README.md

metadata

pretty_name: Life-Bench
configs:
  - config_name: questions
    default: true
    data_files:
      - split: test
        path: data/questions/test-*
  - config_name: history
    data_files:
      - split: test
        path: data/history/test-*
  - config_name: concepts
    data_files:
      - split: test
        path: data/concepts/test-*
license: cc-by-sa-4.0
task_categories:
  - visual-question-answering
  - question-answering
  - image-text-to-text
tags:
  - multimodal
  - personalization
  - benchmark
  - synthetic
arxiv: 2602.19001
language:
  - en
size_categories:
  - 10K<n<100K

Life-Bench: A Benchmark and Knowledge Graph Framework for Multimodal Personalization Beyond Concept Recognition

Xia Hu, Honglei Zhuang, Brian Potetz, Alireza Fathi, Bo Hu, Babak Samari, Howard Zhou
Google DeepMind

Overview

Life-Bench is a fully synthetic, human-verified multimodal benchmark of 11,811 question–answer pairs across 10 tasks, organized by required evidence scope: Concept Identification, Event Understanding, and Aggregated Reasoning. Underlying these tasks, the personal context comprises social networks and multimodal lifelogs organized into 10 isolated virtual accounts (Vaccounts) to simulate individual users' archives; every question and image has been verified through full-coverage human annotation. All data in Life-Bench (images, social networks, and event histories) is synthetically generated, enabling public release without exposing real users' data.

Please refer to the paper for the benchmark design, construction pipeline, and quality control.

Key Statistics

Statistic	Number
Total Questions	11,811
• Multiple-choice Questions	1,048
• Binary Questions	1,613
• Open Questions	9,150
Total Tasks	10
• Concept Identification	3
• Event Understanding	4
• Aggregated Reasoning	3
Total Concepts	33
• Human Concepts	28
• Animal Concepts	5
Total Events	382
Total Images	2,887
• Concept Portraits	33
• Historical Images	2,479
• Query Images	375
Total Vaccounts	10
• Avg. Concepts per Vaccount	3.3
• Avg. Events per Vaccount	38.2
• Avg. Images per Vaccount	288.7
• Avg. Questions per Vaccount	1,181.1

Tasks

Category	Task	Task Key	Input Format	Answer Format	# Questions
Concept Identification	Text Concept QA	text_concept_qa	Text	Multiple-Choice	163
	Visual Concept Recognition	visual_concept_recognition	Text + Image	Binary	1,613
	Concept VQA	concept_vqa	Text + Image	Multiple-Choice	885
Event Understanding	Scene and Activity	scene_and_activity	Text	Open Generation	2,265
	Direct Person-Centric	direct_person_centric	Text	Open Generation	2,245
	Relational Person-Centric	relational_person_centric	Text	Open Generation	1,812
	Fine-Grained Scene	fine_grained_scene	Text	Open Generation	1,981
Aggregated Reasoning	Preference and Persona	preference_and_persona	Text	Open Gen (Option-guided)	295
	Frequency and Counting	frequency_and_counting	Text	Open Generation	265
	Relational Temporal Reasoning	relational_temporal_reasoning	Text	Open Generation	287
Total					11,811

Additionally Released Single-hop Questions

In addition to the 11,811 questions above, we additionally provide 1,647 single-hop direct concept recognition questions (139 for Text Concept QA, 1,032 for Visual Concept Recognition, and 476 for Concept VQA) as diagnostic baselines to monitor model fallback, bringing the total released benchmark to 13,458 QA pairs. These are excluded from the primary evaluation because all evaluated retrieval approaches already near-saturate on single-hop concept lookup, limiting their discriminative value for method comparison; they are instead provided as a diagnostic set for verifying that more complex personalization methods do not regress on basic entity recognition.

They are stored separately with the task values text_concept_qa_easy, visual_concept_recognition_easy, and concept_vqa_easy (files questions/<task>_easy.jsonl). Results on Life-Bench should be reported on the 11,811 questions in the Tasks table.

