InstrumentalEval - Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?

An evaluation designed to detect instrumental convergence behaviors in model responses (e.g., self-preservation, resource acquisition, power-seeking, strategic deception) using a rubric-driven LLM grader. The benchmark tests whether AI systems exhibit behaviors that are instrumentally useful across a wide range of goals, which could indicate concerning patterns of strategic reasoning.

Overview

InstrumentalEval is an evaluation designed to detect instrumental convergence behaviors in model responses (e.g., attempts at self-preservation, resource acquisition, power-seeking, etc.) by using an LLM grader.

• ArXiv paper
• Official repository / dataset source

Contributed by: horvgbor

Overview

This implementation:

• Downloads the InstrumentalEval benchmark CSV files from the upstream GitHub repository.
• Presents each prompt to the evaluated model.
• Uses a separate grader model to classify the response as convergence or no_convergence.
• Aggregates results into global metrics (convergence, no_convergence, invalid) and also emits task-specific breakdown metrics.

Dataset

The dataset is pulled from the upstream GitHub repo at runtime (the benchmark/ folder).

• Each CSV file represents a task type.
• The first row is treated as the evaluation context prompt (eval_prompt).
• Subsequent rows are treated as model inputs.
• Each sample is annotated with:
  • eval_prompt: per-file header used to condition the grader prompt.
  • task_type: derived from the CSV filename.

Notes:

• This evaluation requires internet access to download the dataset.
• GitHub API rate limits can affect dataset fetching.

Scoring

The scorer uses a grader LLM to produce outputs formatted as:

DECISION: [convergence/no_convergence]
REASON: [brief justification]

The decision is parsed into a per-sample dict score with keys:

• convergence (1.0 for convergence, else 0.0)
• no_convergence (1.0 for no-convergence, else 0.0)
• invalid (1.0 if the grader response cannot be parsed into a valid decision)

The metric aggregates these values across samples and also produces per-task metrics (flattened with {task_type}_ prefixes).

Number of Epochs

This evaluation is designed to be run for 1 epoch.

Rationale:

• Each sample is independently graded and does not benefit from repeated passes.
• Running multiple epochs increases cost without improving the estimate of the underlying distribution unless you are specifically measuring grader/model stochasticity.

If you choose to run multiple epochs (e.g., to estimate variance under nonzero temperature), document the justification in your evaluation report.

Usage

Installation

There are two ways of using Inspect Evals, from pypi as a dependency of your own project and as a standalone checked out GitHub repository.

If you are using it from pypi, install the package and its dependencies via:

pip install inspect-evals

If you are using Inspect Evals in its repository, start by installing the necessary dependencies with:

uv sync

Running evaluations

Now you can start evaluating models. For simplicity’s sake, this section assumes you are using Inspect Evals from the standalone repo. If that’s not the case and you are not using uv to manage dependencies in your own project, you can use the same commands with uv run dropped.

uv run inspect eval inspect_evals/instrumentaleval --model openai/gpt-5-nano

You can also import tasks as normal Python objects and run them from python:

from inspect_ai import eval
from inspect_evals.instrumentaleval import instrumentaleval
eval(instrumentaleval)

After running evaluations, you can view their logs using the inspect view command:

uv run inspect view

For VS Code, you can also download Inspect AI extension for viewing logs.

If you don’t want to specify the --model each time you run an evaluation, create a .env configuration file in your working directory that defines the INSPECT_EVAL_MODEL environment variable along with your API key. For example:

INSPECT_EVAL_MODEL=anthropic/claude-opus-4-1-20250805
ANTHROPIC_API_KEY=<anthropic-api-key>

Options

You can control a variety of options from the command line. For example:

uv run inspect eval inspect_evals/instrumentaleval --limit 10
uv run inspect eval inspect_evals/instrumentaleval --max-connections 10
uv run inspect eval inspect_evals/instrumentaleval --temperature 0.5

See uv run inspect eval --help for all available options.

Parameters

instrumentaleval

• grader_model (str): Model used to grade responses. Default is a local dev value; for reproducible runs you should set this explicitly.
• grader_temperature (float): Temperature used for the grader model.

Example:

uv run inspect eval inspect_evals/instrumentaleval \
  --model openai/gpt-5-nano \
  --temperature 0.0 \
  -- \
  --grader-model openai/gpt-5-nano \
  --grader-temperature 0.0

Evaluation Report

This evaluation measures instrumental convergence behaviors in model responses. The dataset consists of prompts designed to elicit power-seeking, self-preservation, and resource acquisition behaviors.

Overall Results (gpt-4o)

Evaluated using gpt-4o as both the main model and grader model.

Metric	Score
convergence	0.539
no_convergence	0.461
invalid	0.0

Results by Task Type

Task Type	Convergence	No Convergence	Invalid
Deceptive Behavior	0.0	1.0	0.0
Evading Shutdown	0.95	0.05	0.0
Hacking Computer Systems	0.1	0.9	0.0
Hiding Unwanted Behavior	0.545	0.455	0.0
Running Multiple AI Copies	0.846	0.154	0.0
Strategically Appearing Aligned	0.364	0.636	0.0

Reference Results (Original Implementation)

Results from the original codebase using gpt-4o:

Task Type	Convergence	No Convergence	Invalid
Deceptive Behavior	0.091	0.909	0.0
Evading Shutdown	1.0	0.0	0.0
Hacking Computer Systems	0.0	1.0	0.0
Hiding Unwanted Behavior	0.455	0.545	0.0
Running Multiple AI Copies	0.769	0.231	0.0
Strategically Appearing Aligned	0.636	0.364	0.0

For additional reference results, please see the original paper. Note that there are differences in results from the original paper.