{
 "body": "\n## What it measures\n\nCL-bench contains expert-crafted contexts in which the information required for each task is contained in the context and absent from pretraining. It spans new domain knowledge, rule systems, procedures and empirical laws.\n\n## How it is scored\n\nThe paper reports task solving rate: frontier models averaged 17.2% and its best reported model 23.7%. Preserve the rubric and task aggregation protocol.\n\n## Dataset and licence\n\nThe primary source establishes the benchmark release, but the opened materials do not establish a single dataset licence. Confirm the current release terms and split accounting before redistribution.\n\n## Who publishes it\n\nCL-bench is introduced in the 2026 arXiv paper; a complete author list was not established from the opened abstract.\n\n## Lineage\n\nNo predecessor or successor was established in the opened primary source.\n\n## Saturation and contamination\n\nThe benchmark materials are public, which creates contamination opportunities. The opened source does not establish a contamination audit or current saturation ceiling.\n\n## How to run it\n\nFollow the primary paper or release protocol, recording the exact model, prompts, evaluator, task version, tool access and timeout. Preserve per-task outcomes when comparing runs.\n\n## Reading the numbers\n\nHigher scores indicate more successful tasks under the selected protocol. Results can depend on evaluator models, prompts, environment setup and aggregation, so compare only matched configurations.\n\n## Protocol cautions\n\nKeep the task set, context, instructions and evaluator fixed. Report domain-level and difficulty-level results whenever the release defines them, because aggregates can hide systematic failures.\n\nThe benchmark should be treated as a protocol rather than a single model score. Store the exact release revision, context or repository snapshot, prompt, tool permissions, timeout, evaluator version and task-level outcomes. If a task depends on external services or packages, record those dependencies and distinguish an infrastructure failure from an agent failure. Aggregate scores are useful for a headline comparison, while per-domain, per-difficulty and per-task results reveal which capabilities remain unreliable. Public examples and reference material also create opportunities for contamination, so report whether the evaluated model or agent had access to the benchmark before the run.\nThese records therefore leave unspecified details explicitly unknown until the release documentation provides them.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "cl_bench_a_benchmark_for_context_learning",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [],
 "page": {
  "aliases": [],
  "category": "long-context",
  "contamination": {
   "note": "The benchmark materials are public; the opened source does not establish a contamination audit or private rotating holdout.",
   "risk": "medium"
  },
  "dataset": {
   "languages": [
    "en"
   ],
   "license": "",
   "modalities": [
    "text"
   ],
   "public_test_set": true,
   "size": 500,
   "size_note": "The paper reports 500 contexts, 1,899 tasks and 31,607 verification rubrics.",
   "splits": "Unknown unless specified by the official release.",
   "url": "https://arxiv.org/abs/2602.03587"
  },
  "freshness": {
   "researched": "2026-09-09",
   "researched_by": "GPT-5.6 Luna, luna-stream-c-007 (Codex coordinated)",
   "reviewed": "",
   "reviewed_by": ""
  },
  "harness": {
   "bigbench": "",
   "helm": "",
   "inspect_evals": "",
   "lm_eval": "",
   "opencompass": "",
   "other": "Use the primary release protocol and record evaluator settings."
  },
  "id": "cl_bench_a_benchmark_for_context_learning",
  "last_updated": "",
  "leaderboard_url": "",
  "lineage": {
   "family": "context learning",
   "predecessor": "",
   "successors": [],
   "variants": []
  },
  "measures": "Context-grounded task solving beyond retrieval or simple in-context pattern learning.",
  "metric": {
   "baseline_note": "No universal random or human baseline was established in the opened primary source.",
   "direction": "higher_is_better",
   "human_baseline": null,
   "max_score": 100,
   "name": "task success rate",
   "random_baseline": null,
   "unit": "%"
  },
  "name": "CL-bench",
  "page_kind": "benchmark",
  "paper": {
   "arxiv": "2602.03587",
   "title": "A Benchmark for Context Learning",
   "url": "https://arxiv.org/abs/2602.03587",
   "year": 2026
  },
  "publisher": {
   "authors": [],
   "org": "CL-bench authors",
   "url": "https://arxiv.org/abs/2602.03587"
  },
  "released": "2026-02",
  "repo_url": "",
  "saturation": {
   "as_of": "",
   "note": "No current saturation ceiling was established in the opened source.",
   "status": "unknown",
   "top_score": null
  },
  "sources": [
   {
    "accessed": "2026-09-09",
    "title": "Primary paper",
    "url": "https://arxiv.org/abs/2602.03587"
   }
  ],
  "status": "active",
  "subcategory": "context learning",
  "summary": "CL-bench tests whether models can learn new domain knowledge, rules and procedures from complex contexts.",
  "tags": [
   "context",
   "reasoning"
  ],
  "task_format": "500 complex contexts, 1,899 tasks and 31,607 verification rubrics."
 }
}