File size: 8,574 Bytes
ea6f59c
 
 
 
 
 
 
 
 
 
 
42aa04f
 
 
 
 
 
 
 
 
 
 
 
 
ea6f59c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42aa04f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea6f59c
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
---
license: cc-by-4.0
task_categories:
  - image-to-text
  - visual-question-answering
tags:
  - Multimodal benchmark
  - Vision-Language Models
  - Compositionality
  - Localism-aware compositionality
  - Multimodal knowledge editing
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/lace_train.parquet
      - split: test
        path: data/lace_test.parquet
  - config_name: keyword_dict
    data_files:
      - split: train
        path: data/train_keyword_dict.parquet
      - split: test
        path: data/test_keyword_dict.parquet
---

# LACE-Bench: Localism-Aware Compositionality Evaluation Benchmark for Vision-Language Models

> **LACE-Bench** is a benchmark for evaluating *localism-aware compositionality* in vision-language models (VLMs) — the ability to selectively integrate local region-level semantics with global scene-level understanding. It comprises two complementary tasks: **LoGoCap** and **MMComE**.

## Dataset Card

| Field | Info |
|---|---|
| **Tasks** | LoGoCap (Local & Global Compositional Captioning), MMComE (Multimodal Compositional Knowledge Editing) |
| **Modality** | Vision-Language |
| **Split** | Train (9,874 images) / Test (2,183 images) |
| **Total** | 12,057 images |
| **Image Source** | [Visual Genome](https://homes.cs.washington.edu/~ranjay/visualgenome/api.html) |
| **License** | CC BY 4.0 |

<!-- 
## Tasks

### 1. LoGoCap — Multi-grained Local and Global Compositional Captioning

LoGoCap evaluates a model's *static selectivity*: can it simultaneously understand the global scene while identifying and grounding constituent local objects?

- **Local captioning**: given an atomic region (a single object marked with a colored bounding box), generate a region-specific caption.
- **Global captioning**: given a compound region (a group of atomic regions), generate a single coherent caption that integrates all constituent local parts while introducing holistic scene-level context (moods, relations, atmosphere) not present in any individual local caption.

Evaluation uses standard captioning metrics (BLEU, ROUGE-1, METEOR) against human-annotated reference captions.

### 2. MMComE — Multimodal Compositional Knowledge Editing

MMComE evaluates a model's *dynamic robustness*: can it apply a localized counterfactual edit (e.g., replacing *referee* with *spectator*) consistently across region-marked images, while preserving all unrelated global semantics?

A multimodal edit request is defined as a tuple `(I, r, ph → ph*)`, where:
- `I` is the image, `r` is the target region
- `ph` is the original phrase to be replaced, `ph*` is the counterfactual substitute

The model is evaluated on whether it correctly reflects `ph*` in both in-scope regions (edited) and correctly retains all out-of-scope regions (unedited).


## Intended Use

LACE-Bench is designed for:

- Evaluating **localism-aware compositionality** — whether VLMs can selectively deploy local and global compositional operations as the task demands
- Measuring **global binding stability**: how consistently local semantic units of atomic regions bind into global captions
- Quantifying **cross-scale interference**: the degree to which local counterfactual edits propagate into unintended global semantic regions
- Benchmarking **fine-tuning strategies** (e.g., LoRA, blur+bbox visual grounding) for compositional captioning -->


## Data Fields

Each record corresponds to one image and contains the following fields:

| Field | Type | Description |
|---|---|---|
| `image_id` | string | Visual Genome image identifier |
| `regions` | list[object] | Annotated bounding box regions (atomic) |
| `narratives` | string | Description of the full image |
| `keywords` | list[object] | Key noun concepts grounded in WordNet |
| `relation_centric_regions` | list[object] | Groups of region IDs with a human-written relational annotation |

### `regions`

Each atomic region corresponds to a single object marked with a distinct colored bounding box.

| Key | Type | Description |
|---|---|---|
| `id` | string | Region identifier (`{image_id}_{region_index}`) |
| `color` | string | Bounding box color used for visual grounding (aqua / yellow / lime / red / blue / orange / magenta) |
| `x`, `y` | float | Top-left corner coordinates of the bounding box |
| `width`, `height` | float | Width and height of the bounding box |
| `captions` | list[object] | Human-annotated region-level captions (see below) |
| `object_ids` | list[int] | Linked object IDs from Visual Genome |
| `relationships` | list[object] | Scene graph relationships associated with this region (see below) |

