{
  "main": [
    {
      "title": "Visual reasoning",
      "metric": "7-benchmark mean accuracy (%)",
      "labels": [
        "Qwen2.5-VL-7B",
        "Mirage",
        "Uni-LaDiR"
      ],
      "roles": [
        "Base model",
        "Modality-specific",
        "Unified"
      ],
      "values": [
        64.6,
        72.2,
        78.5
      ]
    },
    {
      "title": "Visual math & logic",
      "metric": "4-benchmark mean accuracy (%)",
      "labels": [
        "Qwen2.5-VL-7B",
        "Mirage",
        "Uni-LaDiR"
      ],
      "roles": [
        "Base model",
        "Modality-specific",
        "Unified"
      ],
      "values": [
        36.4,
        40.7,
        49.0
      ]
    },
    {
      "title": "Robot manipulation",
      "metric": "RLBench mean success (%)",
      "labels": [
        "OpenVLA (7B)",
        "LaST₀ (3.3B)",
        "Uni-LaDiR (3.3B)"
      ],
      "roles": [
        "Direct policy",
        "Modality-specific",
        "Unified"
      ],
      "values": [
        40.0,
        82.0,
        87.0
      ]
    }
  ],
  "sharing": [
    {
      "title": "Text + image teachers",
      "metric": "4-benchmark mean accuracy (%)",
      "labels": [
        "Separate\nspaces",
        "Text + image\nshared"
      ],
      "values": [
        63.0,
        70.1
      ],
      "gain": "+11.3% relative"
    },
    {
      "title": "Visual + 3D + robot state teachers",
      "metric": "RLBench mean success (%)",
      "labels": [
        "Separate",
        "Visual\n+ 3D",
        "Visual\n+ robot state",
        "3D\n+ robot state",
        "All three"
      ],
      "values": [
        82,
        83.5,
        84.5,
        84.5,
        87
      ],
      "gain": "+6.1% relative"
    }
  ],
  "source": {
    "version": "arXiv:2609.19878v3",
    "date": "2026-09-29",
    "url": "https://arxiv.org/abs/2609.19878v3",
    "main": "Tables 1–2: three named methods per panel. Mirage is the same modality-specific baseline in both VLM groups. OpenVLA is a direct-policy comparator, not an ablation of the Uni-LaDiR backbone. All numbers are reported table values.",
    "sharing": "Figure 3 (rounded scores)",
    "protocol": "Section 4.3 and Appendix B.3: fixed backbone, data, teacher supervision and thought-token dimensions; encoder parameter counts are not matched"
  },
  "paper_figures": {
    "fig1": {
      "source_file": "fig1-0912-motion-detail-v76.pdf",
      "sha256": "15dddc7664553dde14e7ec6cc9e791e0ad9d85c50c33c16d1e66a486c8b7ab9b"
    },
    "fig2": {
      "source_file": "method-overview-v45-reference.pdf",
      "sha256": "f7a50eb424aaa723809a78fb179e88fd1270b1231c708c31873a77e442f938e1"
    },
    "paper-sharing": {
      "source_file": "ablation-sharing-v4.pdf",
      "sha256": "537b73f6963b5320eb1ed50e8525cc2f283e05195a45fa4608bfdcf415c7b29e"
    },
    "objective": {
      "source_file": "ablation-objective-v9-clean-labels.pdf",
      "sha256": "9d89f7438c322b5abb1974634615f9a401ce55a1969860ba65fee82fd021a336"
    },
    "joint-training": {
      "source_file": "ablation-training-v4-staged-joint.pdf",
      "sha256": "97e5744dd2ba49854c60fea310cf4b62b26ba26524f2e9d68b88b3709e0ac984"
    },
    "diffusion": {
      "source_file": "ablation-diffusion.pdf",
      "sha256": "07f7a4cb00bea04e05d42a7de1dc8c4b16b4db231b2f6ef16c44b78946fdca56"
    }
  },
  "interventions": {
    "source": "v3 Table 7",
    "conditions": [
      "Intact",
      "Zeroed",
      "Shuffled",
      "Cross-example"
    ],
    "V*": [
      89.5,
      55.8,
      65.8,
      50.8
    ],
    "RLBench": [
      87,
      55,
      65,
      51
    ]
  },
  "ablation_summary": {
    "source": "v3 Section 4.4, Figures 4–6",
    "continuation_vs_reconstruction_unified_mean_pp": 9.2,
    "joint_vs_separate_stages_mean_relative_pct": 12.7,
    "flow_matching_vs_L2_mean_relative_pct": 16.3,
    "flow_matching_vs_cosine_mean_relative_pct": 21.0,
    "note": "Reported aggregates use unrounded source scores; they are not reconstructed from plotted rounded values."
  }
}