CL-bench Life
Integrity rank #40 of 61 · 13 models scored · top score 22.2 · GPT-5.5
Context learning over messy, real-life, unstructured everyday context
Strongest on contamination resistance, 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-5.5 | 22.2 | unverifiedT2 | 2026-04-23 |
| 2 | GPT-5.4 | 21.7 | unverifiedT2 | 2026-03-05 |
| 3 | GPT-5.1 | 17.3 | unverifiedT2 | 2025-11-13 |
| 4 | Claude Opus 4.6 | 17 | unverifiedT2 | 2026-02-05 |
| 5 | Gemini 3.1 Pro | 16.9 | unverifiedT2 | 2026-02-19 |
| 6 | DeepSeek-V4-Pro | 13.5 | unverifiedT2 | 2026-04-24 |
| 7 | Kimi K2.5 | 13.2 | unverifiedT2 | 2026-02-02 |
| 8 | Qwen 3.5 Plus (hosted 397B-A17B) | 12.4 | unverifiedT2 | 2026-02-16 |
| 9 | Grok 4.20 | 11.9 | unverifiedT2 | 2026-02-17 |
| 10 | GLM-4.7 | 10.9 | unverifiedT2 | 2025-12-22 |
| 11 | DeepSeek-V3.2-Exp | 9.5 | unverifiedT2 | 2025-09-29 |
| 12 | DeepSeek-V3.2 | 7.4 | unverifiedT2 | 2025-12-01 |
| 13 | MiniMax-M2.5 | 6.3 | 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 is not saturated, indicating there is still substantial room for models to differentiate themselves at the top. With a score headroom of approximately 70 points, the current top score leaves ample space for improvement or further differentiation among models. A small cluster of models near the top score suggests that while the field is close, there is still discernible separation among leading models.
3 cited facts
This benchmark evaluates how well models learn from messy, real-life, unstructured everyday context, using 405 expert-curated context-task pairs with 5,348 verification rubrics across three categories of everyday information settings. Rubric grading over messy real-life contexts is a proxy rather than a measurement, and the claim that three categories span everyday information settings is the design's leap of faith. Read scores as conditional on both. A key caveat is that the benchmark's curated 'real-life' scenarios, small scale, strict all-rubric scoring, and judge/rubric noise limit what it can tell us, and the human baseline is not documented.
6 cited facts
CL-bench Life: CL-bench Life is a companion to CL-bench that shifts from clean professional sources to messy real-life context - everyday communication, fragmented notes, and behavioral traces that are socially grounded and temporally dispersed - testing whether models can learn from and reason over unstructured everyday information.
GPT-5.5 leads CL-bench Life at 22.2. The full leaderboard above lists every recorded measurement, not just the headline number.
13 models have recorded scores on CL-bench Life, spanning a score spread of 15.9.
tensor.news grades CL-bench Life B for integrity (score 81/100), ranking #40 of 61 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — CL-bench Life 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 Life are reasonably apples-to-apples.
Every CL-bench Life measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites arxiv.org/abs/2604.27043.