Weekly AI Papers: AgentBaton Cuts API Costs by 65.9%, Criticisms of MCP Governance Monopoly

• AgentBaton reduced API costs by 65.9% and improved accuracy by up to 8.8%p when switching from high-performance models to efficient models.
• Analysis revealed that MCP and AAIF governance concentrates authority in a few companies, causing security debt and legitimacy issues.
• The 'Beyond the Output Layer' study mechanistically elucidated the geometric structure of multimodal reasoning by analyzing the model's internal latent space through causal intervention.
• In the InteractiveWorld benchmark, 12 latest MLLMs were found to fall short of reliable assistant level.

Agent Efficiency and Structuring

• Terminus-4B narrowed the performance gap compared to frontier models in sub-agent roles and reduced token usage by up to 30%.
• The AW-0 framework improved robotic manipulation performance compared to existing VLAs by integrating code planning and video prediction.
• The Trinity optimized agent system scalability and environment setup costs by restructuring tasks by capability units.

Governance and Transparency

• Criticized the structure where the AAIF technical committee is composed solely of 8 platinum companies, including AWS, Anthropic, and OpenAI.
• Attribution Gradients supports evidence tracing and critical review of AI answers through secondary citations.
• A tool selection model based on Qwen3-8B outperformed Llama-4-Maverick in tool environments with over 10,000 tools.