SOAR: System Operations and Reasoning Benchmark

System Operations and Reasoning (SOAR) is a human-curated benchmark designed to evaluate the ability of large language models (LLMs) to interpret and answer questions over IT operational data, with a primary focus on Unix/Linux-style system administration contexts.
It focuses on how well models can interpret, extract, and reason over operational data — including logs, configurations, and system outputs — and on how question formulation and lightweight output-control instructions affect correctness, conciseness, and operational usability.

SOAR is designed around foundational operational micro-tasks that serve as cognitive building blocks for autonomous IT management.
Rather than testing long-horizon planning or multi-source retrieval, each sample is grounded in a single operational source (a CLI output, a log excerpt, a configuration file, or an OS-related document), isolating model behavior at the individual invocation level. This emphasizes foundational reasoning — the kind needed to:

• Parse heterogeneous operational data
• Ground responses in the semantics of IT contexts
• Handle verification, missing evidence, and contradiction
• Reason about structure and meaning within system artifacts

---

Overview

SOAR provides a cognitively structured evaluation space for assessing LLM performance on foundational operational tasks.
Each task presents a piece of realistic system data (e.g., YAML configurations, service status, logs, or command outputs) and asks the model to perform reasoning steps such as:

• Identifying system states or configuration values
• Verifying operational conditions or dependencies
• Inferring trends or behaviors from metrics
• Performing simple calculations or relational analyses

Answers are evaluated in a free-form plain-text format rather than through multiple-choice options, reflecting realistic operational environments where model outputs may be directly integrated into automated workflows. Precision and conciseness are therefore desirable operational properties: returning Kevin is preferable to The user who executed the command is Kevin.

The benchmark aims to answer a core question:

How effectively can LLMs understand and reason over operational data structures that system administrators interact with daily?

SOAR spans a diverse set of operational domains commonly encountered in system administration:

• System information and processes (125 questions): top, uptime, /proc directories, lsb_release, and related system-reporting utilities
• Virtualization and containers (103 questions): docker stats, docker inspect, docker logs, Dockerfile, Kubernetes configuration files, and related container-orchestration elements
• Package and dependency management (72 questions): dpkg, pip, uv, and requirements files, among other package-management artifacts
• Activity monitoring (59 questions): log-analysis tasks and other observability-related resources
• APIs and data formats (51 questions): JSON artifacts, OpenAPI specifications, and comparable structured data formats
• Networking and connectivity (49 questions): netstat, ss, and similar network introspection tools
• Services and daemons (41 questions): service, systemctl, and associated service-management commands

A sample

content	query1	request1	response	lines	description	format	reasoning type	bloom	DOK
apiVersion: v1\nkind: Pod\nmetadata:\n name: nginx\nspec:\n containers:\n - name: nginx\n image: nginx:1.14.2\n ports:\n - containerPort: 80	Which port will be assigned to the container?	Determine the port for the container.	80	20	kubernetes pod	yaml	Identification	Remembering	Recall

Dataset Composition

Each question is annotated according to two complementary cognitive perspectives:

• Bloom-inspired cognitive categories, to indicate the dominant cognitive operation required (from identification to inference)
• Depth of Knowledge (DOK), to capture the depth of reasoning needed to produce a correct answer

This multidimensional annotation enables a fine-grained analysis of how reasoning performance degrades as cognitive complexity increases.

From a statistical perspective, SOAR comprises 500 questions, each provided in two formulations — an explicit interrogative form (query1, e.g., "Is the TTY enabled?") and an imperative command-like form (request1, e.g., "Determine if TTY is enabled.") — resulting in 1,000 question instances in total. The question content was extracted from real data dumps and authentic configuration files, and all questions, answers, and cognitive labels were designed and annotated by the authors.

By cognitive reasoning type:

Type	Count
Identification	194
Verification	84
Simple Calculation	55
Relational Analysis	76
Inference	91

By Depth of Knowledge (DOK) level:

DOK	Count
Recall	367
Skill / Concept	74
Strategic Thinking	59

By content format:

Format	Count
Plain text	220
Tabular text	183
YAML	50
JSON	41
TOML	6

---

Evaluation

The accompanying study evaluates 16 open-weight and proprietary models, spanning dense and mixture-of-experts architectures from 4B to over 600B parameters, under four configurations per sample: two question styles (explicit vs. imperative) and two prompting conditions (no-prompt vs. instructions-prompt).

Key findings:

• Retrieval-oriented tasks (Identification) are generally easier, whereas Verification and Inference remain more challenging, particularly when the evidence is missing, indirect, or contradictory.
• Instruction-based prompting strongly reduces verbosity (average verbosity ratios drop from above 200 to around 4) and substantially improves exact matching.
• Some models still include correct answers within unnecessarily long outputs, highlighting the need for evaluation protocols that account not only for correctness but also for conciseness, evidence sensitivity, and practical usability in IT operational settings.

---

Citation

If you use SOAR in your research or evaluation, please cite:

Pending