papers/navil/sau.json

{
  "paper_id": "navil",
  "paper_title": "NaViL: Rethinking Scaling Properties of Native Multimodal LLMs under Data Constraints",
  "D1": [
    {
      "id": "navil-D1-001",
      "claim": "Visual encoder architecture search: param budget=600M, depth candidates=[3,6,12,24,48] layers, width candidates=[4096,2880,2048,1472,1024] hidden dims, all configurations trained from scratch for architecture comparison",
      "source": "Sec 3.2.3"
    },
    {
      "id": "navil-D1-002",
      "claim": "Scaling study independent variables: LLM sizes=[0.5B,1.8B,7B] params, visual encoder sizes=[75M,150M,300M,600M,1.2B,2.4B] params; fixed 600M encoder used when varying LLM size",
      "source": "Sec 3.3.1"
    },
    {
      "id": "navil-D1-003",
      "claim": "Optimal encoder selection criterion (scaling study): lambda_threshold=0.01, base_encoder=75M; select smallest encoder whose validation loss difference vs 2x larger encoder is <1% of 75M baseline loss",
      "source": "Sec 3.3.2"
    },
    {
      "id": "navil-D1-004",
      "claim": "NaVIL-2B architecture: total_params=4.2B, activated_params=2.4B, visual_encoder=0.6B params, LLM_base=InternLM2-1.8B (activated=1.8B), MoE routing with shared experts",
      "source": "Sec 4.1, Sec 5.1, Tab 6"
    },
    {
      "id": "navil-D1-005",
      "claim": "NaVIL-9B architecture: activated_params=9.2B, visual_encoder=1.2B params, LLM_base=Qwen3-8B (activated=8.0B), scaled-up variant of NaVIL design",
      "source": "Appendix A, Tab 6"
    },
    {
      "id": "navil-D1-006",
      "claim": "NaVIL-2B three-stage training: Stage1.1 (max_patches=4096, steps=70K, batch=7000, peak_lr=5e-5), Stage1.2 (steps=40K), Stage2 (fine-tuning on high-quality multimodal data)",
      "source": "Tab 7"
    },
    {
      "id": "navil-D1-007",
      "claim": "NaVIL-9B three-stage training: Stage1.1 (max_patches=4096, steps=50K, batch=10792), Stage1.2 (steps=33K), Stage2 (steps=6K fine-tuning)",
      "source": "Tab 8"
    },
    {
      "id": "navil-D1-008",
      "claim": "Training data composition: Stage1.1=500M image-text pairs (300M web-sampled + 200M synthetic captions), Stage1.2=185M high-quality multimodal+language, Stage2=68M high-quality multimodal data",
      "source": "Sec 4.2"
    },
    {
      "id": "navil-D1-009",
      "claim": "Visual multiscale packing: downsampling_rate tau=sqrt(2)/2 per iteration, stop when image_area < threshold, <end_of_scale> special token separates scale representations",
      "source": "Sec 4.1"
    },
    {
      "id": "navil-D1-010",
      "claim": "MoE early convergence finding: MoE achieves same validation loss as dense with only 1/10 of training data (data efficiency factor=10x), confirmed in architecture search experiments",
      "source": "Sec 3.2.2"
    },
    {
      "id": "navil-D1-011",
      "claim": "LLM initialization convergence finding: uninitialized (random) LLM requires >10x more training data to reach comparable validation loss vs pre-trained initialization, confirming pre-training criticality",
      "source": "Sec 3.2.1"
    }
  ],
  "D2": [
    {
      "id": "navil-D2-001",
      "claim": "Visual Encoder Meta Architecture: V_{d,w}(I) = C circ (circ_{i=1..d} F_i^w) circ P(I)",
      "source": "Sec 3.1, Eq.(1)"
    },
    {
      "id": "navil-D2-002",
      "claim": "Transformer Parameter Count Approximation: N approx 12 x d x w^2",
      "source": "Sec 3.2.3"
    },
    {
      "id": "navil-D2-003",
      "claim": "MoE-Extended LLM Layer with Modality-Specific Experts: x_{i,m}^{l'} = x_{i,m}^{l-1} + MHA-MMoE(RMSNorm(x_{i,m}^{l-1})); x_{i,m}^{l} = x_{i,m}^{l'} + FFN-MMoE(RMSNorm(x_{i,m}^{l'}))",
      "source": "Sec 3.2.2, Eq.(2)-(3)"
    },
    {
      "id": "navil-D2-004",
      "claim": "Modality-Specific Multi-Head Attention Expert (MHA-MMoE): MHA-MMoE(x_{i,m}) = softmax(Q K^T / sqrt(d)) V W_O^m; Q_{i,m}=x_{i,m}W_Q^m, K_{i,m}=x_{i,m}W_K^m, V_{i,m}=x_{i,m}W_V^m",
