Back to feed
arXiv cs.CL
arXiv cs.CL
7/23/2026
Stateful Guardrails for Multi-Turn LLM Systems: A Conversational Risk Accumulation Framework

Stateful Guardrails for Multi-Turn LLM Systems: A Conversational Risk Accumulation Framework

Short summary

This paper introduces Conversational Risk Accumulation (CRA), a framework for detecting safety failures in multi-turn LLM dialogues that only emerge over multiple benign turns. The authors propose a session-layer system tracking semantic drift, sensitivity-weighted information accumulation, and compliance-gradient signals, along with a learned trajectory model (CRA-Net DA) and a benchmark suite (CRA-Bench) of 1,200-2,000 eight-turn sessions. Results focus on within-distribution session scoring and human-transfer validation.

  • •Introduces Conversational Risk Accumulation (CRA) framework for multi-turn LLM safety failures
  • •Proposes session-layer tracking of semantic drift, sensitivity accumulation, and compliance gradients
  • •Releases CRA-Bench with 1,200-2,000 eight-turn sessions across multiple threat families

Generated with AI, which can make mistakes.

#research-breakthrough#ai-agents#ai-tools
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

CogniConsole: Externalizing Inference-Time Control as a Formal Abstraction for Reliable LLM Interactions
arXiv CS.AI
Adaptive Capitulation: A Structural Failure Mode of LLM Responses in Vulnerability Contexts
arXiv cs.CL
How RLHF Reward Optimization Produces Manipulative Defenses in LLMs
Dev.to
The original title is "Some Large Language Models Exhibit Consistent Risk Attitudes"
arXiv CS.AI