LiveCodeBench (rolling fresh-problem code benchmark)
Integrity rank #16 of 61 · 8 models scored · top score 65.9 · DeepSeek-R1
Competitive programming on problems released after typical training cutoffs
Strongest on discrimination, weakest on saturation headroom. Caveat: the public test set does not rule out training-data contamination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | DeepSeek-R1 | 65.9 | self-reported· optimizedT1 | 2025-01-22 |
| 2 | DeepSeek-R1-Distill-Llama-70B | 57.5 | self-reported· optimizedT1 | 2025-01-22 |
| 3 | DeepSeek-R1-Distill-Qwen-32B | 57.2 | self-reported· optimizedT1 | 2025-01-22 |
| 4 | DeepSeek-R1-Distill-Qwen-14B | 53.1 | self-reported· optimizedT1 | 2025-01-22 |
| 5 | DeepSeek-R1-Distill-Llama-8B | 39.6 | self-reported· optimizedT1 | 2025-01-22 |
| 6 | DeepSeek-R1-Distill-Qwen-7B | 37.6 | self-reported· optimizedT1 | 2025-01-22 |
| 7 | DeepSeek-V3 | 36.2 | self-reported· optimizedT1 | 2025-01-22 |
| 8 | DeepSeek-R1-Distill-Qwen-1.5B | 16.9 | self-reported· optimizedT1 | 2025-01-22 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark evaluates 8 models, spanning a score spread of 49.0 points between the highest and lowest results. This wide dispersion supports the reliability of rank differences, implying that performance gaps are substantial rather than marginal.
2 cited facts
The benchmark's health as a discriminator of frontier models under a disclosed harness is quantified by an integrity score of 98, yielding an integrity grade of A. It holds an integrity rank of 13 out of benchmark universe size benchmarks. Saturation is the weakest component, which means the ceiling is crowded for distinguishing top performers.
5 cited facts
The benchmark is not saturated, indicating that there is still genuine room to separate the field at the top. The current top score stands at 65.9, leaving a headroom of 34.1 points to the ceiling. A single model clusters near the top, meaning today's ranking among the leading entries remains distinct rather than collapsing into noise.
4 cited facts
This benchmark employs a consistent harness, meaning scores from different evaluation setups should not be compared head-to-head. Because the test set is public, high task performance under this disclosed harness deserves more scrutiny for possible contamination than a held-out set would, rather than reflecting deployed capability. Qualitative assessment indicates no known contamination history or specific harness sensitivity that would alter this interpretation.
3 cited facts
This benchmark measures competitive programming proficiency by evaluating models on problems released after typical training cutoffs, utilizing a rolling collection of fresh tasks to construct the test set. The design assumes a consistent evaluation timeline, yet the sharpest caveat is that the effective cutoff depends entirely on the release window chosen by the evaluator, which is rarely disclosed in published reports.
2 cited facts
LiveCodeBench (LiveCodeBench (rolling fresh-problem code benchmark)): LiveCodeBench — a continuously updated code-generation benchmark pulling fresh problems from Codeforces/LeetCode/AtCoder after typical training cutoffs, scored as pass@1 (%); designed to stay contamination-resistant as the collection window advances.
DeepSeek-R1 leads LiveCodeBench at 65.9 — self-reported by the lab. The full leaderboard above lists every recorded measurement, not just the headline number.
8 models have recorded scores on LiveCodeBench, spanning a score spread of 49.
tensor.news grades LiveCodeBench A for integrity (score 98/100), ranking #16 of 61 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — LiveCodeBench still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is public, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on LiveCodeBench are reasonably apples-to-apples.
Every LiveCodeBench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites arxiv.org/abs/2403.04132.