Evaluation dossier

Self-Replication

Whether instructed systems create a separate live copy in controlled trials.

Fudan University / arXiv:2412.12140v1 / Capability

Sources and provenance ↓ Read the research narrative →

CROSS-REPORT MONITORING

PUBLISHED EVIDENCE

Self-Replication · Reported trial success

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

2 observations · 0 dated

● Matched◇ Configuration differs□ Some conditions undisclosed

2 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.

2 observations

Measurements and source conditions
System / model	Result	Condition status	Task / conditions	Evaluator / report	Dates
Qwen 2.5 72B Instruct	90%Reported trial success	Incomplete	Instructed replication trialsPaper-specific planning scaffold (§2.2)	Not disclosedFrontier AI systems have surpassed the self-replicating red line	Release —Report 2024-12-09
Llama 3.1 70B Instruct	50%Reported trial success	Incomplete	Instructed replication trialsPaper-specific planning scaffold (§2.2)	Not disclosedFrontier AI systems have surpassed the self-replicating red line	Release —Report 2024-12-09

Task and measurement

Benchmark protocol profile

Task / subject: Prompted model-driven system in a controlled environment
Measurement target: Trial outcomes for two model-driven systems.
Scoring / outcome: Reported trial success · Completion of the instructed task in the supplied setting.
Interpretation boundary: The supplied environment, explicit instruction and scaffold bound the finding.

A reported evaluation

The source evaluates Llama 3.1 70B Instruct · Instructed replication trials.

Run configuration: Paper-specific planning scaffold (§2.2).

Reported measurement: Reported trial success · 50%.

Completion of the instructed task in the supplied setting. 10 repeated trials per system.

The authors report successful trials under their controlled setup.

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: Can the tested LLM-plus-scaffold create a separate functioning copy without intermediate human assistance?
Task population: The authors combine two open-weight models with an action-feedback scaffold, repeat the copying task, and inspect traces; shutdown and chain variants are additional demonstrations.
Scoring and failure modes: Not established by the retained material
System conditions: Not established by the retained material
Statistical meaning: Not established by the retained material
Validity and interpretation: Two models and a specific scaffold do not represent deployment prevalence. The authors’ broad societal extrapolation exceeds the directly observed copying task.
Appropriate research use: NUWA recommendation: retain access, scaffold, instruction and endpoint definitions before using this work in a replication-risk threshold.

Frontier AI systems have surpassed the self-replicating red line · 2 Experiment designsSelected original resource https://arxiv.org/html/2412.12140v1; sha256:8ae65ad64820f261fcf3f5bc09bfa4e56c5117fc85d487bccff555572a969903; archived source capture; exact capture time not independently established
Frontier AI systems have surpassed the self-replicating red line · 3.1 Summary of resultsSelected original resource https://arxiv.org/html/2412.12140v1; sha256:8ae65ad64820f261fcf3f5bc09bfa4e56c5117fc85d487bccff555572a969903; archived source capture; exact capture time not independently established
Frontier AI systems have surpassed the self-replicating red line · 3.4 Analyzing key factors behind successful self-replicationSelected original resource https://arxiv.org/html/2412.12140v1; sha256:8ae65ad64820f261fcf3f5bc09bfa4e56c5117fc85d487bccff555572a969903; archived source capture; exact capture time not independently established
Frontier AI systems have surpassed the self-replicating red line · 4 DiscussionSelected original resource https://arxiv.org/html/2412.12140v1; sha256:8ae65ad64820f261fcf3f5bc09bfa4e56c5117fc85d487bccff555572a969903; 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.

Frontier AI systems have surpassed the self-replicating red line

Actual evaluators: Not disclosed

2 result records · 2 reported system names

Protocols and comparison decisions (1)

Self-Replication: Reported trial success · Frontier AI systems have surpassed the self-replicating red linesource-report · Scope reviewed

Original source charts and materials

Reported results

Source order · reported configurations

Reported trial success

0–100%

Llama 3.1 70B Instruct10 repeated trials per system. Controlled experiment

Llama 3.1 70B Instruct · 50% · Completion of the instructed task in the supplied setting. · Paper-specific planning scaffold (§2.2) · 10 repeated trials per system.

50%

Qwen 2.5 72B Instruct10 repeated trials per system. Controlled experiment

Qwen 2.5 72B Instruct · 90% · Completion of the instructed task in the supplied setting. · Paper-specific planning scaffold (§2.2) · 10 repeated trials per system.

90%

Showing 2 of 2 systems

Values describe the measured outcome under the source protocol, not a general safety score.

Controlled experiments, 10 trials per system. These rates are not real-world deployment probabilities. Evaluation date not specified; publication / record dates are shown below.

Evaluation conditions and system configurations

Llama 3.1 70B Instruct: 10 repeated trials per system. · Instructed replication trials; modelVersion: not specified · harness: Paper-specific planning scaffold (§2.2) · tools: Controlled computing environment supplied by the authors · budget: not specified

Qwen 2.5 72B Instruct: 10 repeated trials per system. · Instructed replication trials; modelVersion: not specified · harness: Paper-specific planning scaffold (§2.2) · tools: Controlled computing environment supplied by the authors · budget: not specified

Reported results · source order
System / configuration	Reported trial success	Publication / verification
Llama 3.1 70B InstructPaper-specific planning scaffold (§2.2)	50%	Published2024-12-09
Qwen 2.5 72B InstructPaper-specific planning scaffold (§2.2)	90%	Published2024-12-09

References

Numerical results are reported by the cited authors.

[1] Xudong Pan (潘旭东), Jiarun Dai (戴嘉润), et al.. Frontier AI systems have surpassed the self-replicating red line. 2024-12-09. v1 ; Author affiliations on the original paper title page; §1; §3.1

Search NUWA Index