{
  "schema_version": 2,
  "seeds_by_adapter": {
    "lngram-v2": [
      42,
      43,
      44
    ]
  },
  "budget": "paper-specific",
  "papers": {
    "lngram-v2": {
      "manifest": {
        "adapter_key": "lngram-v2",
        "arxiv_id": "2609.03426",
        "title": "Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations",
        "paper_url": "https://arxiv.org/abs/2609.03426",
        "track": "llm",
        "organization": "Beijing University of Posts and Telecommunications / Kuaishou Technology",
        "published": "2026-09-03",
        "code_url": null,
        "topics": [
          "multimodal-foundation-model",
          "conditional-memory",
          "discrete-routing",
          "model-architecture"
        ],
        "local_code_dir": "src/auto_research/reproductions/lngram_v2",
        "fidelity": "core_mechanism",
        "evaluation_tier": "l2_public_dataset",
        "datasets": [
          "Qwen2.5-VL-3B-Instruct checkpoint + deterministic RGB geometry probe"
        ],
        "baseline": "same checkpoint without latent memory branch",
        "metrics": [
          "finite output",
          "route diversity",
          "sink selectivity",
          "activation memory"
        ],
        "default_seeds": [
          42
        ],
        "budget": "adapter-defined fixed budget",
        "device_capabilities": [
          "cuda"
        ],
        "requires_gpu_validation": true,
        "gpu_validation_artifact": "docs/gpu-validations/lngram-v2-a100-20260906.json",
        "online_evidence": [],
        "selection_exception": null,
        "evolve_operators": [
          "memory:latent-ngram-gqa"
        ]
      },
      "seed_results": [
        {
          "paper": {
            "arxiv_id": "2609.03426",
            "title": "Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations",
            "url": "https://arxiv.org/abs/2609.03426",
            "organization": "Beijing University of Posts and Telecommunications / Kuaishou Technology"
          },
          "dataset": {
            "name": "deterministic cross-modal hidden-state fixture",
            "tokens": 64
          },
          "setup": {
            "adapter": "lngram-v2",
            "seed": 42
          },
          "method": {
            "finite_output": true,
            "sink_weight": 0.19995570182800293,
            "unique_route_ids": 8,
            "route_gradient_norm": 8.428582805208862e-05
          },
          "stages": {
            "hard_discrete_forward": true,
            "counterfactual_surrogate": true,
            "gqa_zero_sink": true,
            "activated_memory_parameters": 128
          },
          "paper_results": {
            "parameter_reduction_percent": 82.6,
            "activated_parameter_reduction_percent": 95.2,
            "id_semantic_recovery_min_percent": 65.77,
            "id_semantic_recovery_max_percent": 84.27
          },
          "scope": "CPU fixture executes hard addressing, exact n-gram lookup, GQA sink and surrogate gradients; A100 receipt validates real VLM hidden states.",
          "manifest_ref": "reproduction:lngram-v2",
          "runtime": {
            "requested_device": "auto",
            "cpu_threads": null,
            "platform": "Darwin arm64",
            "resolved_device": "cpu",
            "torch_version": "2.13.0",
            "accelerator": "arm"
          },
          "seed": 42,
          "schema_version": 2,
          "manifest": {
            "adapter_key": "lngram-v2",
            "arxiv_id": "2609.03426",
            "title": "Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations",
            "paper_url": "https://arxiv.org/abs/2609.03426",
            "track": "llm",
            "organization": "Beijing University of Posts and Telecommunications / Kuaishou Technology",
            "published": "2026-09-03",
            "code_url": null,
            "topics": [
              "multimodal-foundation-model",
              "conditional-memory",
              "discrete-routing",
              "model-architecture"
            ],
            "local_code_dir": "src/auto_research/reproductions/lngram_v2",
            "fidelity": "core_mechanism",
            "evaluation_tier": "l2_public_dataset",
            "datasets": [
              "Qwen2.5-VL-3B-Instruct checkpoint + deterministic RGB geometry probe"
            ],
            "baseline": "same checkpoint without latent memory branch",
            "metrics": [
              "finite output",
              "route diversity",
              "sink selectivity",
              "activation memory"
            ],
            "default_seeds": [
              42
            ],
            "budget": "adapter-defined fixed budget",
            "device_capabilities": [
              "cuda"
            ],
            "requires_gpu_validation": true,
            "gpu_validation_artifact": "docs/gpu-validations/lngram-v2-a100-20260906.json",
            "online_evidence": [],
            "selection_exception": null,
            "evolve_operators": [
              "memory:latent-ngram-gqa"
            ]
          },
          "provenance": {
            "created_at": "2026-09-05T19:39:00.360488+00:00",
            "code_commit": "6d97d3b2028b651e46229d2596e866d2a6f567c9",
            "python": "3.12.9",
            "platform": "macOS-26.5.2-arm64-arm-64bit",
            "dataset_dir": "/Users/bytedance/Documents/git_daiwk/auto-research/data",
            "dataset_fingerprint": "e66c261317a7b3179720ef3e3f5f79d7a204b135d1f217e2911b153ad2ce50a0",
