WhitzardNUWA Lab ↗
NUWANUWA INDEX

Search NUWA Index

Research · Benchmarks · Findings · Perspectives

中文
← NUWA Index

Evidence record

Claude Sonnet 4.6 (Claude Code)

PostTrainBench and OpenAI PostTrainBench Lite · Evidence record

ELLIS Institute Tübingen · Max Planck Institute for Intelligent Systems · University of Tübingen · Thoughtful Lab / arXiv:2603.08640v2 / Capability

Sources and provenance ↓ Read the research narrative →

Original source charts and materials

Reported results

Source order · reported configurations

Weighted average performance

0–100
Claude Sonnet 4.6 (Claude Code)
16.4
Showing 1 of 1 systems

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

Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. Evaluation date not specified; publication / record dates are shown below.

Evaluation conditions and system configurations
Claude Sonnet 4.6 (Claude Code)
See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
Reported results · source order
System / configurationWeighted average performancePublication / verification
Claude Sonnet 4.6 (Claude Code)Harness not specified16.4Source checked2026-10-25

Method and scope

Tests whether agents can improve base models through post-training under a fixed compute budget.

Evidence type
Benchmark
Assessment subject
Model or agent system under the original task configuration
Primary dimension
Capability
Evidence snapshot
2026-10-29.1

The main dimension follows the task outcome in the original protocol. Topic relevance is documented in the parent research record; no unreviewed question-level links are inferred.

Source finding

Source-reported measurements for Claude Sonnet 4.6 (Claude Code) in Table 1; values are transcribed, not independently reproduced.

Limitations

Targeted benchmark gains, test contamination and unauthorized shortcuts require separate checks.

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. [1] Ben Rank ∗, Hardik Bhatnagar ∗, et al.. PostTrainBench: Can LLM Agents Automate LLM Post-Training?. 2026-03-09. v2 [arXiv]; Author affiliations on the original paper title page; Table 1