EMNLP 2026 tutorial systematizes agentic AI failure modes, from triggers to fixes
ManlingLi_ · x · 2026-10-28
Manling Li (Northwestern) and colleagues will give a 3.5-hour EMNLP 2026 tutorial (Budapest, Oct 28) titled Failure Modes in Agentic AI.
- Problem: agents are deployed faster than the community can characterize failures — error cascades, reward hacking, reasoning collapse, memory drift, over-reasoning loops lack shared vocabulary, reproducible triggers, and comparable diagnostics.
- Four axes: environments & infrastructure (tool-call protocol, RL systems, token masking, re-tokenization drift, environment errors scored as policy failures, harness as attack surface), training dynamics, test-time scaling & adaptation, evaluation & red teaming, with safety threaded throughout.
- Presenters come from Amazon, Google, Northwestern and UIUC; attendees get reproducible triggers, diagnostic protocols and verifiable mitigations. Materials are already live on GitHub.
Related event: EMNLP 2026 tutorial dissects failure modes in agentic AI(4 posts)→
More from coding & agent
- Matt Pocock shows his /pr skill with a PR that fixes mid-stream edit loss in Article Writer — mattpocockuk · 2026-10-28
- Running SEO entirely from Claude: a 25-prompt workflow with Search Console data — ayushtweetshere · 2026-10-28
- Amp's dev culture: prototype with prompts first, ship only what survives — jobergum · 2026-10-28
- vapache author: parity with Apache2 ACL is possible in principle, or ship a subset — jd_pressman · 2026-10-28
- Fuzzing an AI-written Apache clone: author admits he has no idea when the bug flood ends — jd_pressman · 2026-10-28
- Dumping voice memos to my AI chief of staff rewired my workflow — RachelVT42 · 2026-10-28