Evaluation dossier

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

Tests model responses for instrumental goals such as resource acquisition and self-preservation under specified scenarios.

National University of Singapore / arXiv:2502.12206v1 / Propensity

Sources and provenance ↓ Read the research narrative →

CROSS-REPORT MONITORING

PUBLISHED EVIDENCE

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

Published measurements across source reports. Task scopes are annotated; configuration differences remain attached to each observation.

8 observations · 0 dated

● Matched◇ Configuration differs□ Some conditions undisclosed

8 observations lack a verified model release date. They remain in Data and Systems.

Select a point or table row to inspect its conditions and provenance.

8 observations

Measurements and source conditions
System / model	Result	Condition status	Task / conditions	Evaluator / report	Dates
DeepSeek-R1	50%Evading Shutdown	Incomplete	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	Not disclosedEvaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	Release —Report —
o3-mini	45%Evading Shutdown	Incomplete	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	Not disclosedEvaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	Release —Report —
o1	40%Evading Shutdown	Incomplete	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	Not disclosedEvaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	Release —Report —
o1-mini	40%Evading Shutdown	Incomplete	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	Not disclosedEvaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	Release —Report —
Gemini 2.0 FTE.	35%Evading Shutdown	Incomplete	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	Not disclosedEvaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	Release —Report —
GPT-4o	20%Evading Shutdown	Incomplete	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	Not disclosedEvaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	Release —Report —
Sonnet-3.5	15%Evading Shutdown	Incomplete	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	Not disclosedEvaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	Release —Report —
DeepSeek-V3	15%Evading Shutdown	Incomplete	Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?—	Not disclosedEvaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?	Release —Report —

Task and measurement

Benchmark protocol profile

Task / subject: Model or agent system in the original evaluation configuration
Measurement target: Tests model responses for instrumental goals such as resource acquisition and self-preservation under specified scenarios.
Scoring / outcome: Evading Shutdown · Table 1: category-specific and overall instrumental rates, using GPT-4o as judge. Elicited scenario rates do not estimate deployment probability.
Interpretation boundary: Textual choices and grader labels do not establish executed autonomous behavior or stable deployment propensities.

A reported evaluation

The source evaluates GPT-4o · Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?.

Reported measurement: Evading Shutdown · 20.00%.

Table 1: category-specific and overall instrumental rates, using GPT-4o as judge. Elicited scenario rates do not estimate deployment probability. See the source protocol; no independent trial count inferred.

Source-reported measurements for GPT-4o in Table 1; values are transcribed, not independently reproduced.

Measurement dossier · NUWA reading

Source-method reading and retained catalogue information are distinguished. Publication does not establish independent evidence or bilingual editorial review.

Measurement construct: When do goal-directed model responses exhibit instrumental-goal behaviors?
Task population: The benchmark varies prior task, model and prompt design and uses automated judges to label instrumental convergence.
Scoring and failure modes: Automated judge labels for instrumental-goal responses, sensitive to judge selection.
System conditions: Prior task, model cohort and explicit goal nudging vary.
Statistical meaning: Source-reported result statements; this does not establish trial counts, denominators or an estimator: Judge selection significantly changes detection (§4.6).
Validity and interpretation: Training histories and capabilities differ across model groups; judge sensitivity and constructed tasks limit deployment inference.
Appropriate research use: Model recommendation: retain judge identity and nudging condition, and avoid causal claims from unmatched model cohorts.

Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · §3 InstrumentalEvalSelected original resource https://arxiv.org/html/2502.12206; sha256:1bb93b48504f755173f8730476788c3add572c14d95d32d254b8f7594d7803f1; archived source capture; exact capture time not independently established
Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · §4.5 Explicit Goal Nudging Increases Instrumental ConvergenceSelected original resource https://arxiv.org/html/2502.12206; sha256:1bb93b48504f755173f8730476788c3add572c14d95d32d254b8f7594d7803f1; archived source capture; exact capture time not independently established
Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · §4.6 Judge Model Selection Significantly Impacts Convergence DetectionSelected original resource https://arxiv.org/html/2502.12206; sha256:1bb93b48504f755173f8730476788c3add572c14d95d32d254b8f7594d7803f1; archived source capture; exact capture time not independently established
Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · LimitationsSelected original resource https://arxiv.org/html/2502.12206; sha256:1bb93b48504f755173f8730476788c3add572c14d95d32d254b8f7594d7803f1; archived source capture; exact capture time not independently established

Evidence across reports

The same benchmark name can cover different tasks, scoring rules and system configurations. Each source remains visible; comparison groups explain which conditions are aligned.