      "source": "Sec 3.2.2, Eq.(4)-(5)"
    },
    {
      "id": "navil-D2-005",
      "claim": "Modality-Specific Feed-Forward Network Expert (FFN-MMoE): FFN-MMoE(x_{i,m}) = (SiLU(x_{i,m} W_gate^m) circ (x_{i,m} W_up^m)) W_down^m",
      "source": "Sec 3.2.2, Eq.(6)"
    },
    {
      "id": "navil-D2-006",
      "claim": "Visual Multi-Scale Image Packing: H_i = tau^i H_0, W_i = tau^i W_0. Continue until image area < threshold. Concatenate {x_{i,v}}_{i=0}^{n} with <end_of_scale> between scales.",
      "source": "Sec 4.1"
    },
    {
      "id": "navil-D2-007",
      "claim": "Special Token Insertion for Multimodal Sequence: multimodal token sequence S = [t_1, ..., <begin_of_image>, x^{(1,1)}_{i,v}, <end_of_line>, ..., x^{(r,c)}_{i,v}, <end_of_image>, ..., t_n] for each image i. Multi-scale concatenation: X_v = [x_{0,v}, <end_of_scale>, x_{1,v}, ..., <end_of_scale>, x_{n,v}] where x_{i,v} = V_{d,w}(I_i) with H_i = tau^i H_0, W_i = tau^i W_0, and V_{d,w}(·) is the visual encoder defined in Eq.(1). Special tokens delimit image, row, and scale boundaries.",
      "source": "Sec 4.1"
    },
    {
      "id": "navil-D2-008",
      "claim": "Next-Token-Prediction Training Objective: L_CE(θ) = -Σ_{t=1}^{T} log p_θ(s_t | s_{<t}, I), where s_t is the t-th token in the multimodal token sequence S = [t_1, ..., <begin_of_image>, x^{visual}, <end_of_image>, ..., y_T], I is the input image, x^{visual} are visual encoder outputs, and y_T is the target caption token. The model is trained via teacher-forcing: given all preceding tokens (text + image) and the image encoding, predict each next token.",
      "source": "Sec 3.1"
    },
    {
      "id": "navil-D2-009",
      "claim": "Pixel Shuffle Downsampling Connector: C(x) = MLP(PixelShuffle(x)), downsamples encoded image embeddings and projects to LLM feature space.",
      "source": "Sec 3.1"
    },
    {
      "id": "navil-D2-010",
      "claim": "Two-Stage Training with Parameter Freezing Schedule: Stage 1.1 (50K-70K steps, 500M image-text pairs) — θ_trainable = {θ_visual_encoder, θ_MLP_projector, θ_MoE_visual_experts}, θ_frozen = {θ_text (all self-attention + FFN text params)}; Stage 1.2 (33K-40K steps, 185M high-quality data) — unfreeze self-attention text params: θ_trainable += {θ_text_attention}; Stage 2 (6K steps, 68M data) — all params unfrozen: θ_trainable += {θ_text_FFN}. Stage transition condition: Stage 1.1 completes full 500M-sample pass; model enters Stage 1.2 when visual-text alignment at low-res converges.",
      "source": "Sec 4.2"
    },
    {
      "id": "navil-D2-011",
      "claim": "Optimal Visual Encoder Size Selection Criterion: Optimal encoder size = min{s | (L(s) - L(2s)) / L(75M) < lambda}, where lambda = 0.01.",
      "source": "Sec 3.3.2"
    },
    {
      "id": "navil-D2-012",
      "claim": "Image Preprocessing: Padding and Patch Embedding: Padded_H = ceil(H/32)*32, Padded_W = ceil(W/32)*32. Patch stride = 16.",
      "source": "Sec 5.1"
    }
  ],
  "D3": [
    {
      "id": "navil-D3-001",
      "claim": "Compare native MLLM training from scratch versus initializing from a pre-trained LLM (InternLM2-Base) to evaluate the impact on training convergence and zero-shot multimodal performance. Both variants are trained on web-scale noisy image-caption pairs from LAION-2B [55] using next-token prediction with image captioning as the task. Validation teacher-forcing loss is measured on a held-out subset of the multimodal dataset, and zero-shot image captioning quality is evaluated qualitatively. The uninitialized (scratch) model requires over 10x more training data to reach comparable validation loss and shows a substantial zero-shot captioning gap even with significantly more data, due to lower textual quality and diversity of multimodal training data compared to the LLM pre-training corpus.",
      "source": "Sec 3.2.1"
    },
    {
      "id": "navil-D3-002",
