AI 논문 주간 TOP 10 — 2026-05-28

Top Papers of the Week

1. SkillOpt: Executive Strategy for Self-Evolving Agent Skills (Microsoft Research)

   • Highlights: A text-space optimization system that learns agent skills in an external state, enabling stable updates with zero deployment overhead.
   • Significance: Solves scalability issues for complex agents and proves real-world viability by hitting high marks across various benchmarks.

2. MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing (OpenDataLab)

   • Highlights: A 1.2B parameter vision-language model using a coarse-to-fine strategy to achieve state-of-the-art parsing accuracy.
   • Significance: Drastically improves computational efficiency for document understanding, making it much easier to deploy in real-world scenarios.

3. TriSplat: Simulation-Ready Feed-Forward 3D Scene Reconstruction (Zhejiang University)

   • Highlights: A feed-forward network using oriented triangle primitives to generate simulation-ready meshes directly from a single image.
   • Significance: Cuts out the post-processing stage for 3D reconstruction, streamlining the pipeline for real-time applications.

4. Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super Resolution (Chinese Academy of Science)

   • Highlights: Uses Riemannian geometry and adversarial training to fix spectral alignment issues in image super-resolution.
   • Significance: Sets a new bar for high-quality super-resolution by boosting structural fidelity and reducing artifacts.

5. Beyond Mode Collapse: Distribution Matching for Diverse Reasoning (Intern Large Models)

   • Highlights: Tackles mode collapse in on-policy reinforcement learning via forward KL minimization to maintain output diversity.
   • Significance: Improves performance in combinatorial optimization and reasoning tasks, addressing stability issues in RL-based agents.

6. TradingAgents: Multi-Agents LLM Financial Trading Framework

   • Highlights: A multi-agent framework using LLMs to simulate a real-world financial trading firm.
   • Significance: Demonstrates the practical application of LLM agents by hitting real-world metrics like cumulative returns and the Sharpe ratio.

7. Kronos: A Foundation Model for the Language of Financial Markets

   • Highlights: A specialized pre-training framework for financial K-line data, featuring a unique tokenizer and autoregressive pre-training.
   • Significance: Outperforms existing models in prediction and synthetic data generation, highlighting the potential of domain-specific foundation models.

8. OpenDevin: An Open Platform for AI Software Developers as Generalist Agents

   • Highlights: A development platform for AI agents that interact with the world through coding, command lines, and web browsing.
   • Significance: Sets a new standard for open-source AI agent research by providing multi-agent support and evaluation benchmarks.

9. Instance-Aware Parameter Configuration in Bilevel Late Acceptance Hill Climbing (Featured at IEEE CEC 2026)

   • Highlights: Proposes automated, instance-specific parameter configuration for bilevel optimization in capacitated vehicle routing problems.
   • Significance: Advances the practical application of automated algorithm design by improving efficiency in solving combinatorial optimization problems.

10. ERP-Bench: Enterprise Resource Planning System Benchmark

    • Highlights: Evaluates how well state-of-the-art models handle ERP system tasks, measuring success rates based on explicit constraints and optimal solutions.
    • Significance: Provides a fresh, systematic way to gauge how ready AI models are for real enterprise work.

Research Trends & Technical Analysis

1. Self-Evolution and Scalability in Agent Tech Papers like SkillOpt, OpenDevin, and DMPO show that AI agents are increasingly focused on self-optimization and expanding their own capabilities. The emphasis is on cutting deployment overhead while keeping training stable, which is a huge win for production-grade AI.

2. The Rise of Multimodal and Domain-Specific Foundation Models From MinerU2.5’s document parsing to Kronos for financial markets and TriSplat for 3D, it’s clear that we’re moving beyond general-purpose models toward vision-language-action models optimized for specific domains to balance efficiency and performance.

3. Advanced RL and Optimization Innovations like DMPO’s approach to mode collapse, new bilevel optimization methods, and super-resolution via Riemannian geometry are laser-focused on making training more stable and helping models converge better—addressing the nitty-gritty headaches of large-scale model training.

Upcoming Research to Watch

1. CVPR 2026 (Coming in June) With a record-breaking 16,000+ submissions, CVPR 2026 is set to unveil the latest breakthroughs in computer vision, embodied AI, and multimodal research, especially in 3D reconstruction and real-time processing.

2. arXiv Monthly Highlights (Early June) We’re tracking the cs.AI and cs.LG sections of arXiv, where daily submissions are booming. Expect a deep dive in early June on the most impactful papers of May and the emerging dominance of agent-based and domain-specific models.

3. Automated Scientific Research Platforms Keep an eye out for follow-ups on the "End-to-End AI Research Automation" paper recently published in Nature; we expect to see more real-world implementations that could automate everything from hypothesis generation to verification.