{
  "paper": {
    "arxiv_id": "2608.26105",
    "title": "VBVR-Pro"
  },
  "dataset": {
    "name": "VBVR-Pro procedural task specification",
    "tasks": 300
  },
  "setup": {
    "seed": 42,
    "modalities": [
      "image-state",
      "interleaved",
      "video-state"
    ]
  },
  "variants": {
    "scalar VLM-judge analogue": {
      "reward_mean": 0.5941666666666666,
      "reward_std": 0.3830895964247645
    },
    "task-grounded verifier": {
      "reward_mean": 1.0,
      "reward_std": 0.0
    }
  },
  "relative": {
    "reward_mean_percent": 68.30294530154279
  },
  "diagnostics": {
    "deterministic_scorers": 3,
    "judge_calls": 300,
    "verifier_calls": 300
  },
  "scope": "复刻 300 个程序化状态转移任务与确定性 task-specific scorer，对比易受流畅度影响的标量 judge；不训练或加载 14B/19B 视频生成器，也不冒充七个外部 benchmark 结果。",
  "diagnostic_only": true,
  "runtime": {
    "requested_device": "auto",
    "cpu_threads": null,
    "platform": "Darwin arm64"
  },
  "seed": 42,
  "schema_version": 2,
  "manifest": {
    "adapter_key": "vbvr-pro",
    "arxiv_id": "2608.26105",
    "title": "VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning",
    "paper_url": "https://arxiv.org/abs/2608.26105",
    "track": "llm",
    "organization": "Nanyang Technological University / VBVR Community",
    "published": "2026-08-26",
    "code_url": "https://www.video-reason.com/?v=pro",
    "topics": [
      "multimodal-foundation-model",
      "native-visual-reasoning",
      "verifiable-reward",
      "benchmark"
    ],
    "local_code_dir": "src/auto_research/reproductions/vbvr_pro",
    "fidelity": "core_mechanism",
    "evaluation_tier": "l1_mechanism",
    "datasets": [
      "VBVR-Pro procedural task specification"
    ],
    "baseline": "scalar VLM-as-a-judge analogue on the same 300 generated tasks",
    "metrics": [
      "reward mean",
      "reward standard deviation",
      "deterministic verifier agreement"
    ],
    "default_seeds": [
      42
    ],
    "budget": "adapter-defined fixed budget",
    "device_capabilities": [
      "cpu"
    ],
    "online_evidence": [],
    "selection_exception": null,
    "evolve_operators": []
  },
  "provenance": {
    "created_at": "2026-08-27T16:29:36.470882+00:00",
    "code_commit": "cfb01264c6575f3d1634226940df384a3632a036",
    "python": "3.14.5",
    "platform": "macOS-26.5.2-arm64-arm-64bit-Mach-O",
    "dataset_dir": "/Users/bytedance/Documents/git_daiwk/auto-research/data",
    "dataset_fingerprint": "e66c261317a7b3179720ef3e3f5f79d7a204b135d1f217e2911b153ad2ce50a0",
    "packages": {
      "numpy": "2.4.3"
    },
    "artifact_path": "docs/reproductions/2608.26105-vbvr-pro/metrics/procedural-seed42.json"
  },
  "evaluation_protocol": {
    "tier": "l1_mechanism",
    "tier_label": "L1 核心机制 mini-suite",
    "seeds": [
      42
    ],
    "budget": "paper-specific",
    "formal_comparison": false,
    "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
  },
  "reproduction_fidelity": {
    "level": "core_mechanism",
    "label": "核心机制复现",
    "description": "论文中心算法被实际执行，但生产模型、私有特征或基础设施未复刻。",
    "omitted_core_components": [
      "released large video/image generators",
      "seven external transfer benchmarks",
      "large-scale multi-task RL"
    ]
  },
  "manifest_ref": "reproduction:vbvr-pro"
}