**`regions[].captions`**

| Key | Type | Description |
|---|---|---|
| `caption` | string | Original human-written caption for the region (e.g. `"the tall clock on the street"`) |
| `counterfactual_caption` | string | Minimally edited caption where one noun is replaced with a plausible but incorrect alternative (e.g. `"the tall dart board on the street"`) |

**`regions[].relationships`**

| Key | Type | Description |
|---|---|---|
| `relationship_id` | int | Visual Genome relationship identifier |
| `predicate` | string | Relation predicate between subject and object (e.g. `"on"`) |
| `synsets` | list[string] | WordNet synsets for the predicate (e.g. `["along.r.01"]`) |
| `subject_id` | int | Visual Genome object ID of the subject |
| `object_id` | int | Visual Genome object ID of the object |

---

### `keywords`

Each entry represents a key noun concept extracted from region captions and grounded in WordNet.

| Key | Type | Description |
|---|---|---|
| `synset_id` | string | WordNet synset identifier (e.g. `clock.n.01`) |
| `synonyms` | list[string] | Lemma names belonging to this synset (e.g. `["clock"]`) |
| `nearest_ancestor` | string | Closest hypernym synset in the WordNet hierarchy (e.g. `timepiece.n.01`) |
| `supersense` | string | Broad semantic category from WordNet lexicographer files (e.g. `noun.artifact`, `noun.person`) |
| `counterfactual` | list[object] | Human-annotated counterfactual substitutions for this concept (see below) |

**`keywords[].counterfactual`**

| Key | Type | Description |
|---|---|---|
| `human_annotation` | string | Plausible but incorrect substitute chosen by a human annotator (e.g. `"dart board"`) |
| `candidate` | list[string] | Candidate substitutions presented to the annotator for selection |

> `counterfactual` is empty (`[]`) for concepts where no counterfactual annotation was collected.

---

### `relation_centric_regions`

Each entry groups multiple atomic regions and provides a human-written description of the relational context among them.

| Key | Type | Description |
|---|---|---|
| `human_annotation` | string | Free-form description of the spatial or semantic relationship among the grouped regions (e.g. `"The central clock tower... stands as a focal point against the backdrop of the building's pillars."`) |
| `region_ids` | list[string] | IDs of the atomic regions involved in this relational group (e.g. `["2358647_0", "2358647_1"]`) |


## `keyword_dict` Config

In addition to the per-image records (`default` config), the dataset ships a `keyword_dict` config that maps each WordNet synset ID to the list of surface phrases observed for that concept across the corresponding split's region captions. These dictionaries are useful for keyword-based lookup, counterfactual phrase matching, and lexical normalization of mentions.

**Files**

- `data/train_keyword_dict.parquet`
- `data/test_keyword_dict.parquet`

**Schema**

| Field | Type | Description |
|---|---|---|
| `synset_id` | string | WordNet synset identifier (e.g. `tree.n.01`), matching `keywords[].synset_id` in the `default` config |
| `phrases` | list[string] | All distinct surface phrases (synonyms, plural forms, modifier-noun variants, casing variants) observed for the synset in that split |

**Example rows**

| `synset_id` | `phrases` |
|---|---|
| `leaf.n.01` | `["leaves", "foliage", "leaf", "banana leaf", "dried leaves", "green leaves", ...]` |
| `tree.n.01` | `["tree", "trunk", "evergreen tree", "pine trees", "fir tree", ...]` |
| `bus.n.01`  | `["bus", "city bus", "double decker bus", "motorbus", "coach", ...]` |

**Loading**

```python
from datasets import load_dataset
ds = load_dataset("lacebench/LACE-Bench", "keyword_dict")
```


## Citation

```bibtex
@dataset{anonymous2026lacebench,
  title  = {LACE-Bench: Localism-Aware Compositionality Evaluation Benchmark for Vision-Language Models},
  author = {Anonymous},
  year   = {2026},
}
```