      "claim": "Evaluate the effectiveness of modality-specific Mixture-of-Experts architecture (with MHA-MMoE and FFN-MMoE, 1 expert activated) against a vanilla decoder-only LLM for native MLLM training. Both configurations include a visual encoder and are trained on web-scale noisy image-caption pairs from LAION-2B [55] using next-token prediction; validation teacher-forcing loss is measured on a held-out subset. The MoE-extended LLM achieves the same validation loss as the vanilla LLM with only 1/10 of the training data, without increasing training or inference cost. Only FFN experts caused significant feature scale differences between modalities, motivating the additional introduction of modality-specific attention experts (MHA-MMoE).",
      "source": "Sec 3.2.2"
    },
    {
      "id": "navil-D3-003",
      "claim": "Grid search over five depth-width configurations of the visual encoder (depth d in {3, 6, 12, 24, 48} with corresponding width w in {4096, 2880, 2048, 1472, 1024}) under a fixed 600M parameter budget to determine the optimal architecture for native MLLMs. All configurations are trained from scratch on web-scale noisy image-caption pairs from LAION-2B [55] at multiple training data sizes, using a pre-trained LLM (InternLM2-Base) and next-token prediction. Evaluation uses validation teacher-forcing loss on a held-out subset and zero-shot image captioning benchmarks. Shallower encoders converge faster in early training (<30M samples), while deeper encoders perform slightly better with larger datasets; however, a wide range of depth-width combinations yield near-optimal performance, consistent with prior findings on compute-optimal LLM architectures.",
      "source": "Sec 3.2.3"
    },
    {
      "id": "navil-D3-004",
      "claim": "Investigate scaling properties of the LLM component by varying LLM parameter sizes from 0.5B to 7B (based on the InternLM2 architecture family) while keeping the visual encoder fixed at 600M, to characterize the multimodal scaling law. All models follow the optimal architecture (LLM initialization, MoE, optimal depth-width ratio) and are trained on web-scale noisy image-caption pairs from LAION-2B [55] at multiple training data sizes. The metric is validation teacher-forcing loss on a held-out subset. Results show that scaling up LLM parameters in native MLLMs follows a pattern consistent with conventional LLM scaling law: loss decreases log-linearly as parameter size increases exponentially.",
      "source": "Sec 3.3.1 (Scaling up LLMs)"
    },
    {
      "id": "navil-D3-005",
      "claim": "Investigate scaling properties of the visual encoder by varying encoder sizes across {75M, 150M, 300M, 600M, 1.2B, 2.4B} with a fixed LLM to identify diminishing returns and performance saturation points. All models follow the optimal architecture (LLM initialization, MoE, optimal depth-width ratio) and are trained on web-scale noisy image-caption pairs from LAION-2B [55] at multiple training data sizes. The metric is validation teacher-forcing loss on a held-out subset. Unlike LLM scaling, performance gains from larger visual encoders diminish progressively; beyond a certain encoder size, further scaling yields only marginal loss reduction, indicating that the MLLM performance upper bound is constrained by the LLM's capacity. As training data increases, the loss gap between consecutive encoder sizes narrows to near zero once the visual encoder reaches a sufficient size.",
      "source": "Sec 3.3.1 (Scaling up Visual Encoder)"
    },
    {
      "id": "navil-D3-006",
      "claim": "Determine the optimal visual encoder size for each LLM scale (0.5B, 1.8B, 7B) using the lambda=1% threshold criterion and characterize the log-linear relationship between optimal encoder size and LLM size. Based on validation loss measurements on web-scale image-caption pairs from LAION-2B [55] across multiple encoder sizes for each LLM, the optimal encoder is defined as the smallest encoder whose loss difference versus a 2x larger encoder is less than 1% of the 75M baseline encoder's loss. The logarithm of optimal visual encoder size scales linearly with the logarithm of LLM size, indicating both components should be scaled jointly for balanced performance — a finding that contrasts with the compositional paradigm, which typically uses a single pre-trained visual encoder across all LLM scales.",
