madpyro/beyond-accuracy-loo-scores

Beyond Accuracy: per-scene LOO agent-removal sensitivity scores

Licenses (non-commercial). waymo/ was generated using the Waymo Open Dataset. It is released under the Waymo Dataset License Agreement for Non-Commercial Use (https://waymo.com/open/terms/), and any further downstream use or modification of these data is subject to the terms of that Agreement. av2/ is derived from Argoverse 2 (© 2021 Argo AI, LLC) and released under CC BY-NC-SA 4.0. Details in LICENSE.md.

Scores released with the NeurIPS 2026 Evaluations & Datasets Track paper Beyond Accuracy: Auditing Agent-Removal Sensitivity in Motion Forecasting by Geonyeong Park, Byounghun Park, Nayoung Kim, Kyungmin Kim and Soonmin Hwang (Hanyang University) (project page: https://beyond-accuracy.github.io/, code: https://github.com/beyond-accuracy/beyond-accuracy).

For five motion forecasting models and every validation scene of Argoverse 2 and the Waymo Open Motion Dataset, each surrounding agent $a_i$ is masked in turn. The table records how much the focal agent's forecast error changes against the recorded future $y$:

$${\mathrm{\Delta }}_i=C(f(S\setminus \{a_i\}),y)-C(f(S),y),C\in \{{\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{A}\mathrm{D}\mathrm{E}}_6,{\mathrm{m}\mathrm{i}\mathrm{n}\mathrm{F}\mathrm{D}\mathrm{E}}_6,{\text{Brier-minFDE}}_6\}\backslash Delta\_i\; =\; C\backslash big(f(S\; \backslash setminus\; \backslash \{a\_i\backslash \}),\backslash ,\; y\backslash big)\; -\; C\backslash big(f(S),\backslash ,\; y\backslash big),\; \backslash qquad\; C\; \backslash in\; \backslash \{\backslash mathrm\{minADE\}\_6,\backslash \; \backslash mathrm\{minFDE\}\_6,\backslash \; \backslash text\{Brier-minFDE\}\_6\backslash \}$$

A positive $\Delta_i$ means the error grew when the agent was hidden, and a negative one that it shrank. A score of zero means no change under that cost, not necessarily an identical forecast. These are agent-removal sensitivity scores: they say how a model's output responds to removing an agent, not the agent's causal influence, its semantic importance, or the model's internal reasoning.

Contents

Path	Rows	Size	One row per
av2/agent_scores.parquet	2,685,450	37 MB	(model, scenario, focal agent, surrounding agent)
av2/scene_summary.parquet	114,895	12 MB	(model, scenario, focal agent)
waymo/agent_scores.parquet	5,573,080	84 MB	(model, scenario, focal agent, surrounding agent)
waymo/scene_summary.parquet	220,290	19 MB	(model, scenario, focal agent)

Coverage:

• Models: DeMo, ForecastMAE, Wayformer, AutoBot, MTR, each trained with seed 42.
• Argoverse 2 Motion Forecasting validation: 22,979 scenes, with the dataset's focal agent as the focal agent.
• Waymo Open Motion Dataset v1.1 validation: 44,058 scenes, with the self-driving car as the focal agent.

Every model is evaluated on the same scenes and the same agent slots. A model pair can therefore be joined on (scenario_id, focal_agent_id, slot).

agent_scores.parquet

Column	Type	Meaning
model	category	DeMo, ForecastMAE, Wayformer, AutoBot or MTR
scenario_id	string	scenario id of the source dataset
focal_agent_id	string	track id of the focal agent
slot	int8	position of the agent in UniTraj's 32-slot input, which is sorted by distance to the focal agent; 1 = nearest. Slot 0 is the focal agent and is not listed. Only slots that hold an agent are listed
agent_id	string	track id of the masked agent
delta_minADE6, delta_minFDE6, delta_brier_minFDE6	float32	$\Delta_i$ in metres under each cost

scene_summary.parquet

Column	Meaning
model, scenario_id, focal_agent_id	as above
n_agents	number of surrounding agents listed for this record
baseline_minADE6, baseline_minFDE6, baseline_brier_minFDE6	full-scene error $C(f(S), y)$
{top5,random5,bottom5}_delta_{minADE6,minFDE6,brier_minFDE6}	change in error when the 5 agents with the highest LOO score, 5 random agents, or the 5 with the lowest score are removed together (faithfulness check, paper §5). NaN when the scene has too few agents

Loading

from datasets import load_dataset
scores = load_dataset("madpyro/beyond-accuracy-loo-scores", "waymo_agent_scores", split="validation")
# or directly with pandas
import pandas as pd
df = pd.read_parquet("hf://datasets/madpyro/beyond-accuracy-loo-scores/waymo/agent_scores.parquet")

Configurations: waymo_agent_scores, waymo_scene_summary, av2_agent_scores, av2_scene_summary.

Reproducing

The scores come from tools/experiments/loo_analysis.py in the code repository, with the commands listed in its README and in Appendix A of the paper. They were converted to these tables with tools/export_loo_scores.py. Computed from these files, the mean pairwise Spearman $\rho$ between models (minADE$_6$) is 0.026 on AV2 and 0.021 on Waymo, as reported in the paper.

License

The two folders inherit the terms of their source datasets. They are for non-commercial use only.

• av2/ is derived from the Argoverse 2 Motion Forecasting Dataset (© 2021 Argo AI, LLC). It is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. See av2/LICENSE.
• waymo/ is derived from the Waymo Open Dataset. It is released under the Waymo Dataset License Agreement for Non-Commercial Use (https://waymo.com/open/terms/), and any further downstream use or modification of these data is subject to the terms of that Agreement. See waymo/LICENSE.

Neither folder contains raw sensor, map or trajectory data. To work with the scenes themselves, obtain the datasets from their providers.

Citation

@inproceedings{park2026beyond,
  title     = {Beyond Accuracy: Auditing Agent-Removal Sensitivity in Motion Forecasting},
  author    = {Park, Geonyeong and Park, Byounghun and Kim, Nayoung and Kim, Kyungmin and Hwang, Soonmin},
  booktitle = {The Fortieth Annual Conference on Neural Information Processing Systems Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://openreview.net/forum?id=7axaNwvSMD}
}

When you use the av2/ or waymo/ scores, please also cite Argoverse 2 or the Waymo Open Motion Dataset respectively.