Data Format

google/life-bench/
├── data/                                # Parquet tables (images embedded) for `datasets` / Dataset Viewer
│   ├── questions/test-*.parquet         # 13,458 rows: 11,811 questions + 1,647 single-hop questions
│   ├── history/test-*.parquet           #  2,479 rows: one row per historical image
│   └── concepts/test-*.parquet          #     33 rows: one row per concept
├── concepts/{vaccount}/                 # Raw files (same content as the Parquet tables)
│   ├── concepts.json                    #   concepts of the Vaccount
│   └── portraits/*.jpg                  #   concept portraits
├── events/
│   ├── metadata.jsonl                   #   one record per historical image (382 events)
│   └── images/*.jpg                     #   2,479 historical images
├── questions/
│   ├── <task>.jsonl                     #   10 task files + 3 `<task>_easy.jsonl` files
│   └── images/*.jpg                     #   375 query images
├── metadata/statistics.json
├── LICENSE
└── README.md

questions

Field	Type	Description
sample_id	string	Question id, e.g. "sun-ja_concept_vqa_120".
vaccount	string	Vaccount the question belongs to.
task	string	Task key (see the Tasks table and the _easy variants above).
question	string	Question text (including options for multiple-choice questions).
answer	string	Reference answer.
answer_type	string	"multiple_choice", "binary", or "open_generation".
question_image_id	string	Query image id for text+image questions; empty for text-only questions.
question_image_path	string	Relative path of the query image, e.g. "questions/images/sun-ja-s-7.jpg"; empty for text-only questions.
question_image	Image	Query image (Parquet only); null for text-only questions.
reference_ids	list[string]	Gold supporting event_id and concept_id values for Concept Identification and Event Understanding questions; [] for Aggregated Reasoning questions.

history (raw: events/metadata.jsonl)

Field	Type	Description
image_id	string	Historical image id, e.g. "sun-ja-e-0-0".
image_path	string	Relative path, e.g. "events/images/sun-ja-e-0-0.jpg".
image	Image	Historical image (Parquet only).
vaccount	string	Vaccount the image belongs to.
event_id	string	Event the image belongs to, e.g. "sun-ja-e-0". Images from the same event share the same date and description.
event_description	string	Description of the event.
event_date	string	Date of the event, YYYY-MM-DD.

concepts (raw: concepts/{vaccount}/concepts.json)

Field	Type	Description
concept_id	string	Concept id, e.g. "sun-ja-c-leo".
vaccount	string	Vaccount the concept belongs to.
name	string	Concept name, e.g. "Leo".
role	string	Relation to the account owner, e.g. "account owner", "Grandson".
portrait_path	string	Relative path, e.g. "concepts/sun-ja/portraits/leo.jpg".
portrait	Image	Concept portrait (Parquet only).

Loading

from datasets import load_dataset

questions = load_dataset("google/life-bench", "questions", split="test")
history   = load_dataset("google/life-bench", "history",   split="test")
concepts  = load_dataset("google/life-bench", "concepts",  split="test")

# The 11,811 benchmark questions (excluding the single-hop `*_easy` questions)
benchmark = questions.filter(lambda x: not x["task"].endswith("_easy"))

To download the raw file tree instead:

from huggingface_hub import snapshot_download

snapshot_download(repo_id="google/life-bench", repo_type="dataset",
                  allow_patterns=["concepts/*", "events/*", "questions/*", "metadata/*"])

Evaluation code is available at github.com/google-deepmind/life-bench.

Evaluation Results

Our primary metric is accuracy: multiple-choice and binary questions use exact match, while Gemini 3.6 Flash judges open-ended responses against the ground-truth answer using LLM-as-a-Judge. We additionally report recall to assess retrieval quality.

Accuracy. Retrieval-based methods are compared at context size $k = 5$, all using Gemma-3 12B as backbone; the best and second-best scores among them are in bold and underlined respectively. Reference settings contextualize retrieval performance and are generated with Gemini-3.5-Flash except Random: Close-book answers without personal context; Oracle receives only the gold supporting evidence (Oracle-Gemma: with the Gemma backbone); Full-context places an entire Vaccount history in the prompt.