      "source": "Sec 3.3.2"
    },
    {
      "id": "navil-D3-007",
      "claim": "Evaluate NaViL-2B (2.4B activated params) against both compositional and native MLLMs on 14 multimodal benchmarks. General MLLM understanding benchmarks (7): MMVet, MMMU (val), MMBench-EN (test), MME (perception + cognition sum), MathVista (MINI), OCRBench, CCBench (Table 1). Visual question answering benchmarks (7): TextVQA (val), ScienceQA-IMG (test), GQA (test-dev), DocVQA (test), AI2D (test), ChartQA (test), InfoVQA (test) (Table 2). Key baselines: compositional counterpart InternVL-2.5-2B (same LLM and high-quality data, 300M encoder distilled from 6B encoder), Qwen2VL-2B, InternVL-1.5-2B, MiniCPM-V-2, DeepSeek-VL-1.3B, PaliGemma-3B, MM1-3B-MoE-Chat; native MLLMs Mono-InternVL (1.8B), EVEv2 (7B), SAIL (7B), Emu3 (8B), EVE-7B, VoRA, Chameleon-7B. Metrics: per-benchmark task-specific scores, with average computed by normalizing each metric to a 0-100 range. NaViL-2B achieves average 67.1 on Table 1 and 75.1 on Table 2, outperforming all existing native MLLMs and competitive with compositional baselines.",
      "source": "Sec 5.1-5.2, Table 1, Table 2"
    },
    {
      "id": "navil-D3-008",
      "claim": "Scale NaViL to 9B parameters (9.2B activated, based on Qwen3-8B with 1.2B visual encoder, visual multi-scale packing disabled in Stage 1.1 for acceleration) and evaluate on 12 multimodal benchmarks (MMVet, MMMU, MMBench, MME, MathVista, OCRBench, TextVQA, DocVQA, AI2D, ChartQA, InfoVQA; Table 4), comparing training token efficiency against compositional MLLMs. NaViL-9B uses only 450B total training tokens (all in MLLM training, zero ViT pre-training tokens) versus >4.1T for Qwen2.5VL and >3.5T for InternVL2.5-8B (whose visual encoders separately consume >3.3T pre-training tokens; Table 3). Key baselines from Table 4: compositional counterpart InternVL-2.5-8B (same high-quality data, 300M encoder distilled from 6B), Qwen2.5-VL-8B, Qwen2VL-8B; native MLLMs EVEv2 (7B), SAIL (7B), Mono-InternVL (1.8B), Emu3 (8B), EVE-7B, Chameleon-7B. NaViL-9B achieves average 77.0 on Table 4, outperforming all existing native MLLMs by a large margin and competitive with top-tier compositional baselines.",
      "source": "Appendix A, Table 3, Table 4"
    },
    {
      "id": "navil-D3-009",
      "claim": "Visualize and analyze attention maps of different LLM layers in NaViL (using LLM-1.8B as the base) when using a small (150M) versus large (1.2B) visual encoder to understand how encoder scaling affects cross-modal interaction. Trained models with each encoder size are loaded; sample images are forward-passed, and attention weights are extracted per LLM layer for visualization (Fig. 9). Two findings: (1) with a 150M encoder, shallow-layer attention exhibits obvious locality (attending to spatially adjacent tokens), gradually shifting to global patterns at deeper layers; with a 1.2B encoder, shallow-layer visual tokens already attend to global information, indicating better pre-extraction of high-level semantics; (2) larger encoder facilitates earlier cross-modal interaction — attention weights between visual and text tokens in the first layer are significantly higher with the 1.2B encoder versus 150M, providing an explanatory mechanism for the improved performance from larger encoders.",
      "source": "Sec 5.3"
    },
    {
      "id": "navil-D3-010",
