CL-bench: A Benchmark for Context Learning
Integrity rank #45 of 65 · 19 models scored · top score 27.9 · GPT-5.4
Context learning - acquiring and applying novel knowledge presented only in-context
Strongest on harness comparability, weakest on discrimination. 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 | GPT-5.4 | 27.9 | unverifiedT2 | 2026-03-05 |
| 2 | GPT-5.1 | 23.7 | unverifiedT2 | 2025-11-13 |
| 3 | Grok 4.20 | 22.2 | unverifiedT2 | 2026-02-17 |
| 4 | Claude Opus 4.5 | 21.1 | unverifiedT2 | 2025-11-24 |
| 5 | Gemini 3.1 Pro | 20.8 | unverifiedT2 | 2026-02-19 |
| 6 | Claude Opus 4.6 | 20.7 | unverifiedT2 | 2026-02-05 |
| 7 | Qwen 3.6 Plus | 20.3 | unverifiedT2 | 2026-03-31 |
| 8 | Qwen 3.5 Plus (hosted 397B-A17B) | 19.8 | unverifiedT2 | 2026-02-16 |
| 9 | Kimi K2.5 | 19.3 | unverifiedT2 | 2026-01-27 |
| 10 | GLM-5 | 18.7 | unverifiedT2 | 2026-02-11 |
| 11 | GPT-5.2 | 18.2 | unverifiedT2 | 2025-12-11 |
| 12 | o3 | 17.8 | unverifiedT2 | 2025-04-16 |
| 13 | Kimi K2 Thinking | 17.6 | unverifiedT2 | 2025-11-06 |
| 14 | GLM-4.7 | 15.9 | unverifiedT2 | 2025-12-22 |
| 15 | Gemini 3 Pro | 15.8 | unverifiedT2 | 2025-11-18 |
| 16 | Qwen3-Max | 14.5 | unverifiedT2 | 2025-09-24 |
| 17 | DeepSeek-V3.2-Exp | 13.2 | unverifiedT2 | 2025-09-29 |
| 18 | DeepSeek-V3.2 | 12.4 | unverifiedT2 | 2025-12-01 |
| 19 | MiniMax-M2.5 | 11.4 | unverifiedT2 | 2026-02-12 |
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 scores 19 models, and the gap between its best and worst scores spans 16.5 points. A spread that wide supports reading the ordering as real separation, since rank differences of that magnitude are unlikely to be an artifact of close scoring.
3 cited facts
This benchmark’s Benchmark Integrity Index is 76 (grade B), ranking 42 out of 62 benchmarks. The weakest component in its integrity breakdown is discrimination, which limits how sharply the benchmark separates frontier models and therefore how decisively its scores can order them. This score should be read as the benchmark’s health as a discriminator of frontier models under a disclosed harness, not as any model’s capability.
6 cited facts
This benchmark is not saturated: the top score reaches 27.9, leaving 72.1 points of headroom to the ceiling, and only one model clusters near the top within a couple of points. Because the benchmark is not saturated, there is still genuine room to separate the field at the top, so differences among the leading models are more likely to reflect real capability gaps than noise.
5 cited facts
For this benchmark, scores are reported as task performance under a disclosed harness, and the harness is consistent, so results are directly comparable with one another under that setup rather than across differing evaluation harnesses. The test set is public, which means strong scores deserve more scrutiny for possible contamination than results from a held-out set would, even though the benchmark deliberately uses knowledge absent from pretraining to resist contamination. Grading is rubric-based and likely model-assisted, so harness sensitivity to judge behavior should be kept in mind when interpreting scores.
6 cited facts
This benchmark is designed to measure context learning: the ability to acquire and apply novel knowledge supplied only within the prompt rather than relying on pretrained knowledge. Its test items were built from expert-authored novel contexts paired with many associated tasks and rubric-based verification checks across multiple categories and sub-categories, with the knowledge deliberately held out from pretraining. The design assumes that expert-authored novel contexts plus rubric grading isolate genuine context learning from retrieval or recall, and that the rubrics fully capture task success. The sharpest caveat is that all-rubrics-must-pass scoring is strict, rubric grading often judged by a language model can be noisy, and constructed novelty may still overlap with training distributions; the human baseline is not documented.
5 cited facts
CL-bench (CL-bench: A Benchmark for Context Learning): CL-bench (A Benchmark for Context Learning) tests whether models can absorb and apply genuinely novel in-context knowledge (fictional legal systems, invented game rules, novel financial instruments, empirically derived laws) rather than recall pretrained facts. Average solving rate approx 17.2%; best model GPT-5.1 approx 23.7%.
GPT-5.4 leads CL-bench at 27.9. The full leaderboard above lists every recorded measurement, not just the headline number.
19 models have recorded scores on CL-bench, spanning a score spread of 16.5.
tensor.news grades CL-bench B for integrity (score 76/100), ranking #45 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — CL-bench 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 CL-bench are reasonably apples-to-apples.
Every CL-bench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites arxiv.org/abs/2602.03587.