AI 논문 주간 TOP 10 — 2026-05-25

Top Research Papers of the Week

1. Towards End-to-End Automation of AI Research (Google DeepMind / Published in Nature)

   • Summary: Proposes an architecture for an AI system capable of managing the entire scientific research lifecycle—from hypothesis generation and experiment design to analysis and manuscript writing—and validates its limitations experimentally.
   • Significance: Since its publication in Nature (March 25, 2026), it has become a benchmark for debates on automated research, recently re-examined by The Conversation and others, sparking discussions on the ethics and reliability of AI-generated papers.

   

2. OpenAI o3 Proves Decades-Old Erdős Conjecture (OpenAI)

   • Summary: OpenAI's latest reasoning model, o3, successfully and independently proved a long-standing combinatorics conjecture originally proposed by mathematician Paul Erdős.
   • Significance: Described by New Scientist as "AI’s biggest breakthrough in mathematics," it marks the first time an AI has demonstrated independent proof-solving capabilities in pure mathematics beyond mere assistance. The Erdős conjecture had baffled mathematicians for decades.

   

3. HawkesLLM: Semantic Uncertainty Propagation in Agentic Text Simulation (Accepted at ICML 2026 Workshop)

   • Summary: Models how semantic uncertainty propagates in agent-based text simulations using Hawkes processes, offering a new framework for quantifying uncertainty in LLMs.
   • Significance: Accepted for the ICML 2026 workshop on "Statistical Uncertainty Frameworks for Agentic Systems," it has emerged as a key reference for evaluating the reliability of multi-agent AI systems.

4. Effect-Transparent Governance for AI Workflow Architectures: Semantic Preservation, Expressive Minimality, and Decidability Boundaries (Alan L. McCann)

   • Summary: Introduces a formal theory for AI workflow architectures that balances governance policies with semantic preservation and expressive minimality while maintaining decidability boundaries.
   • Significance: Strengthens the theoretical foundation for AI safety and regulation, providing a mathematical approach to auditability in enterprise AI systems. Currently trending at the top of arXiv cs.AI.

5. Accelerating AI Computing with Photonic-Matter Hybrid Particles (University of Pennsylvania / ScienceDaily)

   • Summary: Implements a new platform using polaritons (hybrid light-matter particles) to process AI calculations via optical circuits instead of electronic ones, achieving significantly lower energy consumption for AI inference.
   • Significance: Highlighted by ScienceDaily as a hardware breakthrough, it suggests a path toward solving AI’s energy crisis through optical computing, sparking interest with headlines like "Forget electrons."

   

6. AI for Scientific Discovery is a Social Problem (arXiv 2509.06580v4, updated)

   • Summary: Argues that AI techniques for scientific discovery—such as graph neural networks, hierarchical attention mechanisms, and physics-informed neural networks—are held back more by social and institutional barriers than by technical limitations.
   • Significance: Reframes the slow adoption of AI-based scientific tools as a social issue, shifting the focus for policymakers and research institutions. The v4 update was recently released.

7. Materials Science + ML: Accelerating Material Discovery (arXiv cs.LG)

   • Summary: Details a pipeline for applying machine learning to materials science (cond-mat.mtrl-sci) to predict and discover new functional materials, supported by a 21-page paper with nine figures.
   • Significance: A prime example of cross-disciplinary research showing how AI can drastically speed up the development of batteries, semiconductors, and biomaterials. References an accompanying theory paper (DOI: 10.5281/zenodo.19237451).

8. DeepSeek Visual Reasoning Breakthroughs (DeepSeek / analyzed by devFlokers)

   • Summary: DeepSeek unveiled a new architecture achieving groundbreaking performance in visual reasoning compared to existing models.
   • Significance: Evaluated in the May comprehensive analysis by devFlokers as a "key development defining a new era for agentic and physical AI." Improved visual understanding has immediate implications for autonomous driving and medical imaging.

9. Review of 2025's Top AI Research Papers: Reasoning Models, Autonomous Agents, and Reinforcement Learning (Analytics Vidhya)

   • Summary: Summarizes ten papers that defined AI research in 2025, categorized by advancements in reasoning models, autonomous agents, and reinforcement learning, with analysis on their impact on 2026.
   • Significance: A "must-read" retrospective for understanding the current trajectory of AI, widely shared within the community after appearing in Analytics Vidhya this week.

   

10. IEEE COMPSAC 2026 Security Workshop Accepted Paper: LLM-based Software Security Threat Analysis

    • Summary: Analyzes new attack vectors where LLMs are used to discover and exploit software vulnerabilities, proposing defensive mechanisms. Accepted for the STPSA workshop at IEEE COMPSAC 2026 (Madrid, Spain, July 7-10).
    • Significance: A proactive study on the risks LLMs pose to the software ecosystem, expected to serve as a foundation for corporate AI security policies.

7 sources

huggingface.co

huggingface.co

huggingface.co

huggingface.co

sciencedaily.com

sciencedaily.com

newscientist.com

newscientist.com

analyticsvidhya.com

analyticsvidhya.com

nature.com

nature.com

sciencedaily.com

sciencedaily.com

Research Trends and Technical Analysis

The core technological themes connecting this week’s top ten papers are as follows:

1. The Real-World Progress of "AI-as-Scientist"

The Nature paper on "End-to-End Automation" and OpenAI o3’s proof of the Erdős conjecture show that AI is crossing the threshold from a tool to an independent scientific agent. As The Conversation noted, AI now generates hypotheses and proves theorems, forcing a fundamental redesign of peer review systems and research ethics.

2. Managing Uncertainty in Agentic AI Systems

HawkesLLM, the governance architecture study, and the security analysis all address the uncertainty and governance challenges inherent in autonomous agents. The inclusion of statistical uncertainty as a dedicated theme at ICML 2026 signals a growing community consensus on the urgency of this issue, shifting AI safety from theory to practical architectural design.

3. Convergence of AI Hardware and Materials Science

Research from the University of Pennsylvania on photonics and the integration of ML into materials science demonstrates that hardware innovation is becoming a central strategy for addressing AI’s energy consumption. The shift from electronic to optical circuits holds the potential to reduce inference costs and carbon footprints, with these papers increasingly bridging the gap between cs.LG and cond-mat.

Upcoming Research to Watch

1. IEEE COMPSAC 2026 STPSA Workshop (Madrid, Spain, July 7-10)
   The paper on LLM security threats will be officially presented. Industry feedback and potential supplemental experimental data are highly anticipated.

2. Full Paper Releases from ICML 2026 "Statistical Uncertainty Frameworks for Agentic Systems" Workshop
   The full list of accepted papers, including HawkesLLM, and the presentation schedule will be released, providing a comprehensive view of the latest directions in agentic AI uncertainty.

3. Forthcoming Policy Research on AI for Scientific Discovery
   Based on the social and institutional barriers identified in the updated "AI for Scientific Discovery is a Social Problem" paper (v4), it is reported that foundations and research institutions are preparing policy proposal documents to govern the acceleration of AI-based scientific discovery.