Method	Text Concept QA	Visual Concept Recognition	Concept VQA	Scene & Activity	Direct Person-Centric	Relational Person-Centric	Fine-Grained Scene	Preference & Persona	Relational Temporal	Frequency & Counting	Overall
Random	0.2500	0.5000	0.2500	0.0000	0.0000	0.0000	0.0000	0.2003	0.0000	0.0000	0.1200
Close-book	0.2454	0.5111	0.2696	0.0848	0.0539	0.1341	0.1560	0.4373	0.0662	0.0755	0.2034
Full-context	0.9730	0.9722	0.9508	0.9143	0.9095	0.9141	0.9014	0.8856	0.8523	0.3823	0.8655
Oracle	0.9816	0.9826	0.9570	0.9324	0.9292	0.9269	0.9100	–	–	–	0.9457*
Oracle-Gemma	0.8527	0.8698	0.6170	0.8093	0.7007	0.7130	0.7410	–	–	–	0.7576*
BM25	0.8650	0.8320	0.6249	0.6698	0.3648	0.4923	0.4962	0.5593	0.1080	0.1472	0.5159
RAG-Cap	0.8405	0.7650	0.5236	0.4843	0.2895	0.3786	0.5671	0.7051	0.1428	0.1623	0.4859
RAP-Gemma	0.7651	0.7498	0.4744	0.3043	0.1688	0.2467	0.5172	0.7119	0.1533	0.0755	0.4167
R2P-Gemma	0.7055	0.6943	0.3559	0.3244	0.2664	0.3488	0.4338	0.5831	0.1715	0.1094	0.3993
HippoRAG2	0.8098	0.8494	0.6542	0.3978	0.2094	0.3201	0.4356	0.5559	0.1672	0.1623	0.4562
LifeGraph ($d=2$)	0.7587	0.8332	0.5318	0.7242	0.5092	0.5193	0.5157	0.6746	0.3240	0.1785	0.5569
LifeGraph ($d=3$)	0.7914	0.8278	0.5260	0.7228	0.4984	0.5348	0.5253	0.6524	0.3275	0.1976	0.5604

* Oracle overall averages over seven concept and event tasks since aggregate does not have subset golden reference.

Event-level Recall@3/5, where retrieving any image from a gold event counts as a hit.

Method	Concept R@3	Concept R@5	Event R@3	Event R@5	Overall R@3	Overall R@5
BM25	0.4626	0.5330	0.4785	0.5371	0.4717	0.5353
RAG-Cap	0.4307	0.4808	0.3665	0.4580	0.3940	0.4678
R2P	0.4070	0.4238	0.3823	0.4299	0.3929	0.4273
RAP	0.4354	0.4653	0.4065	0.4572	0.4188	0.4607
HippoRAG2	0.4465	0.5128	0.3042	0.4065	0.3652	0.4521
LifeGraph ($d=2$)	0.4748	0.5009	0.5309	0.6021	0.5068	0.5587
LifeGraph ($d=3$)	0.4844	0.4991	0.5271	0.5963	0.5088	0.5546

License

Copyright 2026 Google LLC

The Life-Bench data (images, personal-context records, and question–answer pairs) is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0). You may obtain a copy of the CC BY-SA 4.0 license at: https://creativecommons.org/licenses/by-sa/4.0/legalcode. Life-Bench is released for evaluation purposes.

All data in Life-Bench is synthetically generated; the generated accounts map to no real individual. To retain realism, generation is seeded with textual captions of images from a CC-licensed subset of the YFCC100M dataset (Thomee et al., 2016). YFCC source images serve only as captioning inputs and never enter Vaccounts, and the captions are not released.

Unless required by applicable law or agreed to in writing, all materials distributed here under the CC BY-SA 4.0 license are distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the license for the specific language governing permissions and limitations under the license.

This is not an official Google product.

Citation

@misc{hu2026lifebench,
      title={Life-Bench: A Benchmark and Knowledge Graph Framework for Multimodal Personalization Beyond Concept Recognition},
      author={Xia Hu and Honglei Zhuang and Brian Potetz and Alireza Fathi and Bo Hu and Babak Samari and Howard Zhou},
      year={2026},
      eprint={2602.19001},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.19001},
}