---
title: "PredAct-Bench: Benchmarking Tool-Augmented Dialogue under Controlled Tool Noise"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.02372v1"
permalink: "https://feed.astralor.com/items/itm_5cc395bebc6f0c3eaf5c30ef"
published: "2026-08-03T15:15:55.000Z"
score: 75
channel: "tech"
tags: ["PredAct-Bench","Trust Calibration","Large Language Models","Tool Noise","UK Open University"]
---

# PredAct-Bench: Benchmarking Tool-Augmented Dialogue under Controlled Tool Noise

## AI commentary

> Tool noise is the norm in real deployments, yet most benchmarks evaluate LLMs in a sterile environment of perfect tool outputs. This paper advances trust calibration from single decisions to multi-turn dialogue, and even demonstrates with human studies that teachers over-rely on AI suggestions — a must-read for edtech and health AI product teams. Rather than chasing higher accuracy, RAIR/RSR-style metrics may be the better way to decide when the system should hand control back to humans.

## At a glance

Large language models are increasingly deployed in task-oriented dialogue systems for high-stakes domains such as education, healthcare, and finance, yet existing benchmarks typically assume perfectly accurate tool outputs. This paper introduces PredAct-Bench, a benchmark for evaluating tool-augmented dialogue under controlled tool noise, using education as a measurable testbed. The authors propose episode-level Relative AI-Reliance (RAIR) and Relative Self-Reliance (RSR) metrics, extending trust calibration to multi-turn dialogue. They evaluate 13 state-of-the-art closed and open-source LLMs on two educational datasets, OULAD and PREDACT-CS, alongside a human study with instructors and teaching assistants. The results show that when tools are noisy, current SOTA models fail to provide sufficient visibility to teachers, leading to over-reliance on wrong suggestions or hallucinations. The benchmark aims to support building better LLMs as AI decision-support systems for educators.

## Key points

- Introduces PredAct-Bench, a benchmark for evaluating tool-augmented dialogue under controlled tool noise, addressing the gap left by benchmarks assuming perfect tool outputs.
- Proposes RAIR and RSR metrics that extend trust calibration from single decisions to multi-turn dialogue, measuring users' relative reliance on AI versus themselves.
- Evaluates 13 state-of-the-art closed and open-source LLMs on OULAD and PREDACT-CS, complemented by a human study with instructors and teaching assistants.
- Finds that current models fail to provide sufficient visibility under tool noise, leading educators to over-rely on wrong suggestions or hallucinations.

## Full text

Large Language Models (LLMs) are increasingly deployed in task-oriented dialogue systems that support multi-step decision-making in high-stakes domains such as education, healthcare, and finance. However, existing benchmarks typically assume perfectly accurate tool outputs, overlooking the reality that deployed systems must operate with noisy tools and human decision-makers whose trust in the agent is itself uncertain. Such conditions are common in practice, for example, a clinician using a diagnostic prediction tool or an advisor relying on a model that forecasts student outcomes from historical records. We introduce PREDACTBENCH, a benchmark for evaluating dialogue agents paired with statistically imperfect tools, using education as a measurable testbed where ground truth outcomes and clear intervention decisions are available. First, we build a benchmark for AI-assisted human decision-making, where the AI uses noisy predictors to help guide a user. Second, we introduce episode-level Relative AI-Reliance (RAIR) and Relative self-reliance (RSR) metrics, extending prior trust calibration framework to multi-turn dialogue. Third, we evaluate 13 state-of-the-art closed and open source LLMs on two educational datasets, OULAD (real assessment trajectories from the UK Open University) and PREDACT-CS (60 courses with real final grade outcomes and synthetically generated weekly score trajectories), alongside a human study with instructors and teaching assistants. We find that when tools are noisy, SOTA models are supposed to provide visibility to teachers so that they do not over-rely on wrong suggestions or hallucinations, but current models fail to do that. We offer PREDACTBENCH to help build better LLMs as AI decision support systems to help teachers.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_5cc395bebc6f0c3eaf5c30ef