Failure Taxonomies and Mechanisms
Operational definitions, triggering preconditions, minimal reproductions, composable failure primitives, and falsifiable mechanistic hypotheses for agent failures.
ICML 2026 Workshop
Failure Modes in Agentic AI (FMAI): Reproducible Triggers, Trace Diagnostics, and Verified Fixes
Foundation-model agents are now deployed in browsing, scientific analysis, and long-horizon decision-making. FMAI creates a focused venue for turning agent failures into concrete research assets: clear definitions, reproducible triggers, comparable diagnostics, and verified repair strategies.
Requires an OpenReview account.
Focus
Reproducible triggers, trace diagnostics, and verified fixes for agentic failures.
Format
Keynotes, contributed spotlights, posters, and a panel on practical agent failure modes.
Status
OpenReview submissions are now open. Deadline: May 4, 2026, 11:59 PM UTC.
Overview
Why This Workshop
Agent quality is governed by long-horizon interaction. Small stepwise mistakes can compound through tool calls, memory writes, and recovery decisions, shifting the reliability and safety boundaries of the whole system.
FMAI focuses on actionable failure modes in closed-loop agent systems. We treat failures as first-class research objects and push toward four concrete deliverables: operational definitions, reproducible triggers, comparable diagnostics, and verifiable fixes.
We especially welcome work that turns vague failure anecdotes into reusable scientific artifacts that other researchers can reproduce, measure, stress-test, and improve upon.
Current status
OpenReview submissions are now open for FMAI. See the deadline below and submit through the official venue page.
Submission deadline
May 4, 2026, 11:59 PM UTC
The accepted proposal plans a fully in-person workshop with keynote talks, contributed spotlights, posters, and a panel discussion. Livestreaming and fallback remote support may be arranged for exceptional cases.
Workshop date, room, and camera-ready timeline are still to be published through the official ICML 2026 workshop schedule.
Topics of Interest
What Fits FMAI
Closed-loop Evaluation and Trace Diagnostics
Long-horizon or open-world evaluation protocols, interpretable process metrics, counterfactual tests, and logging tools that expose failures beyond terminal success.
Training and Systems Interventions
Mitigations, recovery strategies, tool and memory interface improvements, reward and budget design, and repair mechanisms with clearly verifiable trade-offs.
- OpenReview submissions are now open for FMAI.
- Submission deadline: May 4, 2026, 11:59 PM UTC (May 4, 2026, 6:59 PM CDT).
- Submit through the official OpenReview venue page. An OpenReview account is required.
- High-quality negative results are encouraged when they include controlled failure cases, clear attribution, and transferable lessons for the field.
Invited Speakers
Invited Speakers
The proposed program covers failure mechanisms, diagnostics, evaluation, security, and practical deployment.
Yoshua Bengio
Universite de Montreal, LawZero, and Mila
AI safety, frontier model governance, and deep learning foundationsYoshua Bengio is a Full Professor of Computer Science at Université de Montréal, Founder of Mila (the Quebec AI Institute), and Co-President of LawZero. A recipient of the 2018 Turing Award alongside Geoffrey Hinton and Yann LeCun, his research spans deep learning foundations and, increasingly, AI safety and the governance of frontier systems.
GFlowNetsInternational AI Safety Report
James Zou is an Associate Professor of Biomedical Data Science, Computer Science, and Electrical Engineering at Stanford. His work focuses on making AI more reliable, human-compatible, and statistically rigorous, with major applications in health and biomedicine.
AvaTaRAutoRedTeamer
Bo Li
University of Illinois Urbana-Champaign and Virtue AI
Agent attack surfaces and tool-chain exploitsBo Li is the Wexler AI Scholar and an Associate Professor at UIUC, where she works on trustworthy machine learning, AI safety, security, privacy, and robustness. She is also the founder and CEO of Virtue AI.
ShieldAgentAutoRedTeamer
Logistics
Workshop Logistics
Primary contact
fmaiworkshop@gmail.com
Current sponsor
Sponsors currently listed for the public site are Abaka AI and O2 Lab. The accepted proposal also notes at least $1,000 of Abaka AI support for student support and awards.
Submission status
OpenReview submissions are now open. Deadline: May 4, 2026, 11:59 PM UTC (May 4, 2026, 6:59 PM CDT).
Website
Public workshop site: fmai-workshop.github.io
Contact
Stay in Touch
We are using this site as the working home for the FMAI workshop. If you want to ask about submissions, sponsorship, or program details, reach out directly.