            "packages": {
              "auto-research": "0.1.0",
              "numpy": "2.5.2",
              "torch": "2.13.0",
              "transformers": "5.14.1"
            }
          },
          "evaluation_protocol": {
            "tier": "l2_public_dataset",
            "tier_label": "L2 公开数据集训练",
            "seeds": [
              42
            ],
            "budget": "paper-specific",
            "formal_comparison": false,
            "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
          }
        },
        {
          "paper": {
            "arxiv_id": "2609.03426",
            "title": "Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations",
            "url": "https://arxiv.org/abs/2609.03426",
            "organization": "Beijing University of Posts and Telecommunications / Kuaishou Technology"
          },
          "dataset": {
            "name": "deterministic cross-modal hidden-state fixture",
            "tokens": 64
          },
          "setup": {
            "adapter": "lngram-v2",
            "seed": 43
          },
          "method": {
            "finite_output": true,
            "sink_weight": 0.19999940693378448,
            "unique_route_ids": 8,
            "route_gradient_norm": 7.882686622906476e-05
          },
          "stages": {
            "hard_discrete_forward": true,
            "counterfactual_surrogate": true,
            "gqa_zero_sink": true,
            "activated_memory_parameters": 128
          },
          "paper_results": {
            "parameter_reduction_percent": 82.6,
            "activated_parameter_reduction_percent": 95.2,
            "id_semantic_recovery_min_percent": 65.77,
            "id_semantic_recovery_max_percent": 84.27
          },
          "scope": "CPU fixture executes hard addressing, exact n-gram lookup, GQA sink and surrogate gradients; A100 receipt validates real VLM hidden states.",
          "manifest_ref": "reproduction:lngram-v2",
          "runtime": {
            "requested_device": "auto",
            "cpu_threads": null,
            "platform": "Darwin arm64",
            "resolved_device": "cpu",
            "torch_version": "2.13.0",
            "accelerator": "arm"
          },
          "seed": 43,
          "schema_version": 2,
          "manifest": {
            "adapter_key": "lngram-v2",
            "arxiv_id": "2609.03426",
            "title": "Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations",
            "paper_url": "https://arxiv.org/abs/2609.03426",
            "track": "llm",
            "organization": "Beijing University of Posts and Telecommunications / Kuaishou Technology",
            "published": "2026-09-03",
            "code_url": null,
            "topics": [
              "multimodal-foundation-model",
              "conditional-memory",
              "discrete-routing",
              "model-architecture"
            ],
            "local_code_dir": "src/auto_research/reproductions/lngram_v2",
            "fidelity": "core_mechanism",
            "evaluation_tier": "l2_public_dataset",
            "datasets": [
              "Qwen2.5-VL-3B-Instruct checkpoint + deterministic RGB geometry probe"
            ],
            "baseline": "same checkpoint without latent memory branch",
            "metrics": [
              "finite output",
              "route diversity",
              "sink selectivity",
              "activation memory"
            ],
            "default_seeds": [
              42
            ],
            "budget": "adapter-defined fixed budget",
            "device_capabilities": [
              "cuda"
            ],
            "requires_gpu_validation": true,
            "gpu_validation_artifact": "docs/gpu-validations/lngram-v2-a100-20260906.json",
            "online_evidence": [],
            "selection_exception": null,
            "evolve_operators": [
              "memory:latent-ngram-gqa"
            ]
          },
          "provenance": {
            "created_at": "2026-09-05T19:39:00.482653+00:00",
            "code_commit": "6d97d3b2028b651e46229d2596e866d2a6f567c9",
            "python": "3.12.9",
            "platform": "macOS-26.5.2-arm64-arm-64bit",
            "dataset_dir": "/Users/bytedance/Documents/git_daiwk/auto-research/data",
            "dataset_fingerprint": "e66c261317a7b3179720ef3e3f5f79d7a204b135d1f217e2911b153ad2ce50a0",
            "packages": {
              "auto-research": "0.1.0",
              "numpy": "2.5.2",
              "torch": "2.13.0",
              "transformers": "5.14.1"
            }
          },
          "evaluation_protocol": {
            "tier": "l2_public_dataset",
            "tier_label": "L2 公开数据集训练",
            "seeds": [
              43
            ],
            "budget": "paper-specific",
            "formal_comparison": false,
            "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
          }
        },
        {
          "paper": {
            "arxiv_id": "2609.03426",
            "title": "Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations",
            "url": "https://arxiv.org/abs/2609.03426",
            "organization": "Beijing University of Posts and Telecommunications / Kuaishou Technology"
          },
          "dataset": {
            "name": "deterministic cross-modal hidden-state fixture",
            "tokens": 64
          },
          "setup": {
            "adapter": "lngram-v2",
            "seed": 44
          },
          "method": {
            "finite_output": true,
