HarnessEval-W

Agentifying the Evaluation of Visual Worlds

The era of harnessed benchmarking: evaluation defines the taste of evolution.

Overview

HarnessEval-W

A benchmark should deliver more than a scalar score: what makes an evaluation trustworthy is the reasoning that justifies the score. This is especially critical for world models, where judging a rollout requires understanding whether physics, causality, and world state evolve correctly. Humans spot such violations naturally, yet no existing benchmark automates this capability: metrics are computed brute-force, leaving no reasoning chain that can be examined or verified.

HarnessEval is an agentified evaluation pipeline that brings the harness paradigm from the LLM ecosystem to world model benchmarking. Rather than applying a fixed rubric, HarnessEval interprets the context of each evaluation case, decomposes the evaluation question into measurable sub-questions, and spawns specialized sub-agents, each equipped with tailored context and diagnostic tools to reason over its own sub-question. The parent agent then validates the gathered evidence and aggregates it into the final verdict. Every evaluation becomes a transparent evidence tree whose complete reasoning chain justifies the result.

HarnessEval-W Public Set

To support the long-term stability, reproducibility, and continued growth of the benchmark, HarnessEval-W is organized into two complementary splits: a Public Set and a Private Set. The Public Set provides a stable, inspectable target for development, debugging, and transparent comparison. The Private Set is held out for independent evaluation and future benchmark updates, helping reduce overfitting to published cases while preserving room for regression testing and expansion.

This release is the HarnessEval-W Public Set. It contains 100 cases, each pairing an initial observation with an interaction sequence designed to probe how well a world model maintains or updates a coherent visual world.

Model Weights

We release these model weights to make evaluation with HarnessEval-W more convenient by collecting the required checkpoints in one place. They support reproducible local metric backends and academic research; each checkpoint remains subject to its respective upstream license.

Directory	Contents	Upstream license
AMT-S	AMT-S frame-interpolation checkpoint	CC BY-NC 4.0
CLIP	OpenAI CLIP ViT-B/32 and ViT-L/14	MIT
HPSv3	HPSv3 reward checkpoint and local configuration	HPSv3 code: MIT; Qwen backbone: Qwen license
LAION-Aesthetic-Predictor	LAION aesthetic prediction head for CLIP ViT-L/14	MIT
MUSIQ	MUSIQ-KonIQ checkpoint used by pyiqa	Apache-2.0
MegaSAM	MegaSAM/DROID, Depth Anything, UniDepth, and DINOv2 files	See the individual upstream notices; licenses include Apache-2.0, BSD-3-Clause, CC BY 4.0, and CC BY-NC 4.0
Qwen2-VL-7B-Instruct	Qwen2-VL-7B-Instruct base model	Qwen license
Qwen3-VL-A3B-Visual-Plausibility	Qwen3-VL visual-plausibility reward checkpoint	Apache-2.0

Usage

Download Public Set Only

You can download our public set using the Hugging Face CLI:

pip install huggingface_hub
huggingface-cli download MirroS-Lab/HarnessEval-W \
  --repo-type dataset \
  --include "data/**" \
  --local-dir .

Download Model Weights Only

You can download the model weights using the Hugging Face CLI:

huggingface-cli download MirroS-Lab/HarnessEval-W \
  --repo-type dataset \
  --include "model/**" \
  --local-dir .

Download Everything

To download the public set, model weights, and project assets together:

huggingface-cli download MirroS-Lab/HarnessEval-W \
  --repo-type dataset \
  --local-dir .

Citation

If you find HarnessEval useful, please cite:

@article{mirros2026harnessevalw,
  title   = {HarnessEval-W: Agentifying the Evaluation of Visual Worlds},
  author  = {{MirroS Team}},
  journal = {arXiv preprint arXiv:2608.16859},
  year    = {2026}
}

License

We release our dataset under the Apache 2.0 License.