ICML 2026 FAGEN Workshop Spotlights AI Agent Failure Modes

The ICML 2026 "Failure Modes in Agentic AI" (FAGEN) workshop is currently underway, focusing on the failure modes, security risks, and repair strategies of LLM agent systems. According to on-site information, the workshop attracted 9 speakers and 218 submissions, discussing the risk surfaces of agents in memory, tools, private data, and real-world actions.

Collaboration Failures and Training Bottlenecks

In terms of multi-agent collaboration, research indicates that simply increasing the number of agents does not lead to better teamwork. Instead, it causes communication bottlenecks, redundant labor, error aggregation, and even "false completion claims." For single-agent training, a CMU report pointed out that traditional scalar rewards in the MaxRL framework trigger specific failure modes. They suggested improving this with rich text feedback and self-distillation, and stabilizing training by normalizing updates according to task difficulty.

Decision Control and Safety Attribution

Regarding decision and action control, researchers explored agent logic under uncertainty and cost, such as when to retrieve, ask questions, or stop. The contributed oral session showcased three complementary studies: D-CEM proposed using loss-aware deliberation control to avoid unsafe consensus; Who&When Pro built a large-scale benchmark for multimodal failure attribution; and ATLAS focused on capability-targeted agentic training (TRACE). Additionally, Maarten Sap analyzed the communication of uncertainty in structured social worlds from the perspective of social intelligence.

2026-07-10 ~ 2026-07-10 · 11 related posts

• [source] ICML Workshop on Agent Failure Modes Held — wzenus · 2026-07-10
• ICML 2026 Workshop on Agentic AI Failures — wzenus · 2026-07-10
• FAGEN Workshop to Discuss Agent Failures and Training — wzenus · 2026-07-10
• FAGEN 2026 Focuses on Agentic Failure Modes — wzenus · 2026-07-10
• ICML Workshop Explores Agent Decision-Making and Social Intelligence — wzenus · 2026-07-10
• Reward Improvements in LLM-Agent Training — wzenus · 2026-07-10
• [source] New Methods for LLM Agent Training and Decision-Making — wzenus · 2026-07-10
• [source] ICML 2026 Spotlights Research on Agent Failures — wzenus · 2026-07-10
• Study on Multi-Agent Collaboration Failure Modes — wzenus · 2026-07-10
• Highlights from the Agent Safety Workshop — wzenus · 2026-07-10
• ICML Spotlights Three Studies on Agent Failures — wzenus · 2026-07-10