards so that local perceptual skills support later global reasoning.

From Cache to Belief: Online Probabilistic Memory for Long-Context Language Models

J. Wu, Y. Liu & Y. Yue^* et al.

BeliefMem studies how a language model can retain useful evidence beyond its local attention window while representing uncertainty in compressed history. It adds a lightweight probabilistic memory adapter to a frozen language model.

A fixed set of Bayesian memory slots is updated online. The model reads a query-dependent belief from those slots and combines it with local evidence through a product of Gaussian experts. Shared responsibilities govern both reading and writing, keeping retrieval aligned with memory updates. Experiments on six LongBench tasks evaluate long-context reasoning, while ablations and slot interventions examine whether gains come from distributed content updates rather than a larger local window.

You Only Show Once: Robust GUI Automation via One-Shot Human Demonstration

Y. Huang^#, Y. Liu^#, W. Yan et al.

YOSO-Agent uses a single human demonstration to guide GUI task execution. It preserves action-aligned screenshots and visual references, giving the agent concrete evidence about what to do and where to act when the interface differs from the demonstration.

A Task Manager checks progress against the reference trajectory; an Action Agent grounds the next interaction in the current screen; and a Reflection Agent diagnoses deviations and supports recovery. The grounding model is trained with varied visual appearances. Evaluation covers both target localization under interface perturbations and end-to-end task completion. The framework follows a supplied demonstration for the task; selecting the appropriate demonstration is outside its scope.

A Scalable Benchmark Test Suite for Dynamic Multi-Objective Optimization with a Changing Number of Objectives

K. Shang, Z. Xiao, Y. Liu et al.

Dynamic optimization benchmarks often change the objective functions themselves when the number of objectives changes. This makes it difficult to isolate an algorithm's ability to adapt to adding or removing evaluation criteria.

The benchmark defines a problem with a fixed maximum set of objectives, then activates different subsets over time. The underlying functions remain unchanged. Minus-DTLZ and Minus-WFG formulations avoid degenerate Pareto fronts that can arise when selecting subsets of conventional benchmark objectives. The resulting suite supports controlled comparisons of algorithms as objectives are added or removed.

Read the paper on arXiv →