ICBench

ICBench is a controlled diagnostic benchmark for evaluating linguistic grounding in Vision-Language-Action (VLA) models. It pairs original LIBERO task instructions with contradictory instructions while keeping the visual scene unchanged.

This release accompanies Restoring Linguistic Grounding in VLA Models via Train-Free Attention Recalibration (IGAR).

• Project page: https://ray-nh.github.io/igar/
• Data file: icbench.jsonl
• Format: UTF-8 JSON Lines; one instruction pair per row
• Release contents: instruction text, contradiction labels, and references to LIBERO BDDL task paths. Robot trajectories, images, simulator assets, and model weights are not included.

Dataset composition

The file contains 120 instruction pairs across 30 distinct tasks. Each task has four contradictory instruction variants.

LIBERO suite	Distinct tasks	Instruction pairs
libero_spatial	10	40
libero_object	10	40
libero_goal	10	40
Total	30	120

Variant	Contradiction	Pairs
V1	Manipulated object attribute	30
V2	Target attribute	30
V3	Both manipulated object and target attributes	30
V4	Spatial relation	30

All records are exposed as a single test split for evaluation. The source file does not define a training or validation partition.

Data fields

Field	Type	Description
bddl_path	string	Relative path to the corresponding LIBERO BDDL task specification.
suite	string	LIBERO task suite.
original_instruction	string	Original task instruction.
contradiction.instruction	string	Contradictory instruction used with the same scene.
contradiction.variant	string	Contradiction category: V1, V2, V3, or V4.

Example record (unchanged from the source file):

{
  "bddl_path": "libero_spatial/pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate.bddl",
  "contradiction": {
    "instruction": "Pick the akita white bowl from table center and place it on the plate",
    "variant": "V1"
  },
  "original_instruction": "Pick the akita black bowl from table center and place it on the plate",
  "suite": "libero_spatial"
}

Loading

Install the Hugging Face Datasets library:

pip install datasets

from datasets import load_dataset

dataset = load_dataset("Raynh945/ICBench", split="test")
print(len(dataset))  # 120
print(dataset[0]["contradiction"]["instruction"])

The JSONL file can also be read without third-party libraries:

import json

with open("icbench.jsonl", encoding="utf-8") as f:
    records = [json.loads(line) for line in f if line.strip()]

License

The ICBench instruction dataset is released under CC BY 4.0. Please credit the authors and cite the paper when using this release. The full license text is in LICENSE. LIBERO-derived material retains the upstream MIT notice in NOTICE.

Citation

@inproceedings{zhang2026restoring,
  title={Restoring Linguistic Grounding in VLA Models via Train-Free Attention Recalibration},
  author={Ninghao Zhang and Bin Zhu and Shijie Zhou and Jingjing Chen},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026},
  url={https://arxiv.org/abs/2603.06001}
}