| --- |
| dataset_info: |
| - config_name: imaginary-reference |
| features: |
| - name: role |
| dtype: string |
| - name: content |
| dtype: string |
| splits: |
| - name: test |
| num_bytes: 4485 |
| num_examples: 25 |
| download_size: 4391 |
| dataset_size: 4485 |
| - config_name: indifferent |
| features: |
| - name: role |
| dtype: string |
| - name: content |
| dtype: string |
| splits: |
| - name: test |
| num_bytes: 11732 |
| num_examples: 25 |
| download_size: 10536 |
| dataset_size: 11732 |
| - config_name: math |
| features: |
| - name: role |
| dtype: string |
| - name: content |
| dtype: string |
| splits: |
| - name: test |
| num_bytes: 5440 |
| num_examples: 25 |
| download_size: 4740 |
| dataset_size: 5440 |
| - config_name: redundant |
| features: |
| - name: role |
| dtype: string |
| - name: content |
| dtype: string |
| splits: |
| - name: test |
| num_bytes: 5087 |
| num_examples: 25 |
| download_size: 4096 |
| dataset_size: 5087 |
| - config_name: unanswerable |
| features: |
| - name: role |
| dtype: string |
| - name: content |
| dtype: string |
| splits: |
| - name: test |
| num_bytes: 12501 |
| num_examples: 50 |
| download_size: 8242 |
| dataset_size: 12501 |
| configs: |
| - config_name: imaginary-reference |
| data_files: |
| - split: test |
| path: imaginary-reference/test-* |
| - config_name: indifferent |
| data_files: |
| - split: test |
| path: indifferent/test-* |
| - config_name: math |
| data_files: |
| - split: test |
| path: math/test-* |
| - config_name: redundant |
| data_files: |
| - split: test |
| path: redundant/test-* |
| - config_name: unanswerable |
| data_files: |
| - split: test |
| path: unanswerable/test-* |
| license: cc-by-nc-4.0 |
| language: |
| - en |
| --- |
| # DNR Bench |
|
|
| Don’t Reason Bench (DNR Bench), a novel benchmark designed to expose a vulnerability in current RLMs: their tendency to over-reason by attempting to solve unsolvable |
| problems, leading to excessively long responses. |
|
|
| # Data Summary |
| The DNR Bench dataset contains 150 adversarially crafted prompts divided into five distinct categories: |
| - Imaginary Reference |
| - Indifferent |
| - Math, |
| - Redundant, |
| - Unanswerable. |
|
|
| Each category targets a specific failure mode observed in reasoning-optimized LLMs, such as hallucinating nonexistent references, failing to remain neutral in ambiguous contexts, incorrectly solving flawed math problems, overanalyzing redundant information, or answering questions that lack sufficient data. |
|
|
| # Leaderboard |
| This dataset is used to test reasoning LLMs in [DNR Leaderboard on Huggingface](https://huggingface.co/spaces/ServiceNow-AI/Do-not-reason-bench) |
|
|
|
|
| # Citation |
| ```bibtex |
| @misc{hashemi2025dnrbenchbenchmarkingoverreasoning, |
| title={DNR Bench: Benchmarking Over-Reasoning in Reasoning LLMs}, |
| author={Masoud Hashemi and Oluwanifemi Bamgbose and Sathwik Tejaswi Madhusudhan and Jishnu Sethumadhavan Nair and Aman Tiwari and Vikas Yadav}, |
| year={2025}, |
| eprint={2503.15793}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.LG}, |
| url={https://arxiv.org/abs/2503.15793}, |
| } |
| ``` |