Evaluation

Benchmarks And Transfer Checks

Which public benchmarks prove what, plus how to hold out transfer cases that block quiet gaming.

10 min readIntermediateUpdated 2026-09-19

What each benchmark proves

Coding benchmarks such as SWE bench and Polyglot test whether an agent can read, edit, and repair real code under tests. Reasoning sets such as GSM8K and MATH test step by step problem solving. Decision suites such as HotPotQA and AlfWorld test search plus multi step choices. HumanEval era code tasks test single shot generation with hidden tests.

No single suite proves recursion. A benchmark proves a skill at one moment. Recursion is proven by the diary across cycles, where the same suite plus fresh transfer cases move together.

Holding out transfer the right way

Split every fixture into public checks the generator may see and hidden checks it never sees during search. Promotion runs hidden checks once with full logs. A patch that passes public checks but fails hidden ones returns to the archive with its residue.

Rotate hidden sets on a schedule and track the gap between public and hidden scores. A widening gap means the loop is learning the school instead of the subject, so difficulty steps down and diversity checks tighten.

Benchmark map used across the atlas

BenchmarkSkill under testAtlas use
SWE benchReal code repair under testsArchive search validation
PolyglotMultilingual code tasksSecond validation axis
GSM8KMath word problemsReasoning fixtures
ARC AGIVisual pattern generalizationTransfer probe
HotPotQAMulti hop search answersDecision memory test
AlfWorldHousehold multi step tasksReflection memory test
HumanEvalSingle shot code generationClassic baseline

Continue reading

Found a gap in this page.Open the GitHub repo to propose a fix