StratEval

Comprehensive eval suite for Instrumental convergence

Project Details

StratEval v1 is an evaluation suite of 432 instrumentally convergent scenarios, two patterns of Socratic decomposition probes, and three modalities to control against early lock-in stateful, stateless, and hybrid. A full-run of StratEval creates more than 20,000 individual datapoints, all aimed at measuring instrumental convergence.

Adrian Hau is the principal investigator, and a pre-print of StratEval v1 is available. https://zenodo.org/records/20419741

StratEval v2 has more scenarios, but requires further validation as weaknesses have been identified in StratEval v1. Namely, StratEval v1 underweighted cooperation as an instrumentally convergent behavior, and the schema used to evaluate the relative risk needs to be validated and refactored. Additionally it would be beneficial to study more models, especially frontier models (~70B parameters).

Theory of Impact

Instrumental convergence is the phenomena wherein an intelligent system given a broad goal may develop sub-goals in pursuit of the larger goal. Hypothesized and arguably demonstrated subgoals include, but are not limited to;

Resource acquisition
Deception
Misdirection
Power Seeking
Bribery
Blackmail

This connects directly to AI safety risk. The move into agentic AI has been shown to do exactly this. During OpenAI's cybersecurity evaluations, a combination of models, including GPT-5.6 Sol, and a more capable model, circumvented sandbox controls, obtained internet access, and compromised Hugging Face systems while pursuing benchmark solutions. OpenAI later said the incident was primarily driven by the internal research model. Nevertheless, this demonstrates a real world incident of instrumental convergence. I dont' believe it's correct to say that instrumental convergence remains a pure hypothetical. It's critical that we understand the phenomena better.

The StratEval project was created with a single question, if instrumental convergence is a problem, then how prevalent is it, and how bad could things get when an AI forms said subgoals? The results of StratEval v1 show the problem is significant, and expose that alignment as such is nowhere near a solved problem; in fact, alignment interventions appear at best surface level. Ultimately, in order to mitigate against catastrophic AI risk we need better data; StratEval provides much better data.

People

Adrian Hau

Team Member

Funding Details

Start Date: -
End Date: -
Expected Duration: 6 months
Funding Raised to Date: $750
Annual Budget: -
Monthly Burn Rate: -
Current Runway: -
Funding Goal: -
Funding Stage: -
Fiscal Sponsor: -

Track Record

BlueDot Impact approved a microgrant of $750, but fiscal constraint has prevented moving forward.

Grants Received– no grants recorded

Add grant

Funding Asks

Funding ask

Open

Minimum

$20,000

Ideal

$50,000

Theory of impact

Resource acquisition
Deception
Sandbagging
Misdirection
Power Seeking
Bribery
Blackmail

All of the above were demonstrated in StratEval v1.

This connects directly to AI safety risk. The move into agentic AI has been shown to do exactly this; when OpenAI was testing GPT-6 Astra, Astra breached its sandbox environment in order to obtain data from Hugging Face, said data was to help it achieve a better score on a benchmark.

The StratEval project was created with a single question, if instrumental convergence is a problem, then how prevalent is it, and how bad could things get when an AI forms said subgoals? The results of StratEval v1 show the problem is significant, and expose that alignment as such is nowhere near a solved problem; in fact, alignment interventions appear at best surface level.

Ultimately, in order to mitigate against catastrophic AI risk we need better data; StratEval provides much better data.

How the money will be spent

Approximately:

$3,000 — model API and GPU costs for replication on additional open and closed models.
$3,000 — independent human adjudication of a stratified subset of outputs, including ambiguous and high-escalation cases.
$12,000 — researcher time for experimental execution, statistical analysis and robustness testing.
$2,000 — reproducibility, archival data preparation and publication.

At the $50,000 ideal level, additional funding would expand frontier-model coverage, increase independent human adjudication and replication, support larger robustness and sensitivity analyses, and provide additional researcher time for validation and public artifact preparation.

Discussion

No comments yet. Be the first to share your thoughts.