            "sink_weight": 0.19999323785305023,
            "unique_route_ids": 8,
            "route_gradient_norm": 7.803350308677182e-05
          },
          "stages": {
            "hard_discrete_forward": true,
            "counterfactual_surrogate": true,
            "gqa_zero_sink": true,
            "activated_memory_parameters": 128
          },
          "paper_results": {
            "parameter_reduction_percent": 82.6,
            "activated_parameter_reduction_percent": 95.2,
            "id_semantic_recovery_min_percent": 65.77,
            "id_semantic_recovery_max_percent": 84.27
          },
          "scope": "CPU fixture executes hard addressing, exact n-gram lookup, GQA sink and surrogate gradients; A100 receipt validates real VLM hidden states.",
          "manifest_ref": "reproduction:lngram-v2",
          "runtime": {
            "requested_device": "auto",
            "cpu_threads": null,
            "platform": "Darwin arm64",
            "resolved_device": "cpu",
            "torch_version": "2.13.0",
            "accelerator": "arm"
          },
          "seed": 44,
          "schema_version": 2,
          "manifest": {
            "adapter_key": "lngram-v2",
            "arxiv_id": "2609.03426",
            "title": "Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations",
            "paper_url": "https://arxiv.org/abs/2609.03426",
            "track": "llm",
            "organization": "Beijing University of Posts and Telecommunications / Kuaishou Technology",
            "published": "2026-09-03",
            "code_url": null,
            "topics": [
              "multimodal-foundation-model",
              "conditional-memory",
              "discrete-routing",
              "model-architecture"
            ],
            "local_code_dir": "src/auto_research/reproductions/lngram_v2",
            "fidelity": "core_mechanism",
            "evaluation_tier": "l2_public_dataset",
            "datasets": [
              "Qwen2.5-VL-3B-Instruct checkpoint + deterministic RGB geometry probe"
            ],
            "baseline": "same checkpoint without latent memory branch",
            "metrics": [
              "finite output",
              "route diversity",
              "sink selectivity",
              "activation memory"
            ],
            "default_seeds": [
              42
            ],
            "budget": "adapter-defined fixed budget",
            "device_capabilities": [
              "cuda"
            ],
            "requires_gpu_validation": true,
            "gpu_validation_artifact": "docs/gpu-validations/lngram-v2-a100-20260906.json",
            "online_evidence": [],
            "selection_exception": null,
            "evolve_operators": [
              "memory:latent-ngram-gqa"
            ]
          },
          "provenance": {
            "created_at": "2026-09-05T19:39:00.524507+00:00",
            "code_commit": "6d97d3b2028b651e46229d2596e866d2a6f567c9",
            "python": "3.12.9",
            "platform": "macOS-26.5.2-arm64-arm-64bit",
            "dataset_dir": "/Users/bytedance/Documents/git_daiwk/auto-research/data",
            "dataset_fingerprint": "e66c261317a7b3179720ef3e3f5f79d7a204b135d1f217e2911b153ad2ce50a0",
            "packages": {
              "auto-research": "0.1.0",
              "numpy": "2.5.2",
              "torch": "2.13.0",
              "transformers": "5.14.1"
            }
          },
          "evaluation_protocol": {
            "tier": "l2_public_dataset",
            "tier_label": "L2 公开数据集训练",
            "seeds": [
              44
            ],
            "budget": "paper-specific",
            "formal_comparison": false,
            "claim_policy": "single/few-seed smoke result; do not claim a stable improvement"
          }
        }
      ],
      "aggregate_metrics": {
        "dataset.tokens": {
          "mean": 64.0,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "method.sink_weight": {
          "mean": 0.1999827822049459,
          "std": 2.365427024462447e-05,
          "ci95": 2.6767326614707464e-05,
          "n": 3
        },
        "method.unique_route_ids": {
          "mean": 8.0,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "method.route_gradient_norm": {
          "mean": 8.03820657893084e-05,
          "std": 3.403950076325656e-06,
          "ci95": 3.851932126034342e-06,
          "n": 3
        },
        "stages.activated_memory_parameters": {
          "mean": 128.0,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "paper_results.parameter_reduction_percent": {
          "mean": 82.6,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "paper_results.activated_parameter_reduction_percent": {
          "mean": 95.2,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "paper_results.id_semantic_recovery_min_percent": {
          "mean": 65.77,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        },
        "paper_results.id_semantic_recovery_max_percent": {
          "mean": 84.27,
          "std": 0.0,
          "ci95": 0.0,
          "n": 3
        }
      },
      "formal_comparison": true
    }
  },
  "manifest_ref": "reproduction:lngram-v2",
  "evaluation_protocol": {
    "tier": "l2_public_dataset",
    "seeds": [
      42,
      43,
      44
    ],
    "formal_comparison": true,
    "claim_policy": "formal multi-seed comparison"
  },
  "provenance": {
    "artifact_path": "docs/reproductions/2609.03426-lngram-v2/metrics/public-seeds42-44.json",
    "dataset_fingerprint": "not recorded in historical artifact"
  }
}