ArxivRollBench#

Overview#

ArxivRollBench is a rolling benchmark built from recent arXiv papers. It evaluates whether large language models can reason over fresh scientific text through three task formats: sequencing, cloze, and next-fragment prediction.

Task Description#

Task Type: Multiple-choice scientific text reasoning
Input: Recent arXiv text fragments with four answer choices
Output: Single correct answer letter (A, B, C, or D)
Domains: Computer Science, Quantitative Finance, Mathematics, Physics, Statistics, Quantitative Biology, Economics, and Electrical Engineering/System Science
Releases: 2024b, 2025a, and 2026a rolling snapshots

Key Features#

Time-aware benchmark snapshots reduce contamination-related overestimation
Covers multiple arXiv domains and scientific writing styles
Includes sequencing, cloze, and prediction formats under the SCP framework
Compact -50 split is suitable for cost-controlled API evaluation
Full split is available as arxivrollbench_full

Evaluation Notes#

Default configuration uses 0-shot evaluation
The default arxivrollbench benchmark uses compact -50 datasets
Use arxivrollbench_full for the complete public splits
Each subset is loaded from the public ModelScope mirror under the liangzid namespace
Answers are normalized to A-D and evaluated with accuracy

Properties#

Property	Value
Benchmark Name	`arxivrollbench`
Dataset ID	liangzid/arxivrollbench
Paper	Paper
Tags	`Knowledge`, `MCQ`, `Reasoning`
Metrics	`acc`
Default Shots	0-shot
Evaluation Split	`train`

Data Statistics#

Metric	Value
Total Samples	3,254
Prompt Length (Mean)	1514.19 chars
Prompt Length (Min/Max)	307 / 14112 chars

Per-Subset Statistics:

Subset	Samples	Prompt Mean	Prompt Min	Prompt Max
`2024b_cs_s`	42	949.6	590	1805
`2024b_cs_c`	31	307	307	307
`2024b_cs_p`	50	2617.32	922	7512
`2024b_q_fin_s`	49	1042.31	586	2329
`2024b_q_fin_c`	44	307	307	307
`2024b_q_fin_p`	50	3430.52	872	9106
`2024b_math_s`	34	829.85	593	2115
`2024b_math_c`	15	307	307	307
`2024b_math_p`	51	1957.24	869	6260
`2024b_physics_s`	45	957.11	576	4402
`2024b_physics_c`	28	307	307	307
`2024b_physics_p`	51	2948.1	885	13643
`2024b_stat_s`	45	936.4	582	1678
`2024b_stat_c`	33	307	307	307
`2024b_stat_p`	50	2946.44	861	7026
`2024b_q_bio_s`	43	975	583	2555
`2024b_q_bio_c`	34	307	307	307
`2024b_q_bio_p`	49	3354.53	883	8867
`2024b_econ_s`	48	1021.58	586	2070
`2024b_econ_c`	43	307	307	307
`2024b_econ_p`	50	3257.76	846	8967
`2024b_eess_s`	48	1034.56	574	2922
`2024b_eess_c`	42	307	307	307
`2024b_eess_p`	51	2612.69	882	8609
`2025a_cs_s`	50	921.2	592	1632
`2025a_cs_c`	44	307	307	307
`2025a_cs_p`	51	2895.02	942	6540
`2025a_q_fin_s`	50	931.08	589	2202
`2025a_q_fin_c`	43	307	307	307
`2025a_q_fin_p`	51	2837.86	793	7577
`2025a_math_s`	42	852.52	580	1595
`2025a_math_c`	28	307	307	307
`2025a_math_p`	51	2449.49	889	6893
`2025a_physics_s`	44	939.32	587	1874
`2025a_physics_c`	34	307	307	307
`2025a_physics_p`	49	3568.29	1001	9325
`2025a_stat_s`	48	932.81	600	2063
`2025a_stat_c`	42	307	307	307
`2025a_stat_p`	50	3115.36	822	7349
`2025a_q_bio_s`	49	1074.12	591	1810
`2025a_q_bio_c`	49	307	307	307
`2025a_q_bio_p`	50	3639.26	1038	8890
`2025a_econ_s`	48	982.19	591	2322
`2025a_econ_c`	45	307	307	307
`2025a_econ_p`	51	2860.9	884	6494
`2025a_eess_s`	46	1017.35	588	1807
`2025a_eess_c`	42	307	307	307
`2025a_eess_p`	50	3541.1	943	14112
`2026a_cs_s`	51	944.12	584	1795
`2026a_cs_c`	38	307	307	307
`2026a_cs_p`	51	2629.06	919	5234
`2026a_q_fin_s`	48	1025.44	608	2320
`2026a_q_fin_c`	45	307	307	307
`2026a_q_fin_p`	51	3094.78	872	6644
`2026a_math_s`	44	844.05	575	1381
`2026a_math_c`	30	307	307	307
`2026a_math_p`	51	2160.27	860	12385
`2026a_physics_s`	47	1082.04	599	2522
`2026a_physics_c`	41	307	307	307
`2026a_physics_p`	50	3420.58	894	8788
`2026a_stat_s`	49	1013.47	575	2482
`2026a_stat_c`	46	307	307	307
`2026a_stat_p`	51	2564.47	955	6387
`2026a_q_bio_s`	47	1019.7	584	1707
`2026a_q_bio_c`	40	307	307	307
`2026a_q_bio_p`	48	3030.71	954	6468
`2026a_econ_s`	48	989.67	580	2320
`2026a_econ_c`	47	307	307	307
`2026a_econ_p`	51	2920.76	885	7061
`2026a_eess_s`	51	988.14	579	2231
`2026a_eess_c`	45	307	307	307
`2026a_eess_p`	51	2812.61	922	5589

Sample Example#

Subset: 2024b_cs_s

{
  "input": [
    {
      "id": "7b220fd6",
      "content": "Answer the following ArxivRollBench multiple choice question. The entire content of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of A,B,C,D.\n\nSelect the option that correctly compl ... [TRUNCATED 381 chars] ... m a diagonal matrix into the identity, allows us to write the input matrix as a product of transvections. **C**: Note that row and column operations are effected by left- and right multiplications by transvections\n\nA) BAC\nB) ABC\nC) ACB\nD) BCA"
    }
  ],
  "choices": [
    "BAC",
    "ABC",
    "ACB",
    "BCA"
  ],
  "target": "C",
  "id": 0,
  "group_id": 0,
  "metadata": {
    "original_label": "Selection 3",
    "task_type": "s/c"
  }
}

Prompt Template#

Prompt Template:

Answer the following ArxivRollBench multiple choice question. The entire content of your response should be of the following format: 'ANSWER: [LETTER]' (without quotes) where [LETTER] is one of {letters}.

{question}

{choices}

Usage#

Using CLI#

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets arxivrollbench \
    --limit 10  # Remove this line for formal evaluation

Using Python#

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['arxivrollbench'],
    dataset_args={
        'arxivrollbench': {
            # subset_list: ['2024b_cs_s', '2024b_cs_c', '2024b_cs_p']  # optional, evaluate specific subsets
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)