      "claim": "Evaluate the NLP capability of NaViL-9B on three standard NLP benchmarks (MMLU, CMMLU, MATH) using the OpenCompass evaluation toolkit [18], to verify that modality-specific MoE architecture preserves linguistic competence of the initialization LLM (Qwen3-8B). Baselines from Table 5: InternLM2-Chat (1.8B: MMLU 47.1, CMMLU 46.1, MATH 13.9), Qwen3-8B non-thinking (MMLU 76.5, CMMLU 76.8, MATH 71.1), EVE (7B: MMLU 43.9, CMMLU 33.4, MATH 0.7), Chameleon (7B: MMLU 52.1, MATH 11.5), Mono-InternVL (2B: MMLU 45.1, CMMLU 44.0, MATH 12.3). NaViL-9B scores: MMLU 74.9, CMMLU 75.1, MATH 66.2 — close to Qwen3-8B and far above other native MLLMs, confirming that modality-specific MoE effectively preserves NLP capability despite multimodal training, without requiring large amounts of high-quality text data.",
      "source": "Appendix E, Table 5"
    }
  ],
  "D4": [
    {
      "id": "navil-D4-001",
      "claim": "Experiment phases: Pretrain both variants (scratch vs. LLM-initialized) at increasing data sizes -> Compare validation loss curves -> Evaluate zero-shot image captioning",
      "source": "Sec 3.2.1"
    },
    {
      "id": "navil-D4-002",
      "claim": "Experiment phases: Train vanilla LLM and MoE-extended LLM configurations -> Compare validation loss curves -> Measure data efficiency ratio (10x data reduction with MoE)",
      "source": "Sec 3.2.2"
    },
    {
      "id": "navil-D4-003",
      "claim": "Experiment phases: Grid search all 5 (d,w) configs under 600M budget -> Train each at multiple data sizes -> Compare validation loss curves -> Evaluate zero-shot image captioning at convergence",
      "source": "Sec 3.2.3"
    },
    {
      "id": "navil-D4-004",
      "claim": "Experiment phases: Train MLLMs with LLM sizes {0.5B, 1.8B, 7B} (fixed 600M encoder) -> Measure validation loss across training data sizes -> Fit log-linear scaling law relationship",
      "source": "Sec 3.3.1 (Scaling up LLMs)"
    },
    {
      "id": "navil-D4-005",
      "claim": "Experiment phases: Train MLLMs with encoder sizes {75M, 150M, 300M, 600M, 1.2B, 2.4B} (fixed LLM) -> Measure validation loss across training data sizes -> Identify saturation point where larger encoder yields marginal gain -> Conclude performance upper bound constrained by LLM capacity",
      "source": "Sec 3.3.1 (Scaling up Visual Encoder)"
    },
    {
      "id": "navil-D4-006",
      "claim": "Experiment phases: For each LLM size (0.5B, 1.8B, 7B), sweep encoder sizes -> Measure loss gaps between consecutive sizes -> Apply lambda=0.01 threshold criterion to determine optimal encoder size -> Fit log-log relationship showing optimal encoder scales proportionally with LLM size",
      "source": "Sec 3.3.2"
    },
    {
      "id": "navil-D4-007",
      "claim": "Experiment phases: Stage 1.1 pretraining (500M pairs, frozen text) -> Stage 1.2 continued pretraining (185M, unfrozen attention) -> Stage 2 supervised fine-tuning (68M, all unfrozen) -> Evaluate NaViL-2B on 14 benchmarks -> Compare against compositional and native MLLM baselines",
      "source": "Sec 5.1-5.2, Table 1, Table 2"
    },
    {
      "id": "navil-D4-008",
      "claim": "Experiment phases: Train NaViL-9B with same recipe as NaViL-2B (visual multi-scale packing disabled in S1.1) -> Evaluate on 12 benchmarks (Table 4) -> Compare training token counts against Qwen2.5VL and InternVL-2.5-8B (Table 3) -> Compare benchmark scores against all baselines",
      "source": "Appendix A, Table 3, Table 4"
    },
    {
      "id": "navil-D4-009",
      "claim": "Experiment phases: Load trained NaViL models with 150M and 1.2B encoders -> Forward pass sample images -> Extract attention weights per LLM layer -> Visualize and compare attention patterns (locality/globality, cross-modal interaction)",
      "source": "Sec 5.3"
    },
    {
      "id": "navil-D4-010",
      "claim": "Experiment phases: Pretrain and fine-tune NaViL-9B -> Evaluate on MMLU, CMMLU, MATH using OpenCompass -> Compare against baseline LLMs (InternLM2-Chat, Qwen3-8B) and native MLLMs (EVE, Chameleon, Mono-InternVL)",
      "source": "Appendix E, Table 5"
    }
  ]
}