| --- |
| license: apache-2.0 |
| task_categories: |
| - question-answering |
| language: |
| - en |
| tags: |
| - factuality |
| - search |
| - retrieval |
| - deep research |
| - comprehensiveness |
| - agent |
| - posttraining |
| - benchmark |
| - Google DeepMind |
| pretty_name: DeepSearchQA |
| size_categories: |
| - n<1K |
| configs: |
| - config_name: deepsearchqa |
| default: true |
| data_files: |
| - split: eval |
| path: DSQA-full.csv |
| --- |
| |
| |
|
|
| ▶ [Google DeepMind Release Blog Post](https://blog.google/technology/developers/deep-research-agent-gemini-api/)\ |
| ▶ [DeepSearchQA Leaderboard on Kaggle](https://www.kaggle.com/benchmarks/google/dsqa)\ |
| ▶ [Technical Report](https://storage.googleapis.com/deepmind-media/DeepSearchQA/DeepSearchQA_benchmark_paper.pdf)\ |
| ▶ [Evaluation Starter Code](https://www.kaggle.com/code/andrewmingwang/deepsearchqa-starter-code) |
|
|
|
|
| |
|
|
| DeepSearchQA is a 900-prompt benchmark for evaluating agents on difficult multi-step information-seeking tasks across 17 different fields. Unlike traditional benchmarks that target single-answer retrieval or broad-spectrum factuality, DeepSearchQA features a dataset of challenging, hand-crafted tasks designed to evaluate an agent’s ability to execute complex search plans to generate exhaustive answer lists. |
|
|
| Each task is structured as a "causal chain", where discovering information for one step is dependent on the successful completion of the previous one, stressing long-horizon planning and context retention. All tasks are grounded in the open web with objectively verifiable answer sets. |
|
|
| DeepSearchQA is meant to be used to evaluate LLMs or LLM agents with access to the web. |
|
|
| |
|
|
| This dataset is a collection of 900 examples. Each example is composed of: |
|
|
| * A problem (`problem`) which is the prompt testing parametric knowledge. |
| * A problem category (`problem_category`) specifying which of 17 different domains the problem belongs to. |
| * A gold answer (`answer`) which is used in conjunction with the evaluation prompt to judge the correctness of an LLM's response. |
| * An answer type classification (`answer_type`) specifying whether a single answer or set of answers is expected as a response. This information should NOT be given to the LLM during inference time. 65% of answers are of type `Set Answer`. |
|
|
| See the [Technical Report](https://storage.googleapis.com/deepmind-media/DeepSearchQA/DeepSearchQA_benchmark_paper.pdf) for methodology details. |
|
|
| |
| While DeepSearchQA offers a robust framework for evaluating comprehensive retrieval, it relies on |
| specific design choices that entail certain limitations. By employing an exclusively outcome-based |
| evaluation, we effectively treat any agent that is evaluated as a black box. In the absence of trajectory data, it is difficult |
| to distinguish between an agent that reasoned correctly and one that arrived at the correct list through |
| inefficient or accidental means (e.g., lucky guessing). Additionally, the static web assumption, while |
| necessary for reproducibility, limits the evaluation of “breaking news” retrieval where ground truth is |
| volatile. A task’s ground truth may become outdated if source websites are removed or their content |
| is significantly altered. This is a prevalent challenge for all benchmarks operating on the live web, |
| necessitating periodic manual reviews and updates to the dataset. |
|
|
| Questions, comments, or issues? Share your thoughts with us in the [discussion forum](https://www.kaggle.com/benchmarks/google/dsqa/discussion). |
|
|
| |
| The autorater which should be used for DeepSearchQA is `gemini-2.5-flash` with the grading prompt found in the [starter notebook](https://www.kaggle.com/code/andrewmingwang/deepsearchqa-starter-code) on Kaggle. Using a different autorater model or grading prompt will likely result in statistically significant deviation in results. |
|
|
| |
|
|
| Coming soon. |
