Papers
arxiv:2611.17460

Unfinished-Bench: Measuring Rule Violations by Persistent Agents in Realistic Work Environments

Published on Nov 12
ยท Submitted by
JAMEYUNG YU
on Nov 19
Authors:
,
,

Abstract

AI agents are increasingly equipped with persistence toward task completion to run delegated work autonomously for long periods, under many unstated constraints from internal documents and other people's authority. When completion requires violating a rule, the agent must decide alone, yet benchmarks of instruction-rule conflicts state the rule in the prompt or keep it hidden and do not evaluate persistence. We present Unfinished-Bench, a benchmark that gives models persistence in autonomous runs without a user, in realistic work environments with embedded rules. In each of its 97 scenarios across 7 organizations and 7 risk categories, the goal and the rule cannot both hold. We find that persistence raised violations of a rule in an internal document from 32.2% to 43.6%, five times the 8.5% for the same rule stated by the user. Of the document-rule violations, 98.3% occurred after the rule was displayed, and in 89% of a judged sample the agent first wrote down the rule. Violations of rules placed in the environment without user oversight can lead to serious incidents, so persistence-driven risks need to be evaluated. Benchmark documentation: https://francis21c.github.io/unfinished-bench/

Community

Paper submitter

TL;DR

Unfinished-Bench measures whether persistent agents break rules embedded in realistic work environments to finish their task. Persistence adds violations, and rules in documents are broken far more often than user-stated ones.

Overview

What happens when an agent becomes persistent and finds a workplace rule that blocks the way to its goal? Unfinished-Bench places agents in realistic work environments: 7 organizations with staff, internal documents, and email or chat. A user hands off a task and leaves, so the agent must decide alone. Its 97 scenarios cover 7 risk categories, from data leakage to destructive actions, and in each one the goal and a rule in the environment cannot both hold. We run each scenario without and with persistence, the drive to keep working until the task is done. Across 12 open models, persistence raised violations of rules in internal documents from 32.2% to 43.6%, and nearly all of these violations came after the rule was shown to the agent.

Project page: https://francis21c.github.io/unfinished-bench/
Code: https://github.com/naver-ai/unfinished-bench
alphaXiv: https://www.alphaxiv.org/abs/2611.17460
arXiv: https://arxiv.org/abs/2611.17460

Sign up or log in to comment

Get this paper in your agent:

hf papers read 2611.17460
Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2611.17460 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2611.17460 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2611.17460 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.