CLMCP Agent Benchmark Sandbox

TASK

BENCHMARK OBJECTIVE

Find and enter the watchtower

Find the nearby pillager outpost watchtower, enter it through a valid ground-floor opening, and stand inside its ground-floor interior.

Cross-version zero-shot Minecraft 1.12.2 training corpus → Java 1.14 structure → Paper 1.21.11 world The Stage 1 training data predates pillager outposts, which were introduced in Java 1.14.

Same worldSame spawnSame plannerFinish alive

Outcome—

Best tower rank—

Semantic calls—

Completion tick—

Episode tokens——

SAME TASK · SAME WORLD · SAME PLANNER

CLMCP localizes and enters earlier

—

LIVE 3D INSPECTOR

Vanilla Pillager Outpost

Loading Minecraft assets Preparing official 1.21.11 resource pack…

WebGL · real block models · drag to orbit

Agent trajectory Requested footprint Top candidates Ground truth interior

tick 0 / 0

NATIVE XINCLAW TRACE

Reasoning + tool I/O

Full JSONL.gz 0/0 events

20 Hz state · native model/tool trace

The agent has not acted yet.

AUTHORITATIVE PROGRESS

What has happened?

—

Select an episode to inspect the perception → planning → action boundary.

SEMANTIC RETRIEVAL

CLMCP candidates

0 queries

No semantic searches in this episode.