LongBench-Pro
Realistic bilingual long-context benchmark with 1,500 natural samples across 11 primary and 25 secondary tasks from 8k to 256k tokens.
Newer than LongBench v1/v2 and designed for more realistic, bilingual, fine-grained long-context analysis across 8k-256k token inputs.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
| Model | Score | Date | Source Type | Provenance |
|---|---|---|---|---|
| Gemini-2.5-Pro (paper Overall) | 73.42% | 2024-06-01 | independent | Source ↗ |
| GPT-5 (paper Overall) | 72.61% | 2024-06-01 | independent | Source ↗ |
| Qwen3-235B-A22B-Thinking-2507 (paper Overall) | 66.97% | 2024-06-01 | independent | Source ↗ |
| Qwen3-Next-80B-A3B-Thinking (paper Overall) | 63.95% | 2024-06-01 | independent | Source ↗ |
| DeepSeek-R1-0528 (paper Overall) | 61.89% | 2024-06-01 | independent | Source ↗ |
| DeepSeek-R1 (paper Overall) | 60.07% | 2024-06-01 | independent | Source ↗ |
| Qwen3-30B-A3B-Thinking-2507 (paper Overall) | 59.68% | 2024-06-01 | independent | Source ↗ |
| Claude-4-Sonnet (paper Overall) | 56.07% | 2024-06-01 | independent | Source ↗ |
| Gemini-2.5-Flash (paper Overall) | 55.92% | 2024-06-01 | independent | Source ↗ |
| MiniMax-M2 (paper Overall) | 53.21% | 2024-06-01 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.Measured human performance on professional long-context document analysis.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.normalized task-specific aggregateDataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.$5 – $20 USD for full benchmark evaluation run on frontier APIs.
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.lm_eval --model hf --model_args pretrained=<model_path> --tasks longbench-pro --batch_size autoopencompass --datasets longbench-pro --models <model_config># Standard API Evaluation Loop
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": prompt}],
temperature=0.0,
)When publishing results for LongBench-Pro, always report the exact prompt template, few-shot exemplar ordering, sampling temperature (temperature=0), maximum reasoning budget tokens, and the precise timestamped model snapshot ID.
Contamination & Memorization Analysis
Audit of pretraining exposure risks, memorization vectors, and refresh policies.Static fixed snapshot
Public on web / HuggingFace