benchmarks evals
The Frozen Model Invented Improvements Too
A measured null exposed seven ways transition-level self-improvement ledgers can manufacture gains.
Summary
A measured null exposed seven ways transition-level self-improvement ledgers can manufacture gains.
A frozen control pushed through the same evaluation pipeline showed apparent capability changes, including artifacts tied to greedy decoding and batching. A per-problem exact test with false-discovery-rate control found no changes on held-out null replicates, making matched baseline measurement the paper’s central recommendation.
Why it matters
A measured null exposed seven ways transition-level self-improvement ledgers can manufacture gains.
Limits and context
No additional limitation was separately recorded.
Key claims
A measured null exposed seven ways transition-level self-improvement ledgers can manufacture gains.
Evidence: source-2026-08-23-021
Sources
- arXiv preprint 2608.20290arXiv · primary research
Corrections
No corrections have been recorded for this story.