The TAS framework constrains tail risk in multi-step Agent workflows
Start here
An arXiv paper proposes TAS, a tail-risk-aware agent safety framework that models the safety of multi-step LLM agent workflows as sequences of interdependent safety states, using a shared budget to constrain rare but costly tail risks. The framework first models safety state transitions as an MDP to track risk propagation across steps, then measures tail-sensitive risk and derives step-by-step safety policies. Experiments show that TAS outperforms eleven baselines on three benchmarks, suppressing rare high-cost failures while preserving usability in normal scenarios.
Generated by AI from reports · Updated 2 days ago
Attention and developments
Attention, hourly
Now 24, -2% in 24 hours·Peak 26 (Nov 10 22:00)
Related research
This event's paper
- Tail-risk-aware Agent Safety: Constraining Tail Risks Across Multi-step LLM Agent Workflows
arXiv · 2026-11-07 · Deep reading
Progress timeline
- arXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesTail Risk-Aware Agent Safety (TAS): Constraining Tail Risk in Multi-Step LLM Agent Workflows
The Tail Risk-Aware Agent Safety framework TAS models the safety of multi-step LLM agent workflows as a sequence of interdependent safety states, using a shared budget to constrain rare but costly tail risks. It first models safety state transitions as an MDP to track risk propagation across steps, then measures tail-sensitive risk and derives a step-wise safety policy. Experiments show that TAS outperforms eleven baselines on three benchmarks, suppressing rare high-cost failures while preserving usability in normal scenarios.
Related discussion
Attention trend
New sources per day
- Nov 10: 1 new sources (1 outlets, 0 accounts)
- Nov 11: 0 new sources (0 outlets, 0 accounts)
- Nov 12: 0 new sources (0 outlets, 0 accounts)
- Nov 13: 0 new sources (0 outlets, 0 accounts)