- Who it helps
- Agent-evaluation researchers and developers studying game decision policies.
- Availability
- The repository documents a frozen benchmark edition and replay artifacts. A legally obtained game ROM is required separately; no ROM or ready-to-play service is supplied here.
What a result means
The frozen edition uses Liu Kang against a defined Very Hard CPU schedule. Mean damage is the headline measure and wins are reported separately. Early results described in the README are weak; they should not be presented as agents mastering the game.
Separate replay from model provenance
Replay checks can verify recorded game actions and outcomes. They do not independently prove which model produced those actions or whether reported model cost and latency are accurate. For a comparison, retain the model-call trace alongside the replay and keep the benchmark edition fixed.
What you can do
- Read the frozen-edition rules and artifact requirements. Compare only submissions with matching conditions and separate verified game behavior from self-reported inference details. Open reference ↗
Where the evidence stops
No emulator, model or third-party code was run. The benchmark covers a narrow game setup, and its results do not establish general planning ability.
Sources read. Sources checked October 19, 2026 UTC. No external product workflow, model inference or third-party project was executed.
Sources & further reading
GitHub · implementation
FightBench benchmark repository ↗
Read frozen-edition conditions, replay checks and limits of self-reported model costs. No game or external code executed.
aipdv · page checked 2026-10-19
Coverage note: The cited primary material was read on October 19. A matched original X post, a watched video comparison and independent execution remain unavailable.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
FightBench gives agents a constrained Mortal Kombat II task and records game outcomes. It separates a fixed policy table from live model decisions so results can be compared under stated conditions.