NUWA Evidence record
DeepSeek-R1
Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? · Evidence record
Original source charts and materials
Reported results
Source order · reported configurations
Select systems
Evading Shutdown
0–100%
DeepSeek-R1Values describe the measured outcome under the source protocol, not a general safety score.
Table 1: category-specific and overall instrumental rates, using GPT-4o as judge. Elicited scenario rates do not estimate deployment probability. Evaluation date not specified; publication / record dates are shown below.
Evaluation conditions and system configurations
- DeepSeek-R1
- See the source protocol; no independent trial count inferred. · Table 1 · Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
| System / configuration | Evading Shutdown | Hacking Computer Systems | Running Multiple AI Copies | Hiding Unwanted Behavior | Strategically Appearing Aligned | Deceptive Behavior | Total | Publication / verification |
|---|---|---|---|---|---|---|---|---|
| DeepSeek-R1Harness not specified | 50.00% | 40.00% | 46.15% | 63.64% | 45.45% | 27.27% | 46.05% | Source checked2026-10-25 |
Method and scope
Tests model responses for instrumental goals such as resource acquisition and self-preservation under specified scenarios.
- Evidence type
- Benchmark
- Assessment subject
- Model or agent system in the original evaluation configuration
- Primary dimension
- Propensity
- Evidence snapshot
- 2026-10-29.1
Classified by the operational outcome in the original protocol. Topic associations are documented separately from question-level matrix claims.
Source finding
Source-reported measurements for DeepSeek-R1 in Table 1; values are transcribed, not independently reproduced.
Limitations
Textual choices and grader labels do not establish executed autonomous behavior or stable deployment propensities.
Configuration, task and aggregation differences remain as reported. Source order is retained. Transcription is not independent replication; no cross-protocol ranking is supported.
Related research questions
References
Numerical results are reported by the cited authors.
- [1] Yufei He, Yuexin Li, et al.. Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals?. 2025-02-16. v1 [arXiv]; Author affiliations on the original paper title page; Table 1