Drift-Bench: Diagnosing Cooperative Breakdowns in LLM Agents under Input Faults via Multi-Turn Interaction

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Bao, Han, Zhang, Zheyuan, Jing, Pengcheng, Yuan, Zhengqing, Shi, Kaiwen, Ye, Yanfang
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866917242430554112
author Bao, Han
Zhang, Zheyuan
Jing, Pengcheng
Yuan, Zhengqing
Shi, Kaiwen
Ye, Yanfang
author_facet Bao, Han
Zhang, Zheyuan
Jing, Pengcheng
Yuan, Zhengqing
Shi, Kaiwen
Ye, Yanfang
contents As Large Language Models transition to autonomous agents, user inputs frequently violate cooperative assumptions (e.g., implicit intent, missing parameters, false presuppositions, or ambiguous expressions), creating execution risks that text-only evaluations do not capture. Existing benchmarks typically assume well-specified instructions or restrict evaluation to text-only, single-turn clarification, and thus do not measure multi-turn disambiguation under grounded execution risk. We introduce \textbf{Drift-Bench}, the first diagnostic benchmark that evaluates agentic pragmatics under input faults through multi-turn clarification across state-oriented and service-oriented execution environments. Grounded in classical theories of communication, \textbf{Drift-Bench} provides a unified taxonomy of cooperative breakdowns and employs a persona-driven user simulator with the \textbf{Rise} evaluation protocol. Experiments show substantial performance drops under these faults, with clarification effectiveness varying across user personas and fault types. \MethodName bridges clarification research and agent safety evaluation, enabling systematic diagnosis of failures that can lead to unsafe executions.
format Preprint
id arxiv_https___arxiv_org_abs_2602_02455
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Drift-Bench: Diagnosing Cooperative Breakdowns in LLM Agents under Input Faults via Multi-Turn Interaction
Bao, Han
Zhang, Zheyuan
Jing, Pengcheng
Yuan, Zhengqing
Shi, Kaiwen
Ye, Yanfang
Artificial Intelligence
Computation and Language
Software Engineering
As Large Language Models transition to autonomous agents, user inputs frequently violate cooperative assumptions (e.g., implicit intent, missing parameters, false presuppositions, or ambiguous expressions), creating execution risks that text-only evaluations do not capture. Existing benchmarks typically assume well-specified instructions or restrict evaluation to text-only, single-turn clarification, and thus do not measure multi-turn disambiguation under grounded execution risk. We introduce \textbf{Drift-Bench}, the first diagnostic benchmark that evaluates agentic pragmatics under input faults through multi-turn clarification across state-oriented and service-oriented execution environments. Grounded in classical theories of communication, \textbf{Drift-Bench} provides a unified taxonomy of cooperative breakdowns and employs a persona-driven user simulator with the \textbf{Rise} evaluation protocol. Experiments show substantial performance drops under these faults, with clarification effectiveness varying across user personas and fault types. \MethodName bridges clarification research and agent safety evaluation, enabling systematic diagnosis of failures that can lead to unsafe executions.
title Drift-Bench: Diagnosing Cooperative Breakdowns in LLM Agents under Input Faults via Multi-Turn Interaction
topic Artificial Intelligence
Computation and Language
Software Engineering
url https://arxiv.org/abs/2602.02455