GAIA

What it measures

Assistant questions require coordinating reasoning with tools and external information.

How to read the result

Inspect which tools and inputs were available and whether the agent found and reconciled the necessary evidence.

Limits and comparison traps

The authors gate the dataset to limit leakage. Respect access terms and compare public validation and private test results separately.

Published evaluation

Exact-answer score at the named GAIA difficulty level. Tools and plugins are part of the listed agent setup; human and search-only baselines are excluded.

Where this fits in an evaluation suite

A transcript that says “done” is weaker evidence than the verified final state. Track policy violations, recoveries, and invalid attempts alongside success.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Exact-answer score at the named GAIA difficulty level. Tools and plugins are part of the listed agent setup; human and search-only baselines are excluded.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.