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

TL;DR

The paper addresses the gap in evaluating LLM driven agents under flawed user inputs during multi turn interactions, where existing benchmarks assume Oracle and overlook grounded execution risk.It introduces Drift-Bench, a theoretically grounded taxonomy of four input faults Flaw of Intention Premise Parameter and Expression, a data perturbation pipeline, a persona driven user simulator, and the Rise evaluation protocol.Experiments across state oriented and service oriented environments with multiple frontier models reveal substantial degradation under faults and a Clarification Paradox where clarifications aid white box but hinder black box settings, quantified via $ ext{PD}$ and $ ext{G}$ metrics.The work provides a systematic framework for diagnosing and mitigating cooperative breakdowns, highlighting the need for risk aware and environment sensitive clarification policies to improve safety and reliability in autonomous agents.

Abstract

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.

Figures (40)

• Figure 1: Cooperative Breakdown Taxonomy. The diagram organizes systematic cooperative breakdowns into four high-level categories used throughout this paper: Flaw of Intention, Flaw of Premise, Flaw of Parameter, and Flaw of Expression.
• Figure 2: Pipeline overview of Drift-Bench . Left — Data perturbation: we start from verified tasks in state- and service-oriented environments, extract semantic frames, and generate controlled input faults (flaw of intention, parameter, premise, expression) to produce solvable, diagnostically informative variants. Center — Agent–user interaction: a persona-driven LLM-as-user simulates diverse behaviours while the agent may apply structured clarification actions in multi-turn dialogues to repair cooperative breakdowns. Right — Evaluation: interactions are scored by the Rise protocol, linking clarification behaviour to downstream safety and task effectiveness.
• Figure 3: Clarification Gains ($\mathcal{G}$) on Service-oriented task.
• Figure 4: $\mathcal{SAR}$ by model and fault type for State-Oriented tasks.
• Figure 5: Human Evaluation Screenshot.