BALROG
Not yet importedA shared evaluation framework for language and visual agents across long-horizon reinforcement-learning game environments.
2 task previewsagent · sandboxedReview requiredversion upstream
Complex import. Requires a multi-environment sandbox adapter, parallel evaluation, modality-aware traces, and reward preservation.
What it measures
BALROG is strategically useful as an adapter rather than a single game. One integration can expose several established environments through consistent model, rollout, and scoring machinery.
The import must preserve environment identity, observation modality, action interface, provider settings, and benchmark-specific rewards so unlike runs are never merged into one leaderboard.
not yet imported
Not yet imported. Trunchbull should implement BALROG as shared simulation infrastructure, not duplicate each included environment.
Tasks
2 previewed| id | task | runtime |
|---|---|---|
| balrog/language-agent | Language-agent rolloutRun an environment through structured language observations and actions.tools: BALROG adapter | sandbox |
| balrog/visual-agent | Visual-agent rolloutRun the same environment class through pixel observations.tools: BALROG adapter, visual observation | sandbox |
Tools
1 requested| tool | description | kind | required |
|---|---|---|---|
| BALROG adapter | Normalizes lifecycle, actions, observations, and rewards across environments. | sandbox | yes |