Benchmark for Agentic LLM and VLM Reasoning On Games (NetHack, MiniHack, BabyAI, Crafter, TextWorld, Baba Is You).
40 model results
One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.
Published observations span 0.1 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#5 of 40 in this displayed comparison; 11.35 pp below the leader.
Paglieri, Davide, Cupiał, Bartłomiej, Coward, Samuel, Piterbarg, Ulyana, Wolczyk, Maciej, Khan, Akbir et al. · 2024
All 13 authors
Paglieri, Davide, Cupiał, Bartłomiej, Coward, Samuel, Piterbarg, Ulyana, Wolczyk, Maciej, Khan, Akbir, Pignatelli, Eduardo, Kuciński, Łukasz, Pinto, Lerrel, Fergus, Rob, Foerster, Jakob Nicolaus, Parker-Holder, Jack, Rocktäschel, Tim
Dates & evaluation details
Balrog · Average progress · % · 2026-10-26
Epoch AI public data snapshot, CC BY 4.0. Externally collected result; original evaluation provider is linked. CSV: balrog_external.csv; score column: Average progress; fraction scale: 1. Model version is retained in configuration. Original source authors retain credit and any additional licensing terms.
TRACE THE SCORE · Per-model notes use published breakdowns where available.
Score vs. release date
Epoch AI · 41 observations
Balrog · Average progress · % · higher is better
Dates are model release dates recorded by Epoch AI, not evaluation dates. Each point retains its published score and configuration. Separate releases are not a matched longitudinal experiment. Missing dates are omitted. Explore all reported results ↓
The dashed line traces the best observed score through the model release dates. It describes this selection of reports, with their original configurations. It does not estimate unmeasured models or costs.
View plotted values and sources (41)
BALROG · Balrog · Average progress · %. Published measurements in their original score units.
Benchmark for Agentic LLM and VLM Reasoning On Games (NetHack, MiniHack, BabyAI, Crafter, TextWorld, Baba Is You).
Reading the measurement
Partial credit can distinguish a useful approach from an empty attempt, but it is not the same as a completed task. Read the rubric, full-success count and score distribution together. For mathematical work, preserve the difference between expert review and acceptance by a formal proof checker.
Evidence status
Catalogued from the supplied benchmark research notes. Publisher directories, where linked, are discovery references. Numerical model results are shown only when a source dataset has been imported; a catalog entry alone does not establish an independently reviewed evaluation.
Epoch data and evaluation setting
Epoch AI publishes this benchmark in its capability collection. Imported rows retain the published score column and model configuration. Original authors and evaluation providers remain credited in the result references.
Epoch attribution and score coverage
Data is attributed to Epoch AI under CC BY 4.0, with original evaluation providers credited in the linked source snapshots. Scores are imported only where a published mapping identifies the field and its scale; missing or incompatible records are omitted rather than assigned zero.
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
Record the source publication, benchmark version, scoring units, denominator, tool access and attempt budget. Preserve uncertainty and distinguish developer-reported results from independent